USING SUCCESSIVE RE GUL ARIZATION

CROSS REFERENCE TO OTHER APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application No.

62/417,269, "JOINT MANY-TASK MODEL" (Atty. Docket No. SALE 1182-1/1948PROV1), filed on November 03, 2016. The priority provisional application is hereby incorporated by reference. This provisional application includes a technical paper that supplies additional details of the inventors' work;

[0002] This application claims the benefit of U.S. Provisional Patent Application No.

62/418,070, "JOINT MANY-TASK MODEL'' (Atty. Docket No. SALE 1 182-2/1948PROV2), filed on November 04, 2016. The priority provisional application is hereby incorporated by reference. This provisional application includes a technical paper that supplies additional details of the inventors' work;

[0003] This application claims the benefit of U.S. Nonprovisional Patent Application No. 15/421 ,431 , "TRAINING A JOINT MANY-TASK NEURAL NETWORK MODEL USING SUCCESSIVE REGULARIZATION" " (Atty. Docket No. SALE 1182-5/1946US3), filed on January 31 , 2017. The priority nonprovisional application is hereby incorporated by reference;

[0004] This application incorporates by reference US Nonprovisional Patent Application 15/421 ,407, "JOINT MANY-TASK NEURAL NETWORK MODEL FO MULTIPLE

ATURAL LANGUAGE PROCESSING (NLP) TASKS" (Atty. Docket No. SALE 1 182- 3/1948US1 ), filed on January 31 , 2017; and

[0005] This application incorporates by reference US Nonprovisional Patent Application No. 15/421,424, "DEEP NEURAL NETWORK MODEL FOR PROCESSING DATA THROUGH MUTLIPLE LINGUISTIC TASK HIERARCHIES" (Atty. Docket No. SALE 1182-4/1948US2), filed on January 31 , 2017.

FIELD OF THE TECHNOLOGY DISCLOSED

[0006] The technology disclosed relates generally to an architecture for natural language processing (NLP) using deep neural networks, and in particular relates to multi-task learning using an end-to-end trainable joint many-task neural network model. This architecture is extensible to other multilayer analytical frameworks and tasks.

BACKGROUND

[0007] The subject matter discussed in this section should not be assumed to be prior art merely as a result of its mention in this section. Similarly, a problem mentioned in this section or associated with the subject matter provided as background should not be assumed to have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which in and of themselves can also correspond to implementations of the claimed technology.

[0008] Transfer and multi-task learning have traditionally focused on either a single source- target pair or very few, similar tasks. Ideally, the linguistic levels of morphology, syntax and semantics would benefit each other by being trained in a single model. The technology disclosed provides a so-called ''joint many-task neural network model" to solve a variety of increasingly complex natural language processing (NLP) tasks using a growing depth of layers in a single end-to-end model. The model is successively trained by considering linguistic hierarchies, directly connecting word representations to all model layers, explicitly using predictions in lower tasks, and applying a so-called "successive regularization" technique to prevent catastrophic forgetting. Three examples of lower level model layers are part-of-speech (POS) tagging layer, chunking layer, and dependency parsing layer. Two examples of higher level model layers are semantic reiatedness layer and textual entailment layer. The model achieves the state-of-the-art results on chunking, dependency parsing, semantic reiatedness and textual entailment.

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In the drawings, like reference characters generally refer to like parts throughout the different views. Also, the drawings are not necessarily to scale, with an emphasis instead generally being placed upon illustrating the principles of the technology disclosed. In the following description, various implementations of the technology disclosed are described with reference to the following drawings, in which:

[0010] FIG. 1A illustrates aspects of a joint-many task neural network model that performs increasingly complex NLP tasks at successive layers.

[0011] FIGs. IB and 1C show various modules that can be used to implement the joint- many task neural network model

[ 0012] FIG. 2A depicts a joint-embedding technique that is herein used to robustly encode the input words, especially unknown words.

[0013] FIG. 2B illustrates various tables that demonstrate that the use of the character n- gram embeddings results in improved processing of unknown words.

[0014] FIG. 3 shows one implementation of dimensionality projection.

[0015] FIG, 4A shows one impleme tation of operation of a POS layer of the joint many- task neural network model.

[0016 ] FIG. 4B includes a table that shows the results of POS tagging of the joint many-task neural network model. [ 0017] FIG. 5A shows one implementation of operation of a chunking layer of the joint many-task neural network model.

[0018] FIG. 5B includes a table that shows the results of POS tagging of the joint many -task neural network model

[0019] FIG. 6A shows one implementation of operation of a dependency parsing layer.

[0020] FIGs. 6B, 6C, 6B, 6E, and 6F show one implementation of operation of an attention encoder of the dependency parsing layer.

[0021] FIG. 6G shows one implementation of operation of a dependency relationship label classifier of the dependency parsing layer.

[0022] FIG. 6H shows two example sentences on which model applies dependency parsing.

[0023] FIG. 61 includes a table that shows the results of the dependency parsing layer of the model.

[0024] FIG. 7A shows one implementation of the semantic reiatedness layer.

[0025] FIG, 7B includes a table that shows the results of the semantic reiatedness task.

[0026] FIG. 8A shows one implementation of the entailment layer.

[0027] FIG. 8B includes a table that shows the results of the entailment task.

[0028] FIG. 9 A shows one implementation of training a stacked LSTM sequence processor that is stacked with at least three layers according to an analytical hierarchy.

[0029] FIG. 9B includes a table that demonstrates the effectiveness of the successive regulari zation technique.

[0030] FIG, 10 includes a table that shows results of the test sets on the five different NLP tasks.

[0031] FIG. 11 is a simplified block diagram of a computer system that can be used to implement the joint many-task neural network model. DETAILED DESCRIPTION

[0032] The following discussion is presented to enable any person skilled in the art to make and use the technology disclosed, and is provided in the context of a particular application and its requirements. V arious modifications to the disclosed implementations will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other implementations and applications without departing from the spirit and scope of the technology disclosed. Thus, the technology disclosed is not intended to be limited to the implementations shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein. Introduction

[0033] Multiple levels of linguistic representation are used in a variety of ways in the field of Natural Language Processing (NLP). For example, part-of-speech (POS) tags are applied by syntactic parsers. The POS tags improve higher-level tasks, such as natural language inference, relation classification, sentiment analysis, or machine translation. However, higher level tasks are not usually able to improve lower level tasks, often because systems are unidirectional pipelines and not trained end-to-end.

[0034] In deep learning, supervised word and sentence corpuses are often used to initialize recurrent neural networks (RNNs) for subsequent tasks. However, not being jointly trained, deep NLP models have yet to show benefits from stacking layers of increasingly complex linguistic tasks. Instead, existing models are often designed to predict different tasks either entirely separately or at the same depth, ignoring linguistic hierarchies.

[0035] An overall theme of the technology disclosed is a so-called "joint many-task neural network model" that performs increasingly complex NLP tasks at successive layers. Unlike traditional NLP pipeline sy stems, the joint many-task neural network model is trained end-to-end for POS tagging, chunking, and dependency parsing. It can further be trained end-to-end on semantic relatedness, textual entailment, and other higher level tasks. In a single end-to-end implementation, the model obtains state-of-the-art results on chunking, dependency parsing, semantic relatedness and textual entailment. It also performs competitively on POS tagging. Additionally, the dependency parsing layer of the model relies on a single feed-forward pass and does not require a beam search, which increases parallelization and improves computational efficiency.

[0036] To allow the joint many-task neural network model to grow in depth while avoiding catastrophic forgetting, we also disclose a so-called "successive regularization" technique.

Successive regularization allows multi-layer training of model weights to improve one NLP task's loss without exhibiting catastrophic interference of the other tasks. By avoiding catastrophic interference between tasks, the model allows the lower and higher level tasks to benefit from the joint training,

[0037] To improve generalization and reduce overfitting in the joint many-task neural network model, we farther disclose a so-called "dimensionality projection" technique.

Dimensionality projection includes projecting low-dimensional output of a neural network classifier into a high-dimensional vector space. This projection from a low-dimensional space to a high-dimensional space creates a dimensionality bottleneck that reduces overfitting.

[0038] To robustly encode the input words, especially unknown words, provided to the joint many-task neural network model, we disclose a "joint-embedding" technique. Joint-embedding includes representing an input word using a combination of word embedding of the word and character n-gram embeddings of the word. Joint-embedding efficiently encodes morphological features and information about unknown words.

[0039] FIG. 1 illustrates aspects of a joint-many task neural network model 100 that performs increasingly complex NLP tasks at successive layers. In implementations, model 100 is a stacked long-short-term-memory ("LSTM") sentence processor that is stacked in layers according to a linguistic hierarchy, with bypass connections that deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. The linguistic hierarchy builds from the words in the sentence (e.g., sentence,, or sentence^ ), to the parts of speech, to the chunks of the sentence, to dependency links between the words and their dependency parents, to labels on the dependency links. In the example shown in FIG. 1, model 100 includes two LSTM stacks (i.e., stack a and stack b) with similar architectures. In one implementation, model 100 includes just one LSTM stack. In another implementation, model 100 includes more than two LSTM stacks (e.g., 3, 4, 10, and so on).

[0040] In model 100, the stacked layers include a part-of-speech (POS) label embedding layer (e.g., 104a or 104b), a chunk/chunking label embedding layer (e.g., 106a or 106b) overlying the POS label embedding layer; and a dependency parent identification and dependency relationship label embedding layer (e.g., 108a or 108b) overlying the chunk label embedding layer.

[0041] The POS label embedding layer is implemented as a bi-directional LSTM that uses a POS label classifier. It processes word embedding vectors (e.g., 102a or 102b) representing the words in the input sentence and produces POS label embedding vectors and POS state vectors for each of the words.

[0042] The chunk label embedding layer is implemented as a bi-directional LSTM that uses a chuck label classifier. It processes at least the word embedding vectors, the POS label embedding vectors and the POS state vectors, to produce chunk label embeddings, and chunk state vectors.

[0043] The dependency parent identification and dependency relationship label embedding layer is implemented as a bi-directional LSTM that uses one or more classifiers. It processes the word embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vectors, to identify dependency parents of each of the words in the sentence to produce dependency relationship labels or label embeddings of relationships between the words and respective potential parents of the word. 10044] Also, dimensionalities of the POS label embedding vectors, the chunk label embedding vectors, and the dependency relationship label embedding vectors are similar, within +/- ten percent.

[0045] In some implementations, model 100 includes an output processor that outputs at least results reflecting the identification of dependency parents and production of dependency relationship label embeddings for the words in the sentence. In the example shown in FIG. 1, relatedness encoders (e.g., 110a or 110b) can be considered outside processors that provide the dependency relationship label embeddings to a relatedness layer (e.g., 112). The relatedness layer provides a categorical classification of relatedness between the first and second sentences and delivers the classification to an entailment layer (e.g., 116) via entailment encoders (e.g., 114a or 114b). The entailment layer outputs a categorical classification of entailment between the first and second sentences. In implementations, the relatedness layer and the entailment layer are used as output processors.

[0046] Regarding the bypass connections, a bypass connection supplies an input vector used by an underly ing layer to an overlaying layer without modification. In the example shown in FIG. 1, "type 2" bypass connections provide the word representations directly to each layer in the model 100. In another example of bypass connections, "type 3" bypass connections provide POS label embedding vectors generated at the POS label embedding layer to each of the overlaying layers. In another example of bypass connections, "type 4" bypass connections provide chunk label embeddings generated at the chunk label embedding layer to each of the overlaying layers.

[0047] Model 100 also includes connections that deliver information from an underlying layer to only the successive overlaying layer. For instance, "type 5" connection provides a categorical classification of relatedness between the first and second sentences calcula ted at the semantic relatedness layer to the entailment layer. "Type 6" connection outputs a categorical classification of entailment between the first and second sentences from the entailment layer. Also, "type 1" connections provide hidden state vectors generated at a given layer only to the successive overlaying layer.

[0048] The components in FIG. 1 can be implemented in hardware or software, and need not be divided up in precisely the same blocks as shown in FIG. 1. Some of the components can also be implemented on different processors or computers, or spread among a number of different processors or computers. In addition, it will be appreciated that some of the components can be combined, operated in parallel or in a different sequence than that shown in FIG. 1 without affecting the functions achieved. Also as used herein, the term "component" can include "sub- components", which themselves can be considered herein to constitute components. For example, the POS label embedding layer and the chunk label embedding layer can also be considered herein to be sub-components of a "word level processor" component. Similarly, the dependency parent identification and dependency relationship label embedding layer can also be considered herein to be a sub-component of a "syntactic level processor" component. Likewise, the semantic relatedness layer and the entailment layer can also be considered herein to be subcomponents of a "semantic level processor" component. Furthermore, the blocks in FIG. 1 can also be thought of as flowchart steps in a method. A component or sub-component also need not necessarily have all its code disposed contiguously in memory; some parts of the code can be separated from other parts of the code with code from other components or sub-components or other functions disposed in between.

[0049] In some implementations, model 100 is a stacked LSTM token sequence processor, stacked in layers according to an analytical hierarchy, with bypass connections that deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. In such implementations, the stacked layers of the model 100 include a first embedding layer, a second embedding layer overlying the first embedding layer, and a third embedding layer overlying the second embedding layer.

[0050] In one implementation, the first embedding layer of the model 100, implemented as a bi-directional LSTM and a first label classifier, processes token embeddings representing the tokens in the input sequence, and produces first embeddings and first state vectors of the tokens. In one implementation, the second embedding layer of the model 100, implemented as a bidirectional LSTM and a second label classifier, processes at least the token embeddings, the first label embeddings and first state vectors, to produce second label embeddings and second state vectors. In one implementation, the third embedding layer of the model 100, implemented as a bi-directional LSTM, processes at least the token embeddings, the first label embeddings, the second label embeddings and the second state vectors to produce third label embeddings and third state vectors. In one implementation, an output processor of the model 100 outputs at least results reflecting the third label embeddings for the tokens in the input sequence.

[0051] In some implementations, the first embedding layer further produces first label probability mass vectors, by exponential normalization of the first state vectors, and produces the first label embedding vectors, from the first label probability mass vectors. In some

implementations, the second embedding layer further produces second label probability mass vectors, by exponential normalization of the second state vectors, and produces the second label embedding vectors from the second label probability mass vectors. In some implementations, the third embedding layer further produces third label probability mass vectors, by exponential normalization of the third state vectors, and produces the third label embedding vectors from the third label probability mass vectors. In implementations, dimensionalities of the first label embedding vectors, the second label embedding vectors, and the third label embedding vectors are similar, within +/- ten percent.

[0052] In one implementation, model 100 includes a token embedding processor, underlying the first label embedding layer, that includes a token embedder and a decomposed token embedder. The token embedder maps the tokens in the sequence, when recognized, into a token embedding space represented by a token embedding vector. The decomposed token embedder processes token decompositions of the token at multiple scales, maps each processed token decomposition into an intermediate vector representing a position in a token decomposition embedding space, and combines the intermediate vectors for each unique processed token decomposition to produce token decomposition embedding vectors for each of the tokens. The token embedding processor combines results of the token embedder and the decomposed token embedder, whereby a token not previously mapped into the token embedding space is nonetheless represented by the token decomposition embedding vector, Joint-Embedding

[0053] FIG. 2 A depicts a joint-embedding technique 200 used to robustly encode the input words, especially unknown words. Joint-embedding includes, for each word w _t in the input sequence S of length L , constructing a so-called "word representation" 222 by concatenating a word embedding 210 of the word W _{ and one or more character »-gram embeddings of the word W _f , also referred to herein as "n-character-gram" embeddings. In FIG. 2A, the concatenation operation is denoted by the " + " symbol.

[0054] Regarding the word embeddings, model 100 includes a word embedder 202 that trains a word embedding matrix to create a word embedding space 204. In one implementation, the word embedder 202 uses a skip-gram model to train the word embedding matrix. In another implementation, it uses a continuous bag-of-words (CBOW) model to train the word embedding matrix. In implementations, the word embedding matrix is shard across all the NLP tasks of the model 100. In some implementations, the words which are not included in the vocabulary are mapped to a special " UNK " token.

[0055] Regarding the character n-gr&m embeddings, model 100 includes a character embedder 206 that trains a character embedding matrix to create a character embedding space 208. In one implementation, the character embedder 206 uses a skip-gram model to train the word embedding matrix. In another implementation, it uses a continuous bag-of-words (CBOW) model to train the character embedding matrix. In implementations, the character w-gram embeddings are learned using the same skip-gram objective function as the word vectors.

[0056] Character embedder 206, also referred to herein as an "o-character-gram embedder", constructs the vocabulary of the character »-grams in the training data and assigns an embedding for each character w-gram. In the example shown in FIG. 2, the character embedding space 208 includes a 1-gram embedding 212, a 2-gram embedding 214, a 3-gram embedding 216, and a 4- gram embedding 218. In other implementations, it includes embeddings for different, additional, and/or fewer w-grams.

[0057] The final character embedding 220 combines, element-wise, vectors representing the unique character «-gram embeddings of the word w, , For example, the character «-grams ( n =

1 , 2, 3) of the word "Cat" are {C, a, t, #BEGIN#C, Ca, at, t#END#, #BEGIN#Ca, Cat, at#END#} , where "#BEGIN#" and "#E D#" represent the beginning and the end of each word, respectively. Element- ise combination of vectors representing these substrings can be element- wise averages or maximum values. The use of the character «-gram embeddings efficiently provides morphological features and information about unknown words. Accordingly, each word is represented as word representation X _t 222, the concatenation of its corresponding word embedding 210 and character embedding 220.

[0058] In implementations, the word embedder 202 and the character embedder 206 are part of a so-called "word embedding processor". The POS label embedding layer overlays the word embedding processor. The word embedder 202 maps the words in the sentence, when recognized, into the word embedding space 204 represented by a word embedding vector. The n- character-gram embedder 206 processes character substrings of the word at multiple scales of substring length, maps each processed character substring into an intermediate vector representing a position in the character embedding space 208, and combines the intermediate vectors for each unique processed character substring to produce character embedding vectors for each of the words. The word embedding processor combines results of the word embedder 202 and the »-character-gram embedder 206, whereby a word not previously mapped into the word embedding space is nonetheless represented by the character embedding vector. The handling of unknown or out-of- vocabulary (OoV) words applies well to other NLP tasks, such as question answering.

[0059] In some implementations, the »-character-gram embedder 206 combines the intermediate vectors to produce element wise average in the character embedding vector.

[0060] The POS label embedding layer further processes «-character-gram embedding vectors that represent the words in the input sentence, in addition to the word embedding vectors, and the bypass connections further deliver the rt-character-gram embedding vectors to the chunk label embedding layer and the dependency parent and dependency relationship label embedding layer as input to respective bi-directional LSTMs in those overlaying layers.

[0061] In regards to training, the word embeddings are trained using the skip-gram or the CBOW model with negative sampling, according to one implementation. The character «-gram embeddings are also trained similarly. In some implementations, one difference between the training of the word embeddings and the character «-gram embeddings is that each input word in the skip-gram model is replaced with its corresponding average embedding of the character rt- gram embeddings. Also, these embeddings are fine-tuned during the joint training of the model 100 such that, during backpropagation, the gradients are used to update corresponding character

/7-gram embeddings. The embedding parameters are denoted as " Θ ".

[0062] In one implementation, the vocabulary of the character rt-grams is built on the training corpus, the case-sensitive English Wikipedia text. Such case-sensitive information is important in handling some types of words like named entities. Assuming that the word W _t has its corresponding K character rt-grams {cfl^ , CF -, CH ^ } , where any overlaps and unknown entries are removed. Then, the word W _t is represented with an embedding V (w) , computed as follows:

v _. (w) =™∑v(cii.), where V CH? is the parameterized embedding of the character «-grani CYl. ,

[0063] Furthermore, for each word-context pair (w, w) in the training corpus, negative context words are sampled, and the objective function is defined as follows:

_ N

∑ {-log σ (v ( )) * v(w)) -∑ log σ (- v _c (w) · v(w _f ))} ,

(W, w, ' _j= i

where (7 (·) is the logistic sigmoid function, v(w) is the weight vector for the context word, and W is a negative sample.

[0064] FIG. 2B illustrates various tables that demonstrate that the use of the character rt- gram embeddings results in improved processing of unknown words. This is demonstrated in table 224 of FIG. 2B, which shows the results for the three single tasks, POS tagging, chunking, and dependency parsing, with and without the pre-trained character rt-gram embeddings. The column of "W&C" corresponds to using both of the word and character rt-gram embeddings, and that of "Only W" corresponds to using only the word embeddings. These results clearly show that jointly using the pre-trained word and character n-grarn embeddings is helpful in improving the results. The pre-training of the character n-gram embeddings is also effective; for example, without the pre-training, the POS accuracy drops from 97.52% to 97.38% and the chunking accuracy drops from 95.65% to 95.14%, but they are still better than those of using word2 vec embeddings alone.

[0065] Table 226 of FIG. 2B shows that the joint use of the word and the character w-gram embeddings improves the score by about 19% in terms of the accuracy for unknown words. Table 228 of FIG. 2B shows dependency parsing scores on the development set with and without the character «-gram embeddings, focusing on UAS and LAS for unknown words, UAS stands for unlabeled attachment score. LAS stands for labeled attachment score. UAS studies the structure of a dependency tree and assesses whether the output has the correct head and dependency relationships. In addition to the structure score in UAS, LAS also measures the accuracy of the dependency labels on each dependency relationship. Table 228 clearly indicates that using the character-level information is effective, and in particular, the improvement of the LAS score is large.

Dimensionality Projection

[0066] FIG. 3 show ^r s one implementation of dimensionality projection 300. Dimensionality projection includes conveying intermediate results from an underlying layer to an overlying layer in a neural network stack of bidirectional LSTMs, in which the stack has layers corresponding to an analytical framework that process a sequence of tokens, and the underlying layer produces analy tic framework label vectors for each of the tokens.

[0067] In FIG. 3, the hidden state vectors 314 are generated by a neural network, such as a LSTM or bidirectional LSTM, or any other RNN. Hidden state vectors 314 are encoded in a high dimensional vector space 302 and have a dimensionality of 1 x , which is identified element- wise as {d , . , ., d ., . . . , _* di ,,} such that d represents an individual dimension and the sub-script denotes an ordinal position of the dimension. In one example, jjf j— 200· ^ ^{n οηε} implementation, a classifier 304 classifies the hidden state vectors 314 into an analytic framework label space 306 as label space vectors that have dimensionality about the same as a number of available framework labels. The analytic framework label space 306 encodes linguistic meaniiigfulness. For instance, if the POS label embedding layer has twenty labels, then a= 20· In one implementation, the classifier 304 just includes a dimensionality reduction matrix W■ In another implementation, the classifier 304 includes an exponential normalizer 308 (e.g., a

softmax) in addition to the dimensionality reduction weight matrix W , which normalizes the a

label space vectors produced by the dimensionality reduction weight matrix W .

a

[0068] Once created, the low dimensional label space vectors are projected into an extended dimensionality label space 312 by a dimensionality augmentation weight matrix W 310 to produce extended token label vectors 316, The extended dimensionality label space 312 is a high dimensional vector space. Thus, like the hidden state vectors 314, the label vectors 316 are also mapped to a high dimensional vector space and have a dimensionality of 1 χ jjj j , which is identified element-wise as {/, , /_ , . . ., / . , . . . . , . , } such that / represents an individual

M - 2 ^· - j "

dimension and the sub-script denotes an ordinal position of the dimension. Note that the iabel vectors 316 have dimensionality about the same as dimensionality of the hidden state vectors 314. By about the same, we mean within +/- ten percent. It is not necessary that the

dimensionality be the same, but programming can be easier when they are,

10069] Model 100 uses dimensionality projection at various stages of processing. In one impleme tation, it uses it to project the POS label embeddings in a higher dimensional space such that low dimensional POS analytic label space vectors are projected into a vector space where they have the same dimensionality as the POS hidden state vectors used to generate them. In another implementation, model 100 uses dimensionality projection to project the chunk label embeddings in a higher dimensional space such that low dimensional chunk analytic label space vectors are projected into a vector space where they have the same dimensionality as the chunk hidden state vectors used to generate them. Likewise, in other implementations, other layers use dimensionality projection.

[0070] In one implementation, when the number of available analytical framework labels is one-fifth or less the dimensionality of the hidden state vectors 314, the label space vectors 316 serve as dimensionality bottleneck that reduce overfitting when training the model 100. In another implementation, when the number of available analytical framework labels is one-tenth or less the dimensionality of the hidden state vectors 314, the label space vectors 316 serve as dimensionality bottleneck that reduce overfitting when training the model 100.

[0071] The dimensionality bottleneck also improves processing in other NLP tasks such as machine translation. [0072] FIG. 4A shows one implementation of operation of a POS layer 400 of the model 100.

[0073] The POS label embedding layer, also referred to herein as the "POS layer", produces POS label probability mass vectors (e.g., 404), by exponential normalization (e.g., softmax 406 with a single ReLU layer) of the POS state vectors (e.g., 408), and produces the POS label embedding vectors (e.g., 402), from the POS label probability mass vectors.

[0074] In one implementation, the POS label embedding layer 400 of the model 100 is a bidirectional LSTM 410, whose hidden states are used to predict POS tags. In one implementation, the following LSTM units are used for the forward pass direction: f a (W _t g _t + b _f },

ο, = σ (W _o g _{t +} h _o ),

u = tanh (W g _t + b ), : ^{" l} : ^U , ⁺ f, Γ

h - o tanh (c ),

where the input g. is defined as — [ _¾ · _x ^■ ] , i.e., the concatenation of the previous hidden state and the word representation _x . The backward pass of the LSTM over words is expanded in the same way, but with a different set of weights.

[ 0075] For predicting the POS tags of _w , the concatenation of the forward and backward states is used in a one-layer bi-LSTM layer corresponding to the t -th word: — ; h ^pos ] . Then each ¾ (] < < /_, ) is fed into an exponential normalizer with a single ReLU layer which outputs the probability vector for each of the POS tags.

[0076] FIG. 4B includes a table that shows the results of POS tagging of the model 100. Model 100 achieves the score close to the state-of-the art results.

Word-Level Task - Chunking

[0077] FIG. 5A shows one implementation of operation of a chunking layer 500 of the model 100. [0078] Chunking is also a word-level classification task which assigns a chunking tag (B-NP, I- VP, etc.) for each word. The tag specifies the region of major phrases (or chunks) in the sentence.

[0079] The chunk label embedding layer 500, also referred to herein as the "chunking layer", further produces chunk label probability mass vectors (e.g., 504), by exponential normalization (e.g., softmax 506 with a single ReLU layer) of the chunk label state vectors (e.g., 508), and produces the chunk label embedding vectors (e.g., 502 from the chunk label probability mass vectors.

[0080] In model 100, chunking is performed using a second bi-LSTM layer 510 on top of the POS layer. When stacking the bi-LSTM layers, the LSTM units are provided with the following input:

where }f ^os is the hidden state of the first POS layer. The weight label embedding yP ^0S is defined as follows:

where C is the number of the POS tags, p ( _^ yP ^S = j j hP ⁰ ) is the probability mass for the j -th POS tag is assigned to word _w , and / is the corresponding label embedding. As previously discussed, the label embedding can be at a higher dimensionality than the probability mass. The probability values are automatically predicted by the POS label embedding layer working like a built-in POS tagger, and thus no gold POS tags are needed, in some

implementations.

[0081] For predicting the chunking tags, similar strategy as POS tagging is employed by using the concatenated bi-directional hidden states ¾c — \h ^chk h ^chk ] ^m the chunking layer. In some implementations, a single ReLU hidden layer is used before the exponential classifier.

[0082] FIG, SB includes a table that shows the results of POS tagging of the model 100. Model 100 achieves the state-of-the art results, which show that the lower level tasks are also improved by the joint learning in addition to the higher level tasks.

[0083] FIG. 6A shows one implementation of operation of a dependency parsing layer 600 of the model 100. [0084] Dependency parsing identifies syntactic relationships (such as an adjective modifying a noun) between pairs of words in a sentence,

[0085] The dependency parent identification and dependency relationship label embedding layer 600, also referred to herein as the "dependency layer or dependency parsing layer", produces parent label probability mass vectors by classification and exponential normalization of parent label state vectors 602 produced by the bi-directional LSTM 604, also referred to herein as the "dependency parent analyzer". The dependency parsing layer produces the parent label embedding vectors from the parent label probability mass vectors, produces dependency relationship label probability mass vectors by classification and exponential normalization of the parent label state vectors and the parent label embedding vectors, and produces the dependency relationship label embedding vectors from the dependency relationship label probability mass vectors.

[0086] The dependency parent analyzer 604 processes the words in the input sentences, including processing, for each word, word embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vector to accumulate forward and backward state vectors 602 that represent forward and backward progressions of interactions among the words in the sentence.

[0087] FIGs. 6B, 6C, 6D, 6E, and 6F show one implementation of operation of an attention encoder 610 of the dependency parsing lay er 600. The attention encoder 610 processes the forward and backward state vectors 602 for each respective word in the sentence to encode attention as inner products 612 between each respective word and other words in the sentence, after apply ing a linear transform 608 to the forward and backward state vectors 602 for the word or the other words, whereby weights 606 in the linear transform are trainable. In some implementations, a sentinel vector 622 is used by the attention encoder 610 to encode the root word.

[0088] The attention encoder 610 further applies exponential normalization 614 to vectors 616 of the inner products 612 to produce parent label probability mass vectors 618 and projects the parent label probability mass vectors to produce parent label embedding vectors by mixing or calculating a weighted sum 620 of the linear transformations of the forward and backward state vectors 602 in dependence upon the parent label probability mass vectors 618.

[0089] FIG. 6G shows one implementation of operation of a dependency relationship label classifier 626 of the dependency parsing layer. The dependency relationship label classifier 626, for each respective word in the sentence, classifies and normalizes (using another exponential normalizer such as softmax 628) the forward and backward state vectors 602 and the parent label embedding vectors 624, to produce dependency relationship label probability mass vectors 630, and projects the dependency relationship label probability mass vectors 630 to produce dependency relationship label embedding vectors 632.

[0090] As discussed above, for dependency parsing, model 100 uses a third bi-LSTM layer 604 on top of the POS and chunking layers to classify relationships between ail pairs of words. As shown in FIG. 6 A, the input vector for the dependency parsing LSTM includes hidden states word representations 102 and the label embeddings 402 and 502 for the two previous tasks:

where the chunking vector is computed in a similar fashion as the POS vector in equation (1) above. The POS and chunking tags 402 and 502 are used to improve dependency parsing.

[0091] Like sequential labeling task, model 100 predicts the parent node, also referred to herein as "head", for each word in the sentence. Then, a dependency label is predicted for each of the child-parent node pairs. To predict the parent node of the t -th of the word ^ , model 100 defines a matching function 612 (based on dot product/inner product or bi-linear product) between _w and the candidates of the parent node as: m (t, /) = h† ^{'p T} W _d h† ^p , where is a parameter matrix. As discussed above, for the root, model 100 defines , _f = r as a parameterized sentinel vector 622. As discussed above, to compute the probability that vv

(or the root node) is the parent of H", , the scores are normalized using an exponential normalizer (e.g. softmax 614), as follows:

where L is the sentence length.

Next, the dependency labels are predicted using h ^'e ^ h , ^e ^ as input to another

^ J

exponential normalizer (e.g., softmax 628 with a single ReLU layer). At test time, in one implementation, the parent node and the dependency label are greedily selected for each word in the sentence. That is, model 100 operates without a beam search in the POS label embedding layer, the chunk label embedding layer or the dependency parent identification and dependency relationship label embedding layer. This results because the model 100 calculates the label embeddings on a word-by-word basis, which increases parailelization and improves computational efficiency because it avoids the redundancies and computational latency typically caused by beam search. In addition, the word-by-word computation, during the dependency parsing, allows the model 100 to correct any incorrectly labeled roots such that if a word is predicted as the root of itself, the model 100 can detect that as an incorrect prediction and can automatically calculate a new correct prediction for the word.

[0093] In one implementation, model 100 assumes that each word in the sentence only has one parent node. In another implementation, model 100 assumes that each word can have multiple parent nodes and produces dependency labels using cyclic graphs-like computation. At training time, model 100 uses the gold or ground truth child-parent pairs to train the label predictor.

[0094] FIG. 6H shows two example sentences on which model 100 applies dependency parsing. In the example (a), the two boldfaced words "counsels" and "need" are predicted as child nodes of the root node, and the underlined word "counsels" is the correct one based on the gold annotations. In the example (b), none of the words are connected to the root node, and the correct child node of the root is the underlined word "chairman". Model 100 uses the single parameterized vector r to represent the root node for each sentence and captures various types of root nodes. In some implementations, model 100 uses sentence-dependent root representations.

[0095] FIG. 61 includes a table that shows the results of the dependency parsing layer of the model 100. Model 100 achieves the state-of-the-art results. Note that the greedy dependency parser of the model 100 outperforms the previous state-of-the-art result which based on beam search with global information. This shows that the bi-LSTMs of the model 100 efficiently capture global information necessary for dependency parsing.

Semantic. Task - Semantic. Relatedness

[0096] FIG. 7A shows one implementation of operation of a semantic relatedness layer 700 of the model 100.

[0097] The next two NLP tasks of the model 100 encode the semantic relationships between two input sentences. The first task measures the semantic relatedness between two sentences. The output of the semantic relatedness layer is a real- values relatedness score for the input sentence pair. The second task is a textual entailment task, which requires one to determine whether a premise sentence entails a hypothesis sentence. These are typically three classes: entailment, contradiction, and neutral,

[0098] The two semantic tasks are closely related to each other. In implementations, good semantic relatedness is represented by a low semantic relatedness score. Thus if the semantic relatedness score between two sentences is very low, i.e., the two sentences have high semantic relatedness, they are likely to entail each other. Based on this intuition and to make use of the information from lower layers, model 100 uses the fourth and fifth bi-LSTM layers for the relatedness and entailment task, respectively.

[0099] Since these tasks require sentence-level representation rather than the word-level representation used in the previous tasks, model 100 computes the sentence-level representation } d as the element-wise maximum values across all the word-level representations in the s

fourth layer, as follows:

Ivy ¹ = max (h[ ^el , ψ ¹ , . . ., h _L ^rel ),

where L is the length of the sentence.

[00100] To model the semantic relatedness between s and s' , a feature vector is calculated as follows:

d, _t {s, ') = [ ; ^{V ~} - if' ¹ 1 : If' ¹ (2) where j ] ^e! — ]f ^el j is the absolute value of the element-wise subtraction, and Jf ^el h ^rd * ^s the element-wise multiplication. Both these operations can be regarded as two different similarity metrics of the two vectors. Then, J ( _$f ^') is fed into an exponential normalizer (e.g., softmax) with a single Maxout hidden layer 722 to output a related score (e.g., between 1 and 5) for the sentence pair. The Maxout hidden layer 722 includes a plurality of linear functions with (e.g., pool size is four) that each generate a non-linear projection of the (s, ,< _? ') such that a maximum non-linear projection is fed to the exponential normalizer.

[00101] Turning to FIG. 7A, the semantic relatedness layer 700 with a bi-directional LSTM 702, overlying the dependency parent identification and dependency relationship label embedding layer 600, also includes a relatedness vector calculator 720 and a relatedness classifier 714. The relatedness vector calculator 720 calculates a sentence-level representation 708a and 708b of each of the first and second sentences, including a bi-directional LSTM calculation of forward and backward state vectors 704 for each of the words in the respective sentences and an element-wise max pooling calculation 706 over the forward and backward state vectors 704 for the words in the respective sentences to produce sentence-level state vectors 708a and 708b representing the respective sentences. An alternative implementation could use average pooling. Then, the relatedness vector calculator 720 further calculates an element-wise sentence-level relatedness vector 712 that is processed by the relatedness classifier 714 to derive a categorical classification of relatedness between the first and second sentences. In some implementations, the relatedness vector calculator reports the categorical classification for further processing, such for generated relatedness label embeddings 718. [00102] The relatedness vector calculator 720 includes a feature extraction module 716 that calculates element-wise differences between the sentence-level relatedness vectors 708a and 708b for the first and second sentences, calculates element-wise products between sentence-level relatedness vectors 708a and 708b for the first and second sentences, and uses vectors of absolute values of the element- wise differences and the element- wise products as inputs to the relatedness classifier 714.

[00103] FIG. 7B includes a table that shows the results of the semantic relatedness task. Model 100 achieves the state-of-the-art result.

Semantic Task - Textual Entailment

[00104] For entailment classification between two sentences, model 100 also uses the max- pooling technique as in the semantic related task. To classify the premise-hypothesis pair (s, s ') into one of the three classes, model 100 computes the feature vector d ( _β as in equation

(2), except that it does not use the absolute values of the element-wise subtraction so as to identify which is the premise (or hypothesis). Then, d {s, s') ^ fed into an exponential normaiizer (e.g., softmax) with multiple Maxout hidden layers (e.g., three Maxout hidden layers) 822.

[00105] A Maxout hidden layer applies multiple linear functions and non-linear activations to an input and selects the best result. When multiple Maxout hidden layers are arranged in a stack, the maximum output from a preceding Maxout hidden layer is provided as input to a successive Maxout hidden layer. The maximum output of the last Maxout hidden layer in the stack is provided to the exponential normaiizer for classification. Note that temporal max pooling just evaluates, element-wise, individual dimensions of multiple input vectors and selects a maximum dimension value for each ordinal position to encode in a max pooled vector. In contrast, a Maxout hidden layer subjects an input vector to multiple non-linear transformations and selects just one linear transformation that has maximum dimension values.

[00106] To make direct use of the output from the relatedness layer, model 100 uses the label ernbeddings for the relatedness task. Model 100 computes the class label embeddings for the semantic related task similar to equation (1). The final feature vectors that are concatenated and fed into the entailment classifier are weighted relatedness embedding and the feature vector d ₂ (s, s').

[00107] Turning to FIG. 8A, the entailment layer 800 with a bi-directional I .S I Vi 802, overlying the semantic entailment layer 800, also includes a entailment vector calculator 820 and a entailment classifier 814. The entailment vector calculator 820 calculates a sentence-level representation 808a and 808b of each of the first and second sentences, including a bi-directional LSTM calculation of forward and backward state vectors 804 for each of the words in the respective sentences and an element-wise max pooling calculation 806 over the forward and backward state vectors 804 for the words in the respective sentences to produce sentence- level state vectors 808a and 808b representing the respective sentences. An alternative

implementation could use average pooling. Then, the entailment vector calculator 820 further calculates an element-wise sentence-level entaiiment vector 812 that is processed by the entaiiment classifier 814 to derive a categorical classification of entailment between the first and second sentences. In some implementations, the entailment vector calculator reports the categorical classification for further processing, such for generated entailment label embeddings 818.

[00108] The entailment vector calculator 820 includes a feature extraction module 816 that calculates element-wise differences between the sentence-level entailment vectors 808a and 808b for the first and second sentences, calculates eiement-wise products between sentence-level entaiiment vectors 808a and 808b for the first and second sentences, and uses vectors of absolute values of the element-wise differences and the element-wise products as inputs to the entailment classifier 814.

[00109] FIG, 8B includes a table that shows the results of the textual entailment task. Model 100 achieves the state-of-the-art result. Training - Successive Reguiarization

[00110] In NLP tasks, multi-task learning has the potential to improve not only higher-level tasks, but also lower-level tasks. Rather than treating as fixed the pre-trained model parameters, the disclosed successive reguiarization allows model 100 to continuously train the lower-level tasks without catastrophic forgetting.

[00111] Model 100 is trained jointly over all datasets. During each epoch, the optimization iterates over each full training dataset in the same order as the corresponding tasks described above.

Training the POS Layer

[00112] One training corpus for the POS layer 400 is the Wall Street journal (WSJ) portion of Penn Tree-bank. This corpus includes tokens labelled with POS tags. During training of the POS layer 400, L2-norm reguiarization is applied to the layer parameters of the POS layer 400 because it is the current layer. Successive reguiarization is applied to layer parameters of just one underlying layer, namely, the embedding layer, which includes the word embedding space 204 and the character embedding space 208. POS layer 400, where is the set of the weight matrices in the first bi-LSTM and the pos

classifier, and jj is the set of the bias vectors. The objective function to optimize a is defined as follows:

[po

where p y(P° ^s ) = a s)

is the probability value that the correct label CC is assigned to W ι _† in the sentence S , pos is the L2-norm regularization term, and A is a L2-norm regularization hyperparameter. is the successive reeularization term. The

successive regularization prevents catastrophic forgetting in the model 100 and thus prevents it from, forgetting the information learned for the other tasks. In the case of POS tagging, the regularization is applied to $ , and Θ is the embedding parameters after training the final task in the top-most layer at the previous training epoch, ό is a successive regularization

hyperparameter, which can be different for different layers of the model 100 and can also be val e-assigned variedly to network weights and biases. Training the Chunking Layer

[00114] To train the chunking layer 500, the WSJ corpus was used in which the chunks are labelled. During training of the chunking lay er 500, L2-norm regularization is applied to the layer parameters of the chunking layer 500 because it is the current layer. Successive

regularization is applied to layer parameters of two underlying layers, namely, the embedding layer and the POS layer 400.

[00115] The objective function for the chunking layer is defined as follows:

Λ (* ;h»k) = pos pos

which is similar to that of POS tagging, and Θ ^ is jf ^r , b E , θ ), where J/J _{f ;} and chk ' chk' ' e'

h are the weight and bias parameters including those in Θ . and E is the set of the

-chk pos ^' pos POS label embeddings. _s ^are the POS parameters after training the POS layer 400 at the current training epoch.

116] To train the dependency parsing layer 600, the WSJ corpus with dependency labels was used. During training of the dependency parsing layer 600, L2-norm regularization is applied to the layer parameters of the dependency parsing layer 600 because it is the current layer. Successive regularization was applied to layer parameters of three underlying layers, namely, the embedding layer, the POS layer 400, and the chunking layer 500.

[00117] The objective function for the dependency layer is defined as follows:

wher (X is for W _t and p f]\ h ^dep , h ^dep ) is the probability value assigned to the correct dependency label β for the child-parent pair (w. , (X ) . θ , is defined as W b W r E E Θ ) where l a dep ' den ^' ' rf ' ' ^" pas' ^' chk' e' '

is the

set of the chunking label embeddings.

Training the Relatedness Layer

|00118] At the semantic relatedness layer 700, training used the SICK dataset. During training of the semantic relatedness layer 700, L2-norm regularization was applied to the layer parameters of the semantic relatedness layer 700 because it is the current layer. Successive regularization was applied to layer parameters of four underlying layers, namely, the embedding layer, the POS layer 400, the chunking layer 500, and the dependency parsing layer

[00119] The ob ective function for the relatedness layer is defined as follows:

J4 (0 ,)— - Θ d ' ep

where p($. $') is the gold distribution over the defined relatedness score, p p f ^sl _, Iff) is the predicted distribution given the sentence representations, and ^ ^ j el^ IS the p ip (}

KL - divergence between the two distributions, θ , is defined as (w j _y _ β . Ε _{, ,} θ ) l l reV pos ^*J chk^ e ' Training the Entailment Layer

[00120] To train the entailment layer 800, we also used the SICK dataset. During training of the entailment layer 800, L2-norm regularization is applied to the layer parameters of the entailment layer 800 because it is the current layer. Successive regularization is applied to layer parameters of five underlying layers, namely, the embedding layer, the POS layer 400, the chunking layer 500, the dependency parsing layer 600, and the semantic relatedness layer ' [00121] The objective function for the entailment layer is defined as follows:

E is the set of the relatedness label embeddings.

rel

Epochs of Training

[00122] Turning to FIG. 9 A, FIG. 9A shows one implementation of training a stacked LSTM sequence processor that is stacked with at least three layers according to an analytical hierarchy. In FIG. 9A, first, second and third layers (e.g., POS layer 400, chunking layer 500, and dependency layer 600) are trained by backward propagation using training examples directed to each layer, with regularized pass down to underlying layers during training. The training includes training the first layer using first layer training examples (e.g., POS data), training the second layer using second layer training examples (e.g., chunking data) with regularized pass down training to the first layer, and training the third layer using third layer training examples (e.g., dependency data) with regularized pass down training to the first and second layers. The regularized pass down training is regularized by constraining a training objective function, having a fitness function with at least two regularization terms. The two regularization terms regularize by penalizing growth in a magnitude of weights in coefficient matrices applied to the underlying layers and that successively regularize all changes in the weights in the coefficient matrices applied to the underlying layers. In one implementation, the fitness function is cross- entropy loss. In another implementation, the fitness function is KL-divergence. In yet another implementation, the fitness function is mean squared error.

[00123] In the example shown in FIG. A, two sub-epochs of a single epoch are depicted. In one implementation, model 100 has five sub-epochs corresponding to five NLP tasks, in each sub-epoch, a batch of training examples ... i ϋ corresponding to the current layer's training data is processed. Every time a training example is processed by a current layer, the layer parameters 0S , , . . of the underlying layers and the layer parameters

^{J 5} underlying layers ' ^" ' ' ^!

έ? , of the current layer are updated by back-propagating the gradients.

current layer "

0 denotes the updated value of a parameter Θ of an n under lying layer ^underlying layer

underlying layer as a result of back-propagation on a given training example of the current layer. Also, at the end of each sub-epoch, a "snapshot" is taken of the current state of the embedding parameters of the current layer and of the current state of the embedding parameters of all the underlying layers. The snapshotted values are persisted in memory as Θ , , . , and underlying layers are referred to herein as "current anchor values".

[00124] At the end of each sub-epoch, the successive regularization term

ensures that the update value Θ does not significantly deviate from the current nunderlying layer anchor values θ , , . , of the layer parameters.

underlying layers

[00125] In FIG.9 A, the first sub-epoch at the POS layer 400 starts with current anchored values of the embedding layer and successive regularizes only the embedding layer parameters

Θ . Note that successive regularization is not applied to the parameters of the current layer, i.e.,

2 layer parameters Θ the POS layer 400, and only L2-norm regularization

pos X pos is applied to the current layer's updated parameters to produce regularized current layer parameters

Θ . Successive regularization ensures that the layer parameter values of the underlying layers pos updated during the training of the POS layer 400, i.e. θ , do not significantly deviate from the currently anchored values Θ . This produces successively regularized underlying layer parameters Θ . At the end of the sub-epoch, the most recently regularized current layer parameters 0 ⁿ _)S and the most recently successively regularized underlying layer parameters are subjected to the snapshot operation and persisted in memory as the new current anchored values.

|00126] In the next layer, such as chunking layer 500, the underlying layer parameters now include parameters for the embedding layer and the POS layer. These underlying layer parameters are subjected to successive regularization, while the current layer parameters of the chunking layer are only subjected to L2-norm regularization. This process continues for all the layers of the model 100.

[00127] FIG. 9B includes a table that demonstrates the effectiveness of the successi ve regularization technique. In FIG. 9B, the column of "w/o SR" shows the results of not using successive regularization in the model 100. It can be seen that the accuracy of chunking is improved by the successive regularization, while other results are not affected so much. The chunking dataset used here is relatively small compared with other low-level tasks, POS tagging and dependency parsing. Thus, these results suggest that the successive regularization is effective when dataset sizes are imbalanced.

[00128] FIG. 10 includes a table that shows results of the test sets on the five different NLP tasks. In FIG. 10, the column "Single" shows the results of handling each task separately using single-layer bi-LSTMs, and the column "JMTall" shows the results of the model 100. The single task settings only use the annotations of their own tasks. For example, when treating dependency parsing as a single task, the POS and chunking tags are not used. It can be seen that all results of the five different tasks are improved in the model 100, which shows that model 100 handles the five different tasks in a single model. Model 100 also allows accessing of arbitrary information learned from the different tasks. For example, in some implementations, to use the model 100 just as a POS tagger, the output from the first bi-LSTM layer can be used. The output can be the weighted POS label embeddings as well as the discrete POS tags.

[00129] The table in FIG. 10 also shows the results of three subsets of the different tasks. For example, in the case of "JMTABC", only the first three layers of the bi-LSTMs are used to handle the three tasks. In the case of "JMTDE", only the top two layers are used j ust as a two- layer bi-LSTM by omitting all information from the first three layers. The results of the closely- related tasks show that model 100 improves not only the high-level tasks, but also the low-level tasks.

[00130] Other implementations of the technology disclosed include using normaiizers different than, in addition to, and/or in combination with the exponential normalize!'. Some examples include sigmoid based normaiizers (e.g., multiclass sigmoid, piecewise ramp), hyperbolic tangent based normaiizers, rectified linear unit (ReLU) based normaiizers, identify based normalizers, logistic based normalizers, sine based normalizers, cosine based normalizers, unit sum based normalizers, and step based normalizers. Other examples include hierarchical softmax, differentiated softmax, importance sampling, noise contrastive estimation, negative sampling, gated softmax spherical softmax, Taylor softmax, and sparsemax. In yet other implementations, any other conventional or future-developed normalize!' can be used.

[00131 ] While this technology is discussed with respect to bidirectional LSTMs, there are other emerging forms of LSTMs that may evolve as substitutes of LSTMs. In other

implementation, the technology disclosed uses unidirectional LSTMs in one or more or all layers of the model 100. Examples of some variants of the LSTM include no input gate (NIG) variant, no forget gate (NFG) variant, no output gate (NOG) variant, no input activation function (NIAF) variant, no output activation function (NOAF) variant, coupled input-forget gate (CIFG) variant, peephole (PH) variant, and full gate recurrent (FGR) variant. Yet other implementations include using a gated recurrent unit (GRU), or any other type of RN , or any other conventional or future-developed neural network.

[00132] In yet other implementations, the layers of the model 100 are stacked in the form of a directed acyclic graph. In such implementations, some layers may not be successively on top of others and may instead be acyclically arranged.

[00133] We describe systems, methods, and articles of manufacture for a so-called "joint many-task neural network model" to sol ve a variety of increasingly complex natural language processing (NLP) tasks using growing depth of layers in a single end-to-end model.

Implementations that are not mutually exclusive are taught to be combinable. One or more features of an implementation can be combined with other implementations. This disclosure periodically reminds the user of these options. Omission from some implementations of recitations that repeat these options should not be taken as limiting the combinations taught in the preceding sections - these recitations are hereby incorporated forward by reference into each of the following implementations.

[00134] FIGs. IB and 1C show various modules that can be used to implement the joint- many task neural network model Previously described modules or components of the model 100, such as the word representation layers 102ab, the POS layers 104ab, the chunking layers 106ab, the dependency layers 108ab, the relatedness layers HOab and 112, and the entailment layers 114ab and 116 can alternatively be described using smaller modularized modules or components without changing their principle of operation or the model 100. [00135] The modules in FIGs. IB and 1C can be implemented in hardware or software, and need not be divided up in precisely the same blocks as shown in FIGs. IB and 1C. Some of the modules can also be implemented on different processors or computers, or spread among a number of different processors or computers. In addition, it will be appreciated that some of the modules can be combined, operated in parallel or in a different sequence than that shown in FIGs. 1.4 and IB without affecting the functions achieved. Also as used herein, the term

"module' ^' can include "sub -modules", which themselves can be considered herein to constitute modules. For example, a word embedder module 1021 and a word n-character gram module 1022 can be considered herein to be sub-modules of the word representation modules 102ab. In another example, a POS processing module 1041 and a POS production module 1042 can be considered herein to be sub-modules of the POS modules 104ab. In yet another example, a dependency processing module 1081, a dependency identity module 1082, a dependency production module 1083, an embedding processing module 1084, a mass vector processing module 1085, and a parent label vector production module 1086 can be considered herein to be sub-modules of the dependency modules 108ab. In a further example, an attention encoder 1087, an attention encoder module 1087, a parent label vector module 1089, and a parent labeling module 1086 can be considered herein to be sub-modules of the dependency modules 108ab. In yet another example, a dependency parent analyzer module 1180, an embedding module 1181, a state vector production module 1182, a normalization module 1184, a dependency relationship label vector production module 1187, and a dependency label vector production module 1188 can be considered herein to be sub-modules of the dependency modules 108ab. In yet further example, a sentence input module 1101 , a sentence representation module 1102, a relatedness vector determiner module 1103, and a relatedness classifier module 1104 can be considered herein to be sub-modules of the relatedness encoder modules l lOab and/or the relatedness module 112. In yet another example, an entailment vector determiner module 1141, a pooling module 1142, and an entailment classifier module 1143 can be considered herein to be sub- modules of the entailment encoder modules 114ab and/or the entailment module 116. The blocks in in FIGs, IB and 1C, designated as modules, can also be thought of as flowchart steps in a method. A module also need not necessarily have all its code disposed contiguously in memory; some parts of the code can be separated from other parts of the code with code from other modules or other functions disposed in between.

[00136] In one implementation, a multi-layer neural network system, running on hardware that processes words in an input sentence, is described that includes a stacked long-short-term- memory (abbreviated LSTM) sentence processor, running on hardware, stacked in layers according to a linguistic hierarchy. The stacked LSTM sentence processor can be embodied in a stack of LSTM sentence modules. The stacked LSTM includes bypass connections that deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. The stacked layers include (i) a part-of-speech (abbreviated POS) label embedding layer, (ii) a chunk label embedding layer overlying the POS label embedding layer, and (iii) a dependency parent identification and dependency relationship label embedding layer overlying the chunk label embedding layer. The POS label embedding layer, implemented as a bi-directional LSTM and a POS label classifier, processes word embedding vectors representing the words in the input sentence, and produces POS label embedding vectors and POS state vectors for each of the words. These components of the POS label embedding layer 104 can be embodied in a POS processing module 1041 for processing word embedding vectors

representing the words in the input sentence, and a POS production module 1042 for producing POS label embedding vectors and POS state vectors for each of the words.

[00137] The chunk label embedding layer 106, implemented as a bi-directional LSTM and a chuck label classifier, processes at least the word embedding vectors, the POS label embedding vectors and the POS state vectors, to produce chunk label embeddings and chunk state vectors. These components of the chunk label embedding layer 106 can be embodied in a chunk processing module 1061 for processing at least the word embedding vectors, the POS label embedding vectors and the POS state vectors, and a chunk production module 1062 for producing chunk label embeddings and chunk state vectors.

[00138] The dependency parent identification and dependency relationship label embedding layer 108, implemented as a bi-directional LSTM and one or more classifiers, processes the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors, to identify dependency parents of each of the words in the sentence to produce dependency relationship labels or label embeddings of rela tionships between the words and respective potential parents of the words. These components of the dependency parent identification and dependency relationship label embedding layer 108 can be embodied in a dependency processing module 1081 for processing the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors, and a dependency identification module 1082 for identification of dependency parents of each of the words in the sentence and a dependency production module 1083 for producing dependency relationship labels or label embeddings of relationships between the words and respective potential parents of the words.

[00139] The multi-layer neural network system also includes an output processor that outputs at least results reflecting the identification of dependency parents and production of dependency relationship label embeddings for the words in the sentence. [00140] This system and other implementations of the technology disclosed can each optionally include one or more of the following features and'Or features described in connection with additional systems disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00141] In an implementation of the disclosed multi-layer neural network system, the linguistic hierarchy builds trom the words in the sentence, to the parts of speech, to the chunks of the sentence, to dependency links between the words and their dependency parents, to labels on the dependency links.

[00142] A bypass connection supplies an input vector used by an underlying layer to an overlying layer without modification.

[00143] In some implementations, the POS label embedding layer 104 farther processes n- character-gram embedding vectors that represent the words in the input sentence, in addition to the word embedding vectors. Additionally , the bypass connections deliver the n-character-gram embedding vectors to the chunk label embedding layer and the dependency parent and dependency relationship label embedding layer as input to respective bi-directional LSTMs in those overlaying layers. These further components of the word representation layer 102 can be embodied in a word embedder module 1021 and an n-character-gram embedder module 1022. The bypass connections can be embodied in connections with a chunk processing module and a dependence processing module.

[00144] The POS label embedding layer 1 4 can further produce POS label probability mass vectors, by exponential normalization of the POS state vectors, and produces the POS label embedding vectors, from the POS label probability mass vectors. This functionality can be embodied in a POS module 104. Additionally, the chunk label embedding layer 106 produces chunk label probability mass vectors, by scaling normalization of the chunk label state vectors, and produces the chunk label embedding vectors from the chunk label probability mass vectors. This functionality can be embodied in a chunk production module 1062. Furthermore, the dependency parent identification and dependency relationship label embedding layer 108 produces parent label probability mass vectors by classification and scaling normalization of parent label state vectors produced by the bi-directional LSTM, This functionality can be embodied in a dependency identification module 1082. The dependency parent identification and dependency relationship label embedding layer also produces the parent label embedding vectors from the parent label probability mass vectors, produces dependency relationship label probability mass vectors by classification and exponential normalization of the parent label state vectors and the parent label embedding vectors, and produces the dependency relationship label embedding vectors from the dependency relationship label probability mass vectors. This functionality can be embodied in a dependency production module 1083. The dimensionalities of the POS label embedding vectors, the chunk label embedding vectors, and the dependency relationship label embedding vectors are similar, within +/- ten percent

[00145] The technology disclosed can further include a word embedding layer or processor 102, underlying the POS label embedding layer. The word embedding processor includes a word embedder 1021 and an n-character-gram embedder 1022. The word embedder maps the words in the sentence, when recognized, into a word embedding space represented by a word embedding vector. Additionally, the n-character-gram embedder (i) processes character substrings of the word at multiple scales of substring length, (ii) maps each processed character substring into an intermediate vector representing a position in a character embedding space, and (iii) combines the intermediate vectors for each unique processed character substring to produce character embedding vectors for each of the words. The sentence embedding processor also combines results of the word embedder and the n-character-gram embedder, whereby a word not previously mapped into the word embedding space is nonetheless represented by the character embedding vector. These components of the word embedding layer 102 can be embodied in a word embedder module 1021 for mapping the words in the sentence and an n-character-gram embedder module 1022 for mapping character substrings of different scales in the words, with the POS processing module 1041 further processing output of the n-character-gram embedder module to represent a word not previously mapped into the word embedding space.

[00146] The n-character-gram embedder can combine the intermediate vectors in at least two ways. It can produce element wise averages in the character embedding vector or it can select element wise maximums. The POS label classifier can include a softmax layer or, more generally, an exponential normalizes These alternatives also apply to the chunk label classifier. These alternative features can be embodied in an n-character-gram embedder module and/or a chunk processing or chunk production module.

[00147] The disclosed technology operates well without a beam search in the POS label embedding layer, the chunk label embedding layer or the dependency parent identification and dependency relationship label embedding layer. It could be implemented with a beam search having a narrow span.

[00148] The dependency parent identification and dependency relationship label embedding layer further includes a dependency parent layer and a dependency relationship label classifier. The dependency parent identifier layer includes a dependency parent analyzer, implemented as a bi-directional LSTM, that processes the words in the input sentences. Specifically, the dependency parent analyzer processes, for each word, word embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vector to accumulate forward and backward state vectors that represent forward and backward progressions of interactions among the words in the sentence. The dependency parent identifier layer also includes an attention encoder that processes the forward and backward state vectors for each respective word in the sentence, and encodes attention as inner products between embeddings of each respective word and other words in the sentence, with a linear transform applied to the forward and backward state vectors for the word or the other words prior to the inner product. Additionally, the attention encoder applies exponential normalization to vectors of the inner products to produce parent label probability mass vectors and projects the parent label probability mass vectors to produce parent label embedding vectors. Further, the technology disclosed includes a dependency relationship label classifier that, for each respective word in the sentence, (i) classifies and normalizes the forward and backward state vectors and the parent label embedding vectors and the parent label embedding vectors, to produce dependency relationship label probability mass vectors, and (ii) projects the dependency relationship label probability mass vectors to produce dependency relationship label embedding vectors. These components of the dependency parent identification and dependency relationship label embedding 108 can be embodied in a dependency parent analyzer module for processing the words in input sentences, and an attention encoder module for processing the forward and backward state vectors for producing parent label probability mass vectors and parent label embedding vectors.

[00149] In an implementation, the disclosed multi-layer neural network system further includes a semantic relatedness layer, overlying the dependency parent identification and dependency relationship label embedding layer. The semantic relatedness layer includes a relatedness vector calculator and a relatedness classifier and operates on pairs of first and second sentences processed through the multi-layer neural network system. The relatedness vector calculator, of the technology disclosed, determine a sentence-le vel representation of each of the first and second sentences. The determinations performed by the relatedness vector calculator include (i) a bi-directional 5 S I ^' M calculation of forward and backward state vectors for each of the words in the respective sentences, and (ii) an element-wise max pooling calculation over the forward and backward state vectors for the words in the respective sentences to produce sentence-level state vectors representing the respective sentences. The relatedness vector calculator further calculates an element- wise sentence- level relatedness vector that is processed by the relatedness classifier to derive a categorical c lassification of relatedness between the first and second sentences. This layer can reports the categorical classification for further processing. [00150] Components of the semantic relatedness layer 110 can be embodied in a sentence input module 1101, a sentence representation module 1102, a relatedness vector determiner 1103, and a relatedness classifier 1104; the sentence input module 1101 for inputting pairs of first and second sentences processed through the stack of sentence modules; the relatedness vector determiner 1102 for determining a sentence-le vel representation of each of the first and second sentences, including a bi-directional LSTM for determining forward and backward state vectors for each of the words in the respective sentences and a pooling module for element-wise max pooling over the forward and backward state vectors for the words in the respective sentences, and a sentence representation module 1103 for producing sentence-level state vectors representing the respective sentences; and the relatedness classifier 1104 for categorically classifying relatedness between the first and second sentences.

[00151] The relatedness vector calculator can also (i) determine element-wise differences between the sentence-level relatedness vectors for the first and second sentences, (ii) determine element-wise products between the sentence-level relatedness vectors for the first and second sentences, and (iii) use vectors of absolute values of the element-wise differences and the element-wise products as inputs to the relatedness classifier.

[00152] The technology disclosed can further include an entailment layer that overlays the semantic relatedness layer. The entailment layer includes an entailment vector calculator and an entailment classifier. Further, the entailment vector calculator calculates a sentence-level representation of each of the first and second sentences. The calculations performed by the entailment vector calculator can include (i) a bi-directional LSTM calculation of forward and backward state vectors for each of the words in the respective sentences, and (ii) an element-wise max pooling calculation over the forward and backward state vectors for the words in the respective sentences to produce sentence-level state vectors representing the respective sentences. The entailment vector calculator can further calculate an element-wise sentence-level entailment vector that is processed by the entailment classifier to derive a categorical classification of entailment between the first and second sentences. This layer can reports the categorical classification for further processing.

[00153] Components of the entailment layer 114 can be embodied in an entailment vector determiner 1141 for determining a sentence-level representation of each of the first and second sentences, including a bi-directional LSTM for determining forward and backward state vectors for each of the words in the respective sentences and a pooling module 1142 for element- wise max pooling over the forward and backward state vectors for the words in the respective sentences, and a sentence represen ation module 102 for producing sentence-level state vectors representing the respective sentences; and an entailment classifier 1143 for categorically classifying entailment between the first and second sentences.

[00154] The entailment vector determiner or calculator can further (i) determine element-wise differences between the sentence-level relatedness vectors for the first and second sentences, (ii) determine element-wise products between the sentence-level relatedness vectors for the first and second sentences, and (iii) use vectors of the element- wise differences and the element- wise products as inputs to the relatedness classifier,

[00155] In another implementation, a method is provided that processes words in an input sentence using a stacked layer long-short- term-memory (abbreviated LSTM) sentence processor, running on hardware, stacked in layers according to a linguistic hierarchy. This stack can be embodied in a stack of LSTM token sequence modules. These stacked layers include (i) a part- of-speech (abbreviated POS) label embedding layer, (ii) a chunk label embedding layer overlying the POS label embedding layer, and (iii) a dependency parent identification and dependency relationship label embedding layer overlying the chunk label embedding layer. In particular, this method, of the technology disclosed includes delivering, via bypass connections, input used by underlying layers together with embedding outputs from the underlying layers to overlaying layers. The method also includes, in the POS label embedding layer, applying a bidirectional LSTM and a POS label classifier to process word embedding vectors representing the words in the input sentence, and producing POS label embedding vectors and POS state vectors for each of the words. Additionally, the method includes, in the chunk label embedding layer, applying a bi-directional LSTM and a chuck label classifier to process at least the word embedding vectors, the POS label embedding vectors and the POS state vectors, and producing chunk label embeddings and chunk state vectors. According to the method, in the dependency parent identification and dependency relationship label embedding layer, a bi-directional LSTM and one or more classifiers are applied to process the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors. This is done to identify dependency parents of each of the words in the sentence and producing dependency relationship labels or label embeddings of relationships between the words and respective potential parents of the words. The method also includes outputting results reflecting the dependency relationship labels or label embeddings for the words in the sentence.

[00156] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00157] In an implementation of the disclosed method, the linguistic hierarchy builds from the words in the sentence, to the parts of speech, to the chunks of the sentence, to dependency links between the words and their dependency parents, to labels on the dependency links.

[00158] The delivering, via the bypass connections, can supply an input vector used by an underlying layer to an overlying layer without modification.

[00159] In some implementations, the method, in the POS label embedding layer, processes n- character-gram embedding vectors that represent the words in the input sentence, in addition to the word embedding vectors. Additionally, the bypass connections deliver the n-character-gram embedding vectors to the chunk label embedding layer and the dependency parent and dependency relationship label embedding layer as input to respective bi-directional LSTMs in those overlaying layers.

[00160] The disclosed method can further include producing, in the POS label embedding layer, POS label probability mass vectors, by exponential normalization of the POS state vectors, and producing the POS label embedding vectors, from the POS label probability mass vectors. Additionally, in the chunk label embedding layer, the method produces chunk label probability mass vectors, by scaling normalization of the chunk label state vectors, and produces the chunk label embedding vectors from the chunk label probability mass vectors. A softmax function that applies exponential normalizing can be used for the scaling normalization. Further, in the dependency parent identification and dependency relationship label embedding layer, the disclosed technology (i) produces parent label probabilit mass vectors by classification and scaling normalization of parent label state vectors produced by the bi-directional LSTM, (ii) produces the parent label embedding vectors from, the parent label probability mass vectors, (iii) produces dependency relationship label probability mass vectors by classification and scaling normalization of the parent label state vectors and the parent label embedding vectors, and (iv) produces the dependency relationship label embedding vectors from the dependency relationship label probability mass vectors,

[00161] Optionally, the dimensionalities of the POS label embedding vectors, the chunk label embedding vectors, and the dependency relationship label embedding vectors can be similar, within +/- ten percent,

[00162] In some implementations, the stacked layers can include a sentence embedding layer, underlying the POS label embedding layer. The sentence embedding layer can include a word embedder and an n-character-gram embedder. Additionally, the method includes, mapping, in the word embedder, the words in the sentence, when recognized, into a word embedding space represented by a word embedding vector. The method also includes, in the n-character-gram embedder, (i) processing character substrings of the word at multiple scales of substring length, (ii) mapping each processed character substring into an intermediate vector representing a position in a character embedding space, and (iii) combining the intermediate vectors for each unique processed character substring to produce character embedding vectors for each of the words. The sentence embedding layer can output vectors from the word embedder and the n- character-gram embedder, whereby a word not previously mapped into the word embedding space is nonetheless represented by the character embedding vector. These components of the sentence embedding layer can be embodied in a word embedder module and an n-character-gram embedder module, as described in the accompanying claims.

[00163] The n-character-gram embedder can combine the intermediate vectors in at least two way s. It can produce element wise averages in the character embedding vector or it can select element wise maximums. The POS label classifier can include a softmax layer or, more generally, an exponential normalizer. These alternatives also apply to the chunk label classifier. |00164] The disclosed technology operates well without a beam search in the POS label embedding layer, the chunk label embedding layer or the dependency parent identification and dependency relationship label embedding layer.

[00165] The dependency parent identification and dependency relationship label embedding layer further includes a dependency parent analyzer, an attention encoder, and a dependency relationship label classifier. The disclosed method applies, in the dependency parent analyzer, a bi-directional LSTM to process the words in the input sentences. This processing of the words can include processing, for each word, word and n-character-gram embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vector to accumulate forward and backward state vectors that represent forward and backward progressions of interactions among the words in the sentence. The method can also include processing, in the attention encoder, the forward and backward state vectors for each respective word in the sentence to encode attention as inner products between embeddings of each respective word and other words in the sentence, after applying a linear transform to the forward and backward state vectors for the word or the other words, whereby weights in the linear transform are trainable. This method also includes applying, in the attention coder, exponential normalization to vectors of the inner products to produce parent label probability mass vectors and projecting the parent label probability mass vectors to produce parent label embedding vectors. In the dependency relationship label classifier and for each respective word in the sentence, the method (i) classifies and normalizes the forward and backward state vectors and the parent label embedding vectors and the parent label embedding vectors, to produce dependency relationship label probability mass vectors, and (ii) projects the dependency relationship label probability mass vectors to produce dependency relationship label embedding vectors.

[00166] In an implementation, the stacked layers or stack of modules further include a semantic relatedness layer, overlying the dependency parent identification and dependency relationship label embedding layer. The semantic relatedness layer includes a relatedness vector calculator and a relatedness classifier. The disclosed method further includes, in the semantic relatedness layer, operating on pairs of first and second sentences already processed through the disclosed method. Further, in the relatedness vector calculator, the disclosed method calculates a sentence-level representation of each of the first and second sentences by (i) applying a bi- directional LSTM to calculate forward and backward state vectors for each of the words in the respective sentences, and (ii) calculating an element-wise maximum of the forward and backward state vectors for each of the respective sentences to calculate an element- wise sentence-level relatedness vector. The method also includes processing the sentence-level relatedness vector to derive a categorical classification of relatedness between the first and second sentences. The method can include reporting the categorical classification or the sentence-level relatedness vector for farther processing.

[00167] The disclosed technology, in the relatedness vector determiner or calculator 112, (i) determines element-wise differences between the first and second sentence-level relatedness vectors, (ii) determines element-wise products between the first and second sentence-level relatedness vectors, and (iii) uses vectors of absolute values of the element-wise differences and of the element- wise products as inputs to the relatedness classifier.

[00168] In some implementations, the stacked layers further include an entailment layer, overlying the semantic relatedness layer. The entailment layer 114 includes an entailment vector determiner or calculator 1141 and an entailment classifier 1142. The entailment vector determiner determines a sentence-level representation of each of the first and second sentences by (i) applying a bi-directional LSTM to determine forward and backward state vectors for each of the words in the respective sentences, and (ii) determines an element- wise maximum of the forward and backward state vectors for each of the respective sentences. The described method further includes (i) determining, in the entailment vector determiner, an element-wise sentence- level entailment vector and (ii) processing the sentence-level entailment vector to categorically classify entailment between the first and second sentences.

[00169] The disclosed method can also include the entailment vector determiner (i) determining element-wise differences between the sentence-level relatedness vectors for the first and second sentences, (ii) determining element-wise products between the sentence-level relatedness vectors for the first and second sentences, and (iii) using vectors of the element-wise differences and the element-wise products as inputs to the relatedness classifier.

[00170] In another implementation, a multi-layer neural network system, running on hardware that processes a sequence of tokens in an input sequence, is described that includes a stacked LSTM token sequence processor, running on hardware, stacked in layers according to an analytical hierarchy. This stack can be embodied in a stack of LSTM token sequence modules. The stacked LSTM includes bypass connections that deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. The stacked lay ers include (i) a first embedding layer, (ii) a second embedding layer overlying the first embedding layer, and (iii) a third embedding layer overlying the second embedding layer. The first embedding layer, implemented as a bi-directional LSTM and a first label classifier, processes token embeddings representing the tokens in the input sequence, and produces first enibeddings and first state vectors of the tokens. The second embedding layer, implemented as a bi-directional LSTM and a second label classifier, processes at least the token embeddings, the first label embeddings and first state vectors, to produce second label embeddings and second state vectors. The third embedding layer, implemented as a bi-directional LSTM, processes at least the token embeddings, the first label embeddings, the second label embeddings and the second state vectors to produce third label embeddings and third state vectors. Components of the three embedding layers can be embodied in first, second and third processing modules (e.g., 102, 104, 106) of the respective layers and first, second and third production modules, as described in the accompanying claims. The multi-layer neural network system also includes an output processor that outputs at least results reflecting the third label embeddings for the tokens in the input sequence.

[00171] This system and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional systems disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00172] A bypass connection supplies an input vector used by an underlying layer to an overlying layer without modification.

[00173] In an implementation of the disclosed multi-layer neural network system, the first embedding layer further processes token decomposition embedding vectors that represent the tokens in the input sequence, in addition to the token embedding vectors. Additionally, the bypass connections deliver the token decomposition embedding vectors to the second embedding layer and the embedding third layer as input to respective bi-directional LSTMs in those overlaying layers.

[00174] In some implementations, the first embedding layer farther produces first label probability mass vectors, by exponential normalization of the first state vectors, and produces the first label embedding vectors, from the first label probability mass vectors. Additionally, the second embedding layer produces second label probability mass vectors, by exponential normalization of the second state vectors, and produces the second label embedding vectors from the second label probability mass vectors. Further, the third embedding layer produces third label probability mass vectors, by exponential normalization of the third state vectors, and produces the third label embedding vectors from the third label probability mass vectors. Moreover, dimensionalities of the first label embedding vectors, the second label embedding vectors, and the third label embedding vectors are similar, within +/- ten percent.

[00175] The technology disclosed can further include a token embedding processor, underlying the first label embedding layer. The token embedding processor includes a token embedder and a decomposed token embedder. The token embedder maps the tokens in the sequence, when recognized, into a token embedding space represented by a token embedding vector. Further, the decomposed token embedder (i) processes token decompositions of the token at multiple scales, (ii) maps each processed token decomposition into an intermediate vector representing a position in a token decomposition embedding space, and (iii) combines the intermediate vectors for each unique processed token decomposition to produce token decomposition embedding vectors for each of the tokens. The token embedding processor also combines results of the token embedder and the decomposed token embedder, whereby a token not previously mapped into the token embedding space is nonetheless represented by the token decomposition embedding vector.

[00176] At least one of the label classifiers can include a softmax layer or, more generally, an exponential normalizer.

[00177] The disclosed technology also operates well without a beam search in the first through third label embedding layers.

[00178] In an implementation, the disclosed multi-layer neural network system further includes a fourth label embedding layer, overlying the third label embedding layer. The fourth label embedding layer can be implemented as a bi-directional LSTM to process at least the token embeddings, the first label embeddings, the second label embeddings, the third label embeddings and the third state vectors to produce fourth label embeddings and fourth state vectors.

[00179] The disclosed technology also includes a fifth label embedding layer, overlying the fourth label embedding layer. The fifth label embedding layer can be implemented as a bi- directional LSTM to process at least the token embeddings, the first label embeddings, the second label embeddings, the third label embeddings, fourth label embeddings and the fourth state vectors to produce fifth label embeddings and fifth state vectors.

[00180] In another implementation, a method is provided that processes tokens in an input sequence using a stacked layer long-short-term-memory (abbreviated LSTM) sentence processor, running on hardware, stacked in layers according to an analytical hierarchy. This stack can be embodied in a stack of LSTM token sequence modules. These stacked layers include (i) a first embedding layer, (ii) a second embedding layer overlying the first embedding layer, and (iii) a third embedding layer overlying the second embedding layer. In particular, this method includes delivering, via bypass connections, input used by underlying layers together with embedding outputs of the underlying layers to overlaying layers. The method also includes the first embedding layer applying a bi-directional LSTM and a first label classifier to process token embeddings representing the tokens in the input sequence, and producing first label embeddings and first state vectors of the tokens. Additionally, the method includes the second embedding layer, applying a bi-directional LSTM and a second label classifier to process at least the token embeddings, the first label embeddings and first state vectors, to produce second label embeddings and second state vectors. According to the method the third embedding layer, applies a bi-directional LSTM, to process at least the token embeddings, the first label embeddings, the second label embeddings and the second state vectors to produce third label embeddings and third state vectors. Further, according to the technology discloses the method includes outputting results reflecting stacked LSTM analysis according to the analytical hierarchy, including the third label embeddings for the tokens in the input sequence.

[00181] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00182] The delivering, via the bypass connections, can supply an input vector used by an underlying layer to an overlying layer without modification.

[00183] In some implementations, the method, in the first embedding layer, processes token decomposition embedding vectors that represent the tokens in the input sequence, in addition to the token embedding vectors. Additionally, the bypass connections further deliver the token decomposition embedding vectors to the second embedding layer and the embedding third layer as input to respective bi-directional LSTMs in those overlaying layers. [00184] The disclosed method can further include the first embedding layer producing first label probability mass vectors, by exponential normalization of the first state vectors, and producing the first label embedding vectors, from the first label probability mass vectors.

Additionally, the second embedding layer produces second label probability mass vectors, by exponential normalization of the second state vectors, and produces the second label embedding vectors from the second label probability mass vectors. The third embedding layer further produces third label probability mass vectors, by exponential normalization of the third state vectors, and produces the third label embedding vectors from the third label probability mass vectors. According to the discloses method, dimensionalities of the first label embedding vectors, the second label embedding vectors, and the third label embedding vectors are similar, within +/- ten percent.

[00185] The method disclosed can also include further invoking a token embedding processor, underlying the first label embedding layer that includes a token embedder and a decomposed token embedder. Further, the method can include, in the token embedder, mapping the tokens in the sequence, when recognized, into a token embedding space represented by a token embedding vector. Additionally, in the decomposed token embedder, the method (i) processes token decompositions of the token at multiple scales, (ii) maps each processed token decomposition into an intermediate vector representing a position in a token decomposition embedding space, and (iii) combines the intermediate vectors for each unique processed token decomposition to produce token decomposition embedding vectors for each of the tokens. The method also combines results of the token embedder and the decomposed token embedder, whereby a token not previously mapped into the token embedding space is nonetheless represented by the token decomposition embedding vector.

[00186] At least one of the label classifiers can include a softmax layer or, more generally, an exponential normalizes

[00187] The disclosed technology also operates well without a beam search in the first through third label embedding layers.

[00188] According to the technology disclosed, the stacked layers include a fourth label embedding layer, overlying the third label embedding layer. The method also includes in the fourth label embedding layer, applying a bi-directional LSTM to process at least the token embeddings, the first label embeddings, the second label embeddings, the third label embeddings and the third state vectors to produce fourth label embeddings and fourth state vectors.

[00189] In another implementation, the stacked layers include a fifth label embedding layer, overlying the fifth label embedding layer. Further, the method includes, in the fifth label embedding layer, applying a bi-directional LSTM to process at least the token embeddings, the first label embeddings, the second label embeddings, the third label embeddings, fourth label embeddings and the fourth state vectors to produce fifth label embeddings and fifth state vectors.

[00190] In another implementation, a method is provided that trains a stacked LSTM sequence processor, running on hardware, stacked in at least three layers according to an analytical hierarchy. Bypass connections deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. The method includes training first, second and third layers by backward propagation using training examples directed to each layer, with regularized pass down to underlying layers during training. Specifically , this training includes (i) training the first layer using first layer training examples, (ii) training the second layer using second layer training examples with regularized pass down training to the first layer, and (iii) training the third layer using third layer training examples with regularized pass down training to the first and second layers. Regularized pass down training is regularized by constraining a training objective function, having a fitness function with at least two

regularization terms. In addition, according to the technology disclosed the two regularization terms regularize by penalizing growth in a magnitude of weights in coefficient matrices applied to the underlying layers and successively regularize all changes in the weights in the coefficient matrices applied to the underlying layers,

[00191] This method and other implementations of the technology disclosed can each optionally include one or more of the following teatures and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00192] The fitness function can be a negative log likelihood based cross-entropy, a mean squared error or a Kullback-Leibler divergence (KL-divergence). Further, according to the technology disclosed the fitness function can be represented by

"∑∑ °S ϊ ⁷ (" ^{> :=: a} \ H" ^y ) " ⁺ - regularization __ terms where, (n) denotes the nth layer of the stacked LSTM, and [ ( (X | f ^ denotes the probability value that the correct label OC is assigned to W _t in the sentence S .

[00193] In some implementations the regularization term that penalizes growth in a magnitude of weights in coefficient matrices applied to the underlying layers is where (m), which is the same layer as n, denotes the layers 1 to m of the stacked LSTM, is a -norm regularization hyperparameter, and applies the squaring operation element wise to elements of a weighting matrix for the

,'ers 1 to m of the stacked LSTM.

ί94] implementation of the disclosed method, the successive regularization term

where (m-1) , which is the same layer as n-1 , denotes the layers 1 to m-1 of the stacked LSTM, S is a successive regularization hyperparameter, ® _(m ... _{] }} denotes layer parameters of one or more underlying layers, 0^ _m _-^ denotes layer parameters of one or more underlying layers persisted in a previous sub-epoch, and

t ^ne squaring operation element wise to elements of a weighting matrix for the layers 1 to m-1 of the stacked LSTM.

[00195] Further, in the disclosed method, the analytical hierarchy in the stacked LSTM can include at least five layers or at least ten layers. Additionally, basement layers that are below the stacked LSTM can be trained separately from the stacked LSTM and can produce input used by a lowest layer of the stacked LSTM. Attic layers that are above the stacked LSTM can also be trained separately from the stacked LSTM and can consume output from an upper most layer of the stacked LSTM. The training method can involve training the five or ten layers in the stack. The basement and attic layers can be trained separately.

[00196] In another implementation, method is provided for conveying intermediate results from an underlying layer to an overlying layer in a neural network stack of bi-directional LSTMs. The neural network stack of bi-directional LSTMs includes layers corresponding to an analytical framework that process a sequence of tokens. Further, the underlying layer produces analytic framework label vectors for each of the tokens. Specifically, this method includes, for the sequence, analyzing the tokens using the underlying layer. The analyzing of the tokens can include (i) applying the bi-directional LSTM to compute forward and backward state vectors for each of the tokens, (ii) applying a classifier to the forward and backward state vectors to embed each of the tokens in an analytic framework label space, as label space vectors that have dimensionality about the same as a number of available analytical framework labels, and (iii) projecting the label space vectors of each token into an extended dimensionality label space, which has dimensionality about the same as dimensionality of the forward and backward states, to produce extended token label vectors. Additionally, the method includes conveying, from the underlying layer to the overlying layer, vectors of the forward state, the backward state, and the extended token label, thereby supplying input needed by the overlying layer to perform its role in the analytical framework for processing the tokens.

[00197] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00198] In some implementations, the disclosed method includes conveying by a bypass to the overlaying layer, vectors received as input by the underlying layer, other than state vectors. The underlying layer can be over two deeper layers. Additionally, the disclosed method can include conveying by bypasses to the overlaying layer, vectors received as input by the two deeper layers and embedded label vectors produced as output by the two deeper layers. This conveying by the bypass can cause the conveyed vectors to be conveyed without modification.

[00199] According to the disclosed method, the number of available analytical framework labels is smaller than the dimensionality of the forward and backward states, thereby forming a dimensionalit bottleneck that reduces overfitting when training the neural network stack of bi- directional LSTMs. In some implementations, the dimensionality can be one-fifth or one tenth or less of the dimensionality of the forward and backward states.

[00200] In another implementation a multi-layer neural network system, running on hardware that processes a sequence of tokens in an input sequence, is described that includes a stacked LSTM token sequence processor, running on hardware, stacked in layers according to an analytical hierarchy. The stacked LSTM sentence processor can be embodied in a stack of LSTM sentence modules. The stacked LSTM includes bypass connections that deliver input to underlying layers together with embedding outputs of the underlying layers to overlaying layers. The stacked layers include (i) a first embedding layer and (ii) a second embedding layer overlying the first embedding layer. The first embedding layer is implemented as a bi-directional LSTM and a first label classifier and processes token embeddings representing the tokens in the input sequence. The first embedding layer also produces analytic framework label vectors for each of the tokens. Further, the bi-directional LSTM computes forward and backward state vectors for each of the tokens. This functionality of the first embedding layer can be embodied in an embedding processing module for processing token embeddings representing the tokens in the input sequence and a label vector production module for producing analytic framework label vectors. Additionally, a classifier applied to the forward and backward state vectors embeds each of the tokens in an analytic framework label space, as label space vectors that have

dimensionality about the same as a number of available analytical framework labels. This functionality of the first embedding layer can be embodied in an output port.

[00201] The first embedding layer also can project the label space vectors of each token into an extended dimensionality label space, which has dimensionality about the same as

dimensionality of the forward and backward states, to produce extended token label vectors. This method also includes the first embedding layer sending to the second embedding layer, vectors of the forward state, the backward state, and the extended token label, thereby supplying input needed by the second embedding layer to perform its role in the analytical framework for processing the tokens.

[00202] This system and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional systems disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00203] In some implementations, the method further includes a bypass to the second embedding lay er that conveys vectors received as input by the first embedding layer, other than state vectors.

[00204] In an implementation of the multi-layer neural network system, the first embedding layer is over two deeper layers. The system farther conveys by bypassing to the second embedding layer, vectors received as input by the two deeper layers and embedded label vectors produced as output by the two deeper layers. A bypass can convey vectors without modification.

[00205] A number of available analytical framework labels can be smaller than the dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bi-directional LSTMs. In some implementations, the dimensionality can be one-fifth or one tenth or less of the dimensionality of the forward and backward states.

[00206] In another implementation, a multi-layer neural network system, running on hardware that processes words in an input sentence, including words not previously mapped into a word embedding space is described that includes a word embedder or embedder module and a substring embedder or embedder module, both of which process the words in the input sentence. The word embedder maps previously recognized words into a word embedding space and identifies previously unrecognized words, to produce word embedding vectors for each of the words. The substring embedder (i) processes character substrings of the word at multiple scales of substring length, (ii) maps each processed character substring into an intermediate vector representing a position in a character embedding space, and (iii) combines the intermediate vectors for each unique processed character substring to produce character embedding vectors for each of the words. The multi-layer neural network system also includes an embedder combiner that reports for use by a farther process or processing layer both the word embedding vectors and the character embedding vectors, whereby a word not previously mapped into the word embedding space is nonetheless represented by the character embedding vector.

[00207] This system and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional systems disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00208] In an implementation of the disclosed multi-layer neural network system, the substring embedder or embedder module (i) combines the intermediate vectors by element-wise averaging of the intermediate vectors for each unique processed character substring or (ii) combines the intermediate vectors by element-wise selection of maximum values from the intermediate vectors for each unique processed character substring.

[00209] In some implementations, the substring embedder or embedder module processes the character substrings using substring lengths, not counting sentinels at beginning and ends of the words, of two characters, three characters, and four characters.

[00210] Dimensionality of the intermediate vectors can equal dimensionality of the word embedding vectors.

[00211] The technology disclosed can also project the intermediate vectors into a space of dimensionality that equals dimensionality of the word embedding vectors.

[00212] Additionally the multi-layer neural network system can include the word embedder mapping previously unrecognized words to a reserved word embedding vector for unknown words.

[00213] In another implementation, a method is provided for preparing words in an input sentence, including words not previously mapped into a word embedding space, for processing by multi-layer neural network system running on hardware. The processing can be performed using a word embedder and a substring embedder, both of which process the words in the input sentence. The word and substring embedders can be embodied in a word embedder module and a string embedder module, respectively. The method includes, in the word embedder, mapping previously recognized words into a word embedding space and identifying previously unrecognized words, to produce word embedding vectors for each of the words. The method also includes, in the substring embedder and for each of the words in the input sentence, (i) processing character substrings of the word at multiple scales of substring length, (ii) mapping each processed character substring into an intermediate vector representing a position in a character embedding space, and (iii) combining the intermediate vectors for each unique processed character substring to produce character embedding vectors for each of the words. Additionally, the method includes outputting the word embedding vectors and the character embedding vectors, whereby a word not previously mapped into the word embedding space is nonetheless represented by the character embedding vector.

[00214] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00215] In some implementations, the substring embedder or embedder module can (i) combine the intermediate vectors by element- wise averaging of the intermediate vectors for each unique processed character substring or (ii) combine the intermediate vectors by element-wise selection of maximum values from the intermediate vectors for each unique processed character substring.

[00216] The disclosed method can include the substring embedder or embedder module processing the character substrings using substring lengths, not counting sentinels at beginning and ends of the words, of two characters, three characters, and four characters.

[00217] A dimensionality of the intermediate vectors can equal dimensionality of the word embedding vectors.

[00218] In an implementation, the disclosed method can include (i) projecting the

intermediate vectors into a space of dimensionality that equals dimensionality of the word embedding vectors, and/or (ii) the word embedder or embedder module mapping previously unrecognized words to a reserved word embedding vector for unknown words.

[00219] In another implementation, a dependency parsing layer component of a neural network device, running on hardware, that processes words in an input sentence, is described. The dependency parsing layer overlies a chunk label embedding layer that produces chunk label embeddings and chunk state vectors. Further, the chunk label embedding layer, in turn, overlies a POS label embedding layer that produces POS label embeddings. The dependency parsing layer component includes a dependency parent layer and a dependency relationship label classifier. In addition, the dependency parent layer includes a bi-directional LSTM and one or more classifiers, that process the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors, to produce parent label probability mass vectors by classification and exponential normalization of parent label state vectors produced by the bidirectional LSTM. The dependency parent layer also produces the parent label embedding vectors from the parent label probability mass vectors. The dependency relationship label classifier produces dependency relationship label probability mass vectors by classification and exponential normalization of the parent label state vec tors and the parent label embedding vectors. Further, the dependency relationship label classifier produces the dependency relationship label embedding vectors from the dependency relationship label probability mass vectors. Dimensionalities of the POS label embedding vectors, the chunk label embedding vectors, and the dependency relationship label embedding vectors are similar, within +/- ten percent. The dependency parsing layer component further includes an output processor that outputs at least the dependency relationship label embedding vectors or dependency relationship labels based thereon.

[00220] Parts of the dependency parsing layer component 108 can be embodied in an embeddings processing module 1084, a mass vector production module 1085, and a parent label vector production module 1086: the embeddings processing module for processing the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors; the a mass vector production module for producing parent label probability mass vectors from parent label state vectors produced by the bi-directional LSTM; and the parent label vector production module for producing the parent label embedding vectors from the parent label probability mass vectors. The dependency relation label classifier can be embodied in a normalizing module and a dependency label vector production module: the normalizing module for scale normalizing the parent label state vectors and the parent label embedding vectors; and the dependency label vector production module for producing the dependency relationship label embedding vectors from the parent relationship label probability mass vectors,

[00221] This component and other implementations of the technology disclosed can each optionally include one or more of the following features and'Or features described in connection with additional components disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations. [00222] In some implementations, the bi-directional LSTM produces forward and backward parent label state vectors for each respective word in the sentence, which represent forward and backward progressions of interactions among the words in the sentence, from which the parent label probability mass vectors are produced. The disclosed dependency parsing layer component 108 of the neural network further includes an attention encoder 1087 that (i) processes the forward and backward state vectors for each respective word in the sentence, (ii) encodes attention as vectors of inner products between embeddings of each respective word and other words in the sentence, with a linear transform applied to the forward and backward state vectors for the word or the other words prior to the inner product, and (iii) produces the parent label embedding vectors from the encoded attention vectors. The attention encoder components can be embodied in an attention coder module 1088 and a parent label vector module 1089 for producing the parent label embedding vectors from encoded attention vectors.

[00223] The linear transform applied prior to the inner product is trainable during training of the dependency parent layer and the dependency relationship classifier.

100224] According to the disclosed dependency parsing layer component (i) a number of available analytical framework labels, over which the dependency relationship probability mass vectors are determined, is smaller than dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs or (ii) the number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one- tenth or less a dimensionality of the forward and backward states, thereby forming a

dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs. In some implementations, the dimensionality can be one- fifth or less of the dimensionality of the forward and backward states,

[00225] In one implementation a dependency parsing layer component of a neural network device, running on hardware, for processing words in an input sentence, is described. The dependency parsing layer overlies a chunk label embedding layer that produces chunk label embeddings and chunk state vectors. The chunk label embedding layer, in turn, overlies a POS label embedding layer that produces POS label embeddings and POS state vectors. The dependency parsing layer component includes a dependency parent layer and a dependency relationship label classifier. In addition, the dependency parent layer includes a dependency parent analyzer, implemented as a bi-directional LSTM, that processes the words in the input sentences. The bi-directional LSTM processes, for each word, word embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vector to accumulate forward and backward state vectors that represent forward and backward progressions of interactions among the words in the sentence. The dependency parent analyzer 1180 components can be embodied in an embedding module or processor 1181 for processing, for each word, word embedding®, the POS label embeddings, the chunk label embeddings, and the chunk state vector and a state vector production module 1182 for producing forward and backward state vectors that represent forward and backward progressions of interactions among the words in the sentence.

[00226] The dependency parent layer also includes an attention encoder for (i) processing the forward and backward state vectors for each respective word in the sentence, (ii) encoding attention to potential dependencies., and (iii) applies scaling normalization to vectors of the inner products to produce parent label probability mass vectors and projects the parent label probability mass vectors to produce parent label embedding vectors. The functional of these components of the attention encoder 1087 can embodied in a normalization module 1184 for applying scaling normalization to produce parent label probability mass vectors and projects the parent label probability mass vectors, and a parent labelling module 1186 for producing parent label embedding vectors.

[00227] Additionally, the dependency relationship label classifier, for each respective word in the sentence, (i) classifies and normalizes the forward and backward state vectors and the parent label embedding vectors and the parent label embedding vectors, to produce dependency relationship label probability mass vectors, and (ii) projects the dependency relationship label probability mass vectors to produce dependency relationship label embedding vectors. The dependency parsing layer component also includes an output processor that outputs at least results reflecting classification labels for the dependency relationship of each word, the dependency relationship label probability mass vectors, or the dependency relationship label embedding vectors. The dependency relationship label classifier 1186 can be embodied in a dependency relationship label vector production module 1187 for producing dependency relationship label probability mass vectors from embedding vectors and the parent label embedding vectors; and in a dependency label vector production module 1188 for producing dependency relationship label embedding vectors from the dependency relationship label probability mass vectors.

[00228] Attention to potential dependencies can be determined as inner products between embeddings of each respective word and other words in the sentence, with a linear transform applied to the forward and backward state vectors for the word or the other words prior to the inner product.

[00229] This component and other implementations of the technology disclosed can each optionally include one or more of the following features and'or features described in connection with additional components disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations,

[00230] The linear transform applied prior to the inner product is trainable during training of the dependency parent layer and the dependency relationship classifier.

[00231 ] In some implementations, a dimensionality bottleneck can be created by restricting a number of available analytical framework labels, as described above, with the benefit of reducing overfitting when training the stack. In alternative implementations, (i) a number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one-fifth or less a dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs or (ii) the number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one- tenth or less a dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs.

[00232] In another implementation, a method is provided for dependency parsing using a neural network system or device, running on hardware, that processes words in an input sentence. A dependency parsing layer overlies a chunk label embedding layer that produces chunk label embeddings and chunk state vectors. The chunk label embedding layer, in turn, overlies a POS label embedding layer that produces POS label embeddings. Further, the dependency parsing layer includes a dependency parent layer and a dependency relationship label classifier. The disclosed method includes, in the dependency parent layer, applying a bidirectional LSTM and one or more classifiers, that process the word embeddings, the POS label embeddings, the chunk label embeddings and the chunk state vectors, to produce parent label probability mass vectors by classification and scaling normalization of parent label state vectors produced by the bi-directional LSTM. The scaling normalization can be implemented using a softmax component that performs exponential normalization. The method also includes producing the parent label embedding vectors from the parent label probability mass vectors. The disclosed method further includes, in the dependency relationship label classifier, (i) producing dependency relationship label probability mass vectors by classification and scaling normalization of the parent label state vectors and the parent label embedding vectors, and (ii) producing the dependency relationship label embedding vectors from the dependency relationship label probability mass vectors. According to the disclosed method, reporting, outputting or persisting at least the dependency relationship label embedding vectors or dependency relationship labels are based thereon.

[00233] Optionally, dimensionalities of the POS label embedding vectors, the chunk label embedding vectors, and the dependency relationship label embedding vectors are similar, within +/- ten percent.

[00234] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00235] In some implementations, the method includes the bi-directional I .S I M producing forward and backward parent label state vectors for each respective word in the sentence, which represent forward and backward progressions of interactions among the words in the sentence, from which the parent label probability mass vectors are produced. The method also includes, in an attention encoder for processing the forward and backward state vectors for each respective word in the sentence, encoding attention to potential dependicies as vectors.

[00236] This can include determining inner products between embeddings of each respective word and other words in the sentence and applying a linear transform applied to the forward and backward state vectors for the word or the other words prior to the inner product, and producing the parent label embedding vectors from the encoded attention vectors,

[00237] The linear transform can be applied prior to the inner product is trainable during training of the dependency parent layer and the dependency relationship classifier.

[00238] According to the disclosed method, a dimensionality bottleneck can be created by restricting a number of available analytical framework labels, as described above, with the benefit of reducing overfitting when training the stack. In alternative implementations, (i) a number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one-fift or less a dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs or (ii) the number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one-tenth or less a dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs. [00239] In another implementation, method is provided that dependency parses using a neural network device, running on hardware, that processes words in an input sentence, A dependency parsing layer overlies a chunk label embedding layer that produces chunk label embeddings and chunk state vectors. The chunk label embedding layer, in turn, overlies a POS label embedding layer that produces POS label embeddings. Further, the dependency parsing layer includes a dependency parent layer and a dependency relationship label classifier. The disclosed method includes, in the dependency parent layer, in a dependency parent analyzer, applying a bidirectional LSTM to process the words in the input sentences. These processes include processing, for each word, word embeddings, the POS label embeddings, the chunk label embeddings, and the chunk state vector to accumulate forward and backward state vectors that represent forward and backward progressions of interactions among the words in the sentence. The disclosed method also includes, in the dependency parent layer, in an attention encoder that processes the forward and backward state vectors for each respective word in the sentence, (i) encoding attention as inner products between embeddings of each respective word and other words in the sentence, with a linear transform applied to the forward and backward state vectors for the word or the other words prior to the inner product and (ii) applying scaling normalization to vectors of the inner products to produce parent label probability mass vectors and projects the parent label probability mass vectors to produce parent label embedding vectors. Further, according to the disclosed method, in the dependency relationship label classifier and for each respective word in the sentence, (i) classifying and normalizing the forward and backward state vectors and the parent label embedding vectors and the parent label embedding vectors, to produce dependency relationship label probability mass vectors, and (ii) projecting the dependency relationship label probability mass vectors to produce dependency relationship label embedding vectors. The disclosed method also includes outputting at least results reflecting classification labels for the dependency relationship of each word, the dependency relationship label probability mass vectors, or the dependency relationship label embedding vectors.

[00240] This method and other implementations of the technology disclosed can each optionally include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this section can readily be combined with sets of base features identified as implementations.

[00241] The linear transform applied prior to the inner product is trainable during training of the dependency parent layer and the dependency relationship classifier. [00242] According to the disclosed method, a dimensionality bottleneck can be created by restricting a number of available analytical framework labels, as described above, with the benefit of reducing overfitting when training the stack. In alternative implementations, (i) a number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one-fifth or less a dimensionality of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs or (ii) the number of available analytical framework labels, over which the dependency relationship probability mass vectors are calculated, is one-ten th or less a dimensionali ty of the forward and backward states, thereby forming a dimensionality bottleneck that reduces overfitting when training the neural network stack of bidirectional LSTMs.

[00243] Other implementations may include a tangible non-transitory computer readable medium impressed with instructions that are combraable with a processor and memory coupled to the processor. The instructions, when executed on a computer device and one or more servers, perform any of the methods described earlier. In yet other implementations, a tangible non- transitory computer readable medium with instructions that are combinable with a processor and memory coupled to the processor carry out the systems described earlier.

[00244] Yet another implementation may include a computing system including at least one server comprising one or more processors and memory, coupled to the processors, containing computer instructions that, when executed on the processors, cause the computing system to perform any of the processes described earlier,

[00245] While the technology disclosed is disclosed by reference to the preferred

embodiments and examples detailed above, it is to be understood that these examples are intended in an illustrative rather than in a limiting sense. It is contemplated that modifications and combinations will readily occur to those skilled in the art, which modifications and combinations will be within the spirit of the in vention and the scope of the following claims.

Computer System

[00246] FIG. 11 is a simplified block diagram of a computer system 1100 that can be used to implement the joint many-task neural network model 100. Computer system 1100 typically includes one or more CPU processors 1120 that communicate with a number of peripheral devices via bus subsystem 1132. These peripheral devices can include a memory subsystem 1112 including, for example, memory devices and a file storage subsystem 1118, user interface input devices 1130, user interface output devices 1124, a network interface subsystem 1122, and a GPU 1126 with multiple GPU processing cores or GPU processors 1128. The input and output devices allow user interaction with computer system 1100. Network interface subsystem 1122 provides an interface to outside networks, including an interface to corresponding interface devices in other computer systems.

[00247] The operations of the joint many- task neural network model 100 are performed by the GPU processing cores 1128, according to some implementations.

[00248] User interface input devices 1130 or clients or client devices can include a keyboard; pointing devices such as a mouse, trackball, touchpad, or graphics tablet; a scanner; a touch screen incorporated into the display; audio input devices such as voice recognition systems and microphones; and other types of input devices. In general, use of the term "input device" is intended to include all possible types of devices and ways to input information into computer system 1100.

[00249] User interface output devices 1124 can include a display subsystem, a printer, a fax machine, or non-visual displays such as audio output devices. The display subsystem can include a cathode ray tube (CRT), a flat-panel device such as a liquid crystal display (LCD), a projection device, or some other mechanism for creating a visible image. The display subsystem can also provide a non-visual display such as audio output devices. In general, use of the term "output device" is intended to include all possible types of devices and ways to output information from computer system 1100 to the user or to another machine or computer system.

[00250] Storage subsystem 1110 stores programming and data constructs that provide the functionality of some or all of the modules and methods described herein. These software modules are generally executed by CPU processors 1120 alone or in combination with other processors like GPU processors 1128.

[00251] Memory subsystem 1112 in the storage subsystem can include a number of memories including a main random access memory (RAM) 1116 for storage of instructions and data during program execution and a read only memory (ROM) 1114 in which fixed instructions are stored, A file storage subsystem 1118 can provide persistent storage for program and data files, and can include a hard disk drive, a floppy disk drive along with associated removable media, a CD- ROM drive, an optical drive, or removable media cartridges. The modules implementing the functionality of certain implementations can be stored by file storage subsystem 1118 or the memory subsystem 1112, or in other machines accessible by the processor.

[00252] Bus subsystem 1132 provides a mechanism for letting the various components and subsystems of computer system 1100 communicate with each other as intended. Although bus subsystem 1132 is shown schematically as a single bus, alternative implementations of the bus subsystem can use multiple busses. In some implementations, an application server (not shown) can be a tramework that aliows the applications of computer system 1100 to run, such as the hardware and/or software, e.g., the operating system.

[00253] Computer system 1100 itself can be of varying types including a personal computer, a portable computer, a workstation, a computer terminal, a network computer, a television, a mainframe, a server farm, a widely-distributed set of loosely networked computers, or any other data processing system or user device. Due to the ever-changing nature of computers and networks, the description of computer system 1100 depicted in FIG. 11 is intended only as a specific example for purposes of illustrating the preferred embodiments of the present invention. Many other configurations of computer system 1100 are possible having more or less components than the computer system depicted in FIG, 11.

[00254] The preceding description is presented to enable the making and use of the technology disclosed. Various modifications to the disclosed implementations will be apparent, and the general principles defined herein may be applied to other implementations and applications without departing from the spirit and scope of the technology disclosed. Thus, the technology disclosed is not intended to be limited to the implementations shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein. The scope of the technology disclosed is defined by the appended claims.

A JOINT MANY- TASK MODEL: GROWING A NEURAI NETWORK FOR MULTIPLE NLP TASKS

Kazuma Hashimoto,* Calming XiongJ Yoshimasa Tsuruoka, Richard Socher

The University of Tokyo

j hassy, t suruoka} @ logos . t . u-tokyo . ac . jp

Salesforce Research

{cxiong, rs ocher } @ sa le s f o rce . com

ABSTRACT

Transfer and multi- task learning have traditionally focused on either a single source-target pair or very few, similar tasks. Ideally, the linguistic levels of morphology, syntax and semantics would benefit each other by being trained in a single model. We introduce such a joint many-task model together with a strategy for successively growing its depth to solve increasingly complex tasks. All layers include shortcut connections to both word representations and lower-level task predictions. We use a simple regul rization term to allow for optimizing all model weights to improve one task's loss without exhibiting catastrophic interference of the other tasks. Our single end-to-end trainable mode! obtains state-of-the-art results on chunking, dependency parsing, semantic relatedness and textual entailment. It also performs competitively on POS tagging. Our dependency parsing layer relies only on a single feed-forward pass and does not require a beam search.

1 INTRODUCTION

The potential for leveraging multiple levels of representation has been demonstrated in a variety of ways in the field of Natural Language Processing (NLP). For example, Part-Of-Speech (POS) tags are used to train syntactic parsers. The parsers are used to improve higher-level tasks, such as natural language inference (Chen et al., 2016), relation classification (Socher et al., 2012), sentiment analysis (Socher et al, 2013; Tai et a!., 2015a), or machine translation (Eriguchi et al., 2016). However, higher level tasks are not usually able to improve lower level tasks, often because systems are pipelines and not trained end-to-end.

In deep learning, unsupervised word vectors are useful representations and often used to initialize recurrent neural networks for subsequent iasks (Pennington et al., 2014). However, not being jointly trained, deep NLP models have yet shown benefits from predicting many ( > 4) increasingly complex linguistic tasks each at a successively deeper layer. Instead, existing models are often designed to predict different tasks either entirely separately or at the same depth (Co!lobert et al, 203 1), ignoring linguistic hierarchies.

We introduce a Joint Many-Task (JMT) model, outlined in Fig. 1, which predicts increasingly complex NLP tasks at successively deeper layers. Unlike traditional NLP pipeline systems, our single JMT model can be trained end-to-end for POS tagging, chunking, dependency parsing, semantic relatedness, and textual entailment. We propose an adaptive training and regularization strategy to grow this model in its depth. With the help of this strategy we avoid catastrophic interference between tasks, and instead show that both lower and higher level iasks benefit from the joint training. Our model is influenced by the observation of S0gaard & Goldberg (2016) who showed that predicting two different tasks is more accurate when performed in different layers than in the same layer (CoUoberi et ai., 2011).

^* Work was done while the first author was an intern ai Salesforce Research.

^Corresponding author.

1 Under review as a conference paper at ICLR 2017

Figure 1: Overview of the joint many-task model predicting different linguistic outputs at successively deeper layers

2 T HE JOINT MAN Y- TA SK MOD EL

In this section, we assume that the model is trained and describe its inference procedure. We begin at the lowest level and work our way to higher layers and more complex tasks.

2. 1 WORD REPRESENTATIONS

For each word w _t in the input sentence s of length L, we construct a representation by concatenating a word and a character embedding.

Word em eddings: We use Skip-gram (Mikolov et al., 2033) to train a word embedding matrix, which will be shared across all of the tasks. The words which are not included in the vocabulary are mapped to a special UNK token.

Character ra-gram embeddings: Character ??,-gram embeddings are learned using the same skip- gram objective function as the word vectors. We construct the vocabulary of the character ??,-grams in the training data and assign an embedding for each character ra-gram. The final character embedding is the average of the unique character -re-gram embeddings of a word ti> _t . ¹ For example, the character n-grams (n = 1, 2, 3) of the word "Cat" are {C, a, i, #BEGIN#C, Ca, at, t#END#, #BEGlN#Ca, Cat, at#END#|, where "#BEGIN#" and "#END#" represent the beginning and the end of each word, respectively. The use of the character n-gram embeddings efficiently provides morphological features and information about unknown words. The training procedure for character n-grams is described in Section 3.1. Each word is subsequently represented as x _t , the concatenation of its corresponding word and character vectors.

2.2 WORD- LEVEL TASK: POS TAGGING

The first layer of the model is a bi-directional LSTM (Graves & Schniidhuber, 2005; Hochreiter & Schniidhuber, 3997) whose hidden states are used to predict POS tags. We use the following Long Short-Term Memory (LSTM) units for the forward direction:

i _t ^ σ (W _t g _t + ) . f _t - ---- σ ( /<¾ + bf ) , o _{ = σ (W ₀ g _t + b ₀ ) , lit— tanh (W _u g _t + b _u ) , _t - ---- i _t u _t f _t Θ <¾_ i , h _t = o _{ © tanh (<¾) ,

Wieting et al. (2016) used a nonlineanty, but vve have observed that the simple averaging also works well.

2 Under review as a conference paper at 1CLR 2017

Figure 2: Overview of the POS tagging and chunking tasks in the first and second layers of the JMT model. where we define the first layer g _t as g _t = \ h _t . _~ i; x+j , i.e. the concatenation of the previous hidden state and the word representation of w _t . The backward pass is expanded in the same way, but a different set of weights are used.

For predicting the POS tag of w _t , we use the concatenation of the forward and backward states in a one-layer bi-LSTM layer corresponding to the t-th word: h _t = [ h _t : h _t ] . Then each ht ( 1 < t < L) is fed into a standard softmax classifier with a single ReLU layer which outputs the probability vector xr ¹ ' for each of the POS tags.

2.3 WORD-LEVEL TASK: CHUNKING

Chunking is also a word-ievei classification task which assigns a chunking tag (B-NP, I -VP, etc.) for each word. The tag specifies the region of major phrases (or chunks) in the sentence.

2.4 SYNTACTIC TASK: DEPENDENCY PARSING

Dependency parsing identifies syntactic relationships (such as an adjective modifying a noun) between pairs of words in a sentence. We use the third bi-LSTM layer on top of the POS and chunking layers to classify relationships between all pairs of words. The input vector for the LSTM includes hidden states, word representations, and the label embeddings for the two previous tasks: i ^' 3) r , i ^' 3) , (2) ,· (nos) (chk) . , , . , , , . . . .. gl = |ft _t _ _i ; hi ; x-t'. {yt ~r % ) i, where we computed the chunking vector its a similar fashion as the POS vector in Eq. (2). The POS and chunking tags are commonly used to improve dependency parsing (Attardi & DellOrletta, 2008).

Like a sequential labeling task, we simply predict the parent node (head) for each word in the sentence. Then a dependency label is predicted for each of the child-parent node pairs. To predict the parent node of the t-th word w _t , we define a matching function between w _t and the candidates of the parent node as m (t, j) Wdh ', where We. is a parameter matrix. For the root, we

3 Under review as a conference paper at ICLR 2017

Figure 3: Overview of dependency parsing its the third layer of the JMT model

Figure 4: Overview of the semantic tasks in the top layers of the JMT model. define <¾V ^' + _] ~ ^{r as 3} parameterized vector. To compute the probability that w _j , (or the root node) is the parent of vj _t , the scores are normalized:

exp (m (t, j))

(3) where L is the sentence length.

Next, the dependency labels are predicted using \h _t : h _j \ as input to a standard softmax classifier with a si ngle ReLU layer. At test time, we greedily select the parent node and the dependency label for each word in the sentence. ² At training time, we use the gold child-parent pairs to trai n the label predictor.

2.5 SEMANTIC TASK: SEMANTIC RELATEDNESS

The next two tasks model the semantic relationships between two input sentences. The first task measures the semantic relatedness between two sentences. The output is a real -valued relatedness score for the input sentence pair. The second task is a textual entailment task, which requires one to determine whether a premise sentence entails a hypothesis sentence. There are typically three classes: entai l, contradict ion, and neutral.

The two semantic tasks are closely related to each other. If the semantic relatedness between two sentences is very low, they are unlikely to entail each other. Based on this intuition and to make use of the information from lower layers, we use the fourth and fifth bi-LSTM layer for the relatedness and entailment task, respectively.

^■ "This method current!)' assumes thai each word has only one parent node, but it can be expanded to handle multiple parent nodes, which leads to cyclic graphs.

4 6(

Under review as a conference paper at 1CLR 2017

Now it is required to obtain the sentence-level representation rather than the word-level represen- tation h _t ' used in the first three tasks. We compute the sentence-level representation ¾ as the element-wise maximum values across ail of the word-level representations in the fourth layer:

¾< ⁴⁾ - Diax fhf , - . . , ) - (4)

To model the semantic reiaiedness between s and s', we follow Tai et al. (2015b). The feature vector for representing the semantic re!atedness is computed as follows:

di | ¾ ⁴⁾ --- ¾iT ⁾ ; ^{/ 4¾} © ] ₁ (5)

tence pa r.

2.6 SEMANTIC TASK: TEXTUAL ENTAILMENT

For entailment cl ssification between two sentences, we also use the max-poo!ing technique as in the semantic relatedness task. To classify the premise-hypothesis pair into one of the three classes, we compute the feature vector d ₂ (s, *') as in Eq. (5) except that we do not use the absolute values of the element-wise subtraction, because we need to identify which is the premise (or hypothesis). Then <¾(s, *') is fed into a standard softmax classifier.

To make use of the output from the relatedness layer directly, we use the label embeddings for the relatedness task. More concretely, we compute the class label embeddings for the semantic relatedness task similar to Eq. (2). The final feature vectors thai are concatenated and fed into the entailment classifier are the weighted relatedness label embedding and the feature vector d _' jXs, s').-' We use three Maxout hidden layers before the classifier.

3 TRAINING THE JMT MODEL

The model is trained jointly over all data sets. During each epoch, the optimization iterates over each full training dataset in the same order the corresponding tasks are described in the modeling section.

3.1 PRE-TRAINING WORD REPRESENTATIONS

We pre -train word embeddings using the Skip-gram model with negative sampling (Mikolov et al., 2013). We also pre -train the character n-gram embeddings using Skip-gram. The only difference is that each input word embedding in the Skip-gram model is replaced with its corresponding average embedding of the character n-gram embeddings described in Section 2.1. These embeddings are fine-tuned during the training of our JMT model. We denote the embedding parameters as 9 _e .

3.2 TRAINING THE POS LAYER

Lei (?pos = (Wpos _> &POS ₎ 0 _e ) denote the set of model parameters associated with the POS layer, where IVpos is the set of the weight matrices in the first bi-LSTM and the classifier, and 6pos is the set of the bias vectors. The objective function to optimize ijpos is defined as follows:

-∑∑ ³ °gP W ^1} - + AllWpos f + δ\ - θ> , ₍₆₎ s t

where p(y) ^1! ------ a _Wt \h^) is the probability value that the correct label a is assigned to w _t in the sentence s, Ajj pos !! ² is the L2-norm regularization term, and λ is a hyperparameter.

""This modification does not affect the LSTM iransiiions, and thus it is still possible to add other single- sentence-level tasks on top of our model.

5 Under review as a conference paper at ICLR 2017

We call the second regularization term δ\\θ _Β — 6 jj ² a successive regularization term. The successive regularization is based on the idea that we do not want the model to forget the information learned for the other tasks. In the case of POS tagging, the regularization is applied to 9 _e , and ff _e is the embedding parameter after training the final task in the top-most layer at the previous training epoch. δ is a hyperparameter.

3.3 TRAINING THE CHUNKING LAYER

The objective function is defined as follows:

- - 0'pos ii ² > (7)

which is similar to that of the POS layer, and (?chunk is defined as (Wd _lurl k, &chunk> E os, 0 _e )» where Wchunk and debunk ^are the weight and bias parameters including those in 0pos _> and Epos is the set of the POS label embeddings. * _pOS is the one after training the POS layer at the current training epoch.

3.4 TRAINING THE DEPENDENCY LAYER

The objective function is defined as follows:

- ll ² , (8)

where p(a \h^ ) is the probability value assigned to the correct parent node a for w _t , and pi ;3| ¾. _j ^' h _a ' ) is the probability value assigned to the correct dependency label β for the child- parent pair (w a) , 0 _dep is defined as (Wd _ep , &de _P , Wd, r, Epos, Eomnk, & _e ) _> where Wd _ep and 6 _dep are the weight and bias parameters including those in 0 _C hunk _> and ¾hunk is the set of the chunking label embeddings.

3.5 TRAINING THE RELATEDNESS LAYER

Following Tai et ai. (2015b), the objective function is defined as follows:

3.6 TRAINING THE ENTAILMENT LAYER

The objective function is defined as follows:

-∑ ^k >gP(¾¾ = - ¾ _e! |! ² , ₍₁₀₎ where p(y) , h ) is the probability value that the correct label a is assigned to the premise-hy s, s'} . 0 _ent is defined as (VV _ent , b _ent , Epos, i¾mnk _> 0 _e )-

4 RELATED WORK

Many deep learning approaches have proven to be effective in a variety of NLP tasks and are becoming more and more complex. They are typically designed to handle single tasks, or some of them are designed as general- purpose models (Kumar et al., 2016; Sutskever et al., 2014) but applied to different tasks independently.

6 6

Under review as a conference paper at ICLR 201';

For handling multiple NLP tasks, multi-task learning models with deep neural networks have been proposed (Collobert et ai., 201 1 ; Luong et al,, 2016), and more recently Sogaard & Goldberg (2016) have suggested that using different layers for different tasks is more effective than using the same layer in jointly learning closely-related tasks, such as POS tagging and chunking. However, the number of tasks was limited or they have very similar lask settings like word-level Sagging, and it was not clear how lower-level tasks could be also improved by combining higher-level tasks.

In the field of computer vision, some transfer and multi-task learning approached have also been proposed (Li & Hoierrt, 2016; Misra et al ., 2016). For example, Misra et ai. (2016) proposed a multi-task learning model to handle different tasks. However, they assume that each data sample has annotations for the different tasks, and do not explicitly consider task hierarchies.

Recently, Rusu et al. (2016) have proposed a progressive neural network model to handle multiple reinforcement tasks, such as Atari games. Like our JMT model, their model is also successively trained according to different tasks using different layers called columns in their paper. Its their model, once the first task is completed, the model parameters for the first task are fixed, and then the second task is handled by adding new model parameters. Therefore, accuracy of the previously trained tasks is never improved. In NLP tasks, multi-task learning has a potential to improve not only higher-level tasks, but also lower-level tasks. Rather than fixing the pre-trained model parameters, our successive regularization allows our model to continuously train the lower-level tasks without significant accuracy drops.

5 EXPERIMENTAL SETTINGS

5.1 DATASETS

POS tagging: To train the POS tagging layer, we used the Wall Street Journal (WSJ) portion of Penn Treebank, and followed the standard split for the training (Section 0- 18), development (Section 19- 21 ), and test (Section 22-24) sets. The evaluation metric is the word-level accuracy.

Chunking: For chunking, we also used the WSJ corpus, and followed the standard split for the training (Section 15-18) and test (Section 20) sets as in the CoNLL 2000 shared task. We used Section 19 as the development set, following Sogaard & Goldberg (2016), and employed the IOBES tagging scheme. The evaluation metric is the Fl score defined in the shared task.

Dependency parsing: We also used the WSJ corpus for dependency parsing, and followed the standard split for the training (Section 2-21), development (Section 22), and test (Section 23) sets. We converted the treebank data to Stanford style dependencies using the version 3.3.0 of the Stanford converter. The evaluation metrics are the Unlabeled Attachment Score (UAS) and the Labeled Attachment Score (LAS), and punctuations are excluded for the evaluation.

Semantic relatedness: For the semantic relatedness task, we used the SICK dataset (Marelli et al ., 2014), and followed the standard split for the training (SICK_train.txt), development (SICK_trial . txt), and test (SICK_test_annotated . txt) sets. The evaluation metric is the Mean Squared Error (MSB) between the gold and predicted scores.

Textual entailment: For textual entailment, we also used the SICK dataset and exactly the same data split as the semantic relatedness dataset. The evaluation metric is the accuracy.

5.2 TRAINING DETAILS

Pre-training embeddings: We used the word2vec toolkit to pre-train the word embeddings.

We created our training corpus by selecting lowercased English Wikipedia text and obtained 100- dimensional Skip-gram word embeddings trained with the context wi ndow size 1, the negative sam pling method (15 negative samples), and the sub- sampling method (10 ^{~ 5} of the sub-sampling coefficient). ⁴ We also pre-trained the character rt-gram embeddings using the same parameter settings with the case-sensitive Wikipedia text. We trained the character n-gram embeddings for n = 1, 2. 3, 4 in the pre-training step.

Ti is empirically known that such a small window size in leads to better results on syntactic tasks than large window sizes. Moreover, we have found thai such word embeddings work well even on the semantic tasks. Under review as a conference paper at ICLR 201'/

Embedding initialization: We used the pre-trained word embeddings io initialize the word embeddings, and the word vocabulary was built based on the training data of the five tasks. All words in the training data were included in the word vocabulary, and we employed the word-dropout method (Kiperwasser & Goldberg, 2016) to train the word embedding for the unknown words. We also built the character n-gram vocabulary for n ----- 2, 3, 4, following Wieting et at. (2016), and the character ra-gram embeddings were initialized with the pre-irained embeddings. All of the label embeddings were initialized with uniform random values in [— /6/(dim + C), /6/(dim + C)], where dim = 100 is the dimensionality of the label embeddings and C is the number of labels.

Weight initialization: The dimensionality of the hidden layers in the bi-LSTMs was set io 100. We initialized all of the softmax parameters and bias vectors, except for the forget biases in ihe LSTMs, with zeros, and the weight matrix W _< j and the root node vector r for dependency parsing were also initialized with zeros. All of ihe forget biases were initialized with ones. The other weight matrices were initialized with uniform random values in [— 6/(row + col), / G/{r<jw + cot)], where row and col are the number of rows and columns of the matrices, respectively.

Optimization: At each epoch, we trained our model in the order of POS tagging, chunking, dependency parsing, semantic re!atedness, and textual entailment. We used mini -batch stochastic gradient decent to train our model. The mini-batch size was set to 25 for POS tagging, chunking, and the SICK tasks, and 15 for dependency parsing. We used a gradient clipping strategy with growing clipping values for the different tasks; concretely, we employed ihe simple function: min(3.0, depth), where depth is the number of bi-LSTM layers involved in each task, and 3.0 is the maximum value. The learning rate at the fc-th epoch was set to V- _Q II-^— - v, where ε is the initial learning rate, and p is the hyperparameter to decrease the learning rate. We set ε to 1.0 and p to 0.3. At each epoch, the same learning rate was shared across all of the tasks.

egularization: We set the regularization coefficient to 10 ^~ ° for the LSTM weight matrices, 10 ^~ ° for the weight matrices in ihe classifiers, and 10 ^~ '' for the successive regularization term excluding the classifier parameters of the lower-level tasks, respectively. The successive regularization coefficient for the classifier parameters was set to 10 ^"2 . We also used dropout (Hinton et ai., 2012). The dropout rate was set to 0.2 for the vertical connections in the multi-layer bi-LSTMs (Pham et ai., 2014), the word representations and the label embeddings of the entailment layer, and the classifier of the POS tagging, chunking, dependency parsing, and entailment. A different dropout rate of 0.4 was used for the word representations and the label embeddings of the POS, chunking, and dependency layers, and the classifier of the re!atedness layer.

6 RESULTS AN D D ISCU SS ION

6.1 S UMMARY OF MULTI-TASK RESULTS

Table 1 shows our results of the test set results on the five different tasks. ⁵ The column "Single" shows the results of handling each task, separately using single-layer bi-LSTMs, and the column "JMT _a ii" shows the results of our JMT model. The single task settings only use the annotations of their own tasks. For example, when treating dependency parsing as a single task, the POS and chunking tags are not used. We can see that all results of the five different tasks are improved in our JMT model, which shows that our JMT model can handle the five different tasks in a single model. Our JMT model allows us to access arbitrary i formation learned from the different tasks. If we want to use the model just as a POS tagger, we can use the output from the first bi-LSTM layer. The output can be the weighted POS label embeddings as well as the discrete POS tags.

Table i also shows the results of three subsets of the different tasks. For example, in the case of "JMTABC". only the first three layers of the bi-LSTMs are used to handle the three tasks. In the case of "JMTDE" _> ^on *y the top two layers are used just as a two-layer bi-LSTM by omitting all information from the first three layers. The results of the closely-related tasks show that our JMT model improves not only the high-level tasks, but also the low-level tasks.

³ The development and test sentences of the chunking dalasel are included in the dependency parsing dataset, although our model does not explicitly use the chunking annotations of the development and test data. In such cases, we show the results in parentheses.

8 Under review as a conference paper at ICLR 2017

Single JMT _a ]i JMTAB JMTABC JMT _DE

A POS 97.45 97.55 97.52 97.54 n/a

B Chunking 95.02 (97.12) 95.77 (97.28) n/a

Dependency UAS 93,35 94.67 n/a 94.71 n/a

C

Dependency LAS 91.42 92.90 n/a 92,92 n/a

D Relatedness 0.247 0.233 n/a n/a 0.238

E Entailment 81 .8 86.2 n/a n/a 86.8

Table 1 : Test set results for the five tasks. In the relatedness task, the lower scores are better.

Method Acc.

97,55 Method F 1

Ling at al. (2015) 97.78 95.77

Kumar et aL (2016) 97.56 S0gaard & Goldberg (2016) 95.56

M_ & Hovy (20! 6) 97.55 S-zuki & isozak: (2008) 15

S0gaard (2011) 97.50 Coiiobsrt et al. (2011) 94.32

Coiiobsrt et al. (2011) 97 29 Kudo & MatsLimoto (2001) 93.91

Tsunioka et al. (2011) 97 8 Tsumofca et al. (2011) 93.81

Toutat!ova st al. (2003) 97 27

Table 3: Chunking results.

Table 2: POS taggi g results. Table 4: Dependency results.

Method MSB Method Acc.

aii 86.2

0.238

Zhou et ai. (2016) 0,243 Yin et al. (2016) 86,2

Tai et ai. (2015b) 0,253 Lai & Hockenmaie; (2014) 84,6

Table 5: Semantic relatedness results. Table 6: Textual entailment results.

6.2 COMPARISON WITH PUBLISHED RESULTS

tagging: Table 2 shows the results of POS tagging, and our JMT model achieves the score close to the state-of-the-art results. The best result to date has been achieved by Ling et al (2015), which uses character-based LSTMs. Incorporating the character- based encoders into our JMT model would be an interesting direction, but we have shown that the simple pre-trained character re-gram embeddings lead to the promising result.

Chunking: Table 3 shows the results of chunking, and our JMT model achieves the state-of-the-art result. S0gaard & Goldberg (2016) proposed to jointly learn POS tagging and chunking in different layers, but they only showed improvement for chunking. By contrast, our results show that the low-level tasks are also i mproved by the joint learning.

Dependency parsing: Table 4 shows the results of dependency parsing by using only the WSJ corpus in terms of the dependency annotations, and our JMT model achieves the state-of-the-art result. ⁶ It is notable that our simple greedy dependency parser outperforms the previous state-of- the-art result which is based on beam search with global information. The result suggests that the bi-LSTMs efficiently capture global information necessary for dependency parsing. Moreover, our single task result already achieves high accuracy without the POS and chunking information.

Semantic relatedness: Table 5 shows the results of the semantic relatedness task, and our JMT model achieves the state-of-the-art result. The result of "JMTDE" is already better than the previous state-of-the-art results. Both of Zhou et al. (2016) and Tai et al. (2015b) explicitly used syntactic tree structures, and Zhou et al. (2016) relied on attention mechanisms. However, our method uses the simple max-pooling strategy, which suggests that it is worth investigating such simple methods before developing complex methods for simple tasks. Currently, our JMT model does not explicitly use the learned dependency structures, and thus the explicit use of the output from the dependency layer should be an interesting direction of future work.

°C ^' hoe & Chainiak (2016) employed the in-training technique to expand the training data w automatically-generated 400,000 trees in addition to the WSJ corpus, and reported 95.9 UAS and 94.1 LAS

9 Under review as a conference paper at ICLR 2017

JMT _a!! w/o SR w/o SC w/o LE w/o SC&LE Single Single*

POS 97.88 97.85 97.79 97.85 97.87 97.52

Chunking 97.59 97.13 97.08 97.40 97.33 95.65 96.08

Dependency UAS 94.51 94.46 94.52 94.09 94,04 93.38 93,88

Dependency LAS 92.60 92.57 92,62 92.14 92.03 91.37 91.83

Relatedness 0.236 0.239 0.698 0.261 0.765 0.239 0.665

Entailment 84.6 84.2 75,0 81.6 71.2 83.8 66.4

Table 7: Development set results of JMT _a j] with different joint-training strategies. The use of the Successive Regularization (SR), Shortcut Connections (SC), and Label Embeddings (LE) are considered. For comparison, the single task results are also shown.

Textual eoiailmeot: Table 6 shows the results of textual entailment, and our JMT model achieves the state-of-the-art result. ⁷ The previous state-of-the-art result in Yin et al. (2016) relied on attention mechanisms and dataset-specific data pre-processing and features. Again, our simple max-pooling strategy achieves the state-of-the-art result boosted by the joint training. These results show the importance of jointly handling related tasks.

6.3 ANALYSIS ON MULTI-TASK LEARNING ARCHITECTURES

Here, we first investigate the effects of using deeper layers for the five different single tasks. We then show the effectiveness of our proposed model and training strategy: the successive regularization, the shortcut connections of the word representations, the embeddings of the output labels, the character n-gram embeddings, and the use of the different layers for the different tasks. All of the results shown in this section are the development set results.

Depth: The single task settings shown in Table 1 are obtained by using single layer bi-LSTMs, but in our JMT model, the higher-level tasks use successively deeper layers. To investigate the gap between the different number of the layers for each task, we also show the results of using multilayer bi-LSTMs for the single task settings, in the column of "Single*" in Table 7, More concretely, we use the same number of the layers with our JMT model; for example, three layers are used for dependency parsing, and five layers are used for textual entailment. As shown in these results, deeper layers do not always lead to better results, and the joint learning is more important than making the models complex for only single tasks.

Successive regularization In Table 7, the column of "w/o SR" shows the results of omitting the successive regularization terms described in Section 3. We can see that the accuracy of chunking is improved by the successive regularization, while other results are not affected so much. The chunking dataset used here is relatively small compared with other low-level tasks, POS tagging and dependency parsing. Thus, these results suggest thai the successive regularization is effective when dataset sizes are unbalanced.

Shortcut connections: Our JMT model feeds the word representations into all of the bi-LSTM layers, which is called the shortcut connection. Table 7 shows the results of "JMT with and without the shortcut connections. The results without the shortcut connections are shown in the column of "w/o SC". These results clearly show that the importance of the shortcut connections in our JMT model, and in particular, the semantic tasks in the higher layers strongly rely on the shortcut connections. That is, simply stacking the LSTM layers is not sufficient to handle a variety of NLP tasks in a single model. In Appendix A, we show how the shared word representations change according to each task (or layer).

Output label embeddings: Table 7 also shows the results without using the output labels of the POS, chunking, and relatedness layers, in the column of "w/o LE". These results show that the explicit use of the output information from the classifiers of the lower layers is important in our JMT model. The results in the last column of "w/o SC&LE" are the ones without both of the shortcut connections and the label embeddings.

' The result of "JMT _» n" is slightly worse than that of "JMT _DE ", but the difference is not significant because the training data is small.

10 _l

Under review as a conference paper at ICLR 2017

Single W&C Only word JMTABC w/o S&E Ali-3

POS 97.52 96.26 POS 97.90 97.87 97.62

Chunking 95.65 94.92 Chunking 97.80 97.41 96.52

Dependency UAS 93.38 92,90 Dependency UAS 94,52 94.13 93,59

Dependency LAS 91.37 90.44 Dependency LAS 92.61 92.16 91.47

Table 8: Development set results for the Table 9: Development set results of three different three single tasks with and without the multi-task learning strategies for the three tasks, character n-gram embeddings.

Character n-gram embeddings: Table 8 shows the results for ihe three single tasks, POS tagging, chunking, and dependency parsing, with and without the pre-trained character rc-gram embeddings. The column of "C&W" corresponds to using both of the word and character n-gram embeddings, and that of "Only word" corresponds to using only the word embeddings. These results clearly show that jointly using the pre-trained word and character n-graoi embeddings is helpful in improving the results. The pre-training of the character n-gram embeddings is also effective; for example, without the pre-training, the POS accuracy drops from 97.52% to 97.38% and the chunking accuracy drops from 95.65% to 95.14%, but they are still better than those of using word2 vec embeddings alone.

Different layers for differeni tasks: Table 9 shows the results for ihe three tasks of our "JMTABC" setting and that of not using the shortcut connections and ihe label embeddings as in Table 7. In addition, in the column of "All-3", we show ihe results of using the highest (i.e., the third) layer for all of the three tasks without any shortcut connections and label embeddings, and thus the two settings "w/o SC&LE" and "All-3" require exactly the same number of the model parameters. The results show that using the same layers for the three differeni tasks hampers ihe effectiveness of our JMT model, and the design of the model is much more important than the number of ihe model parameters.

7 CONCLUS ION

We presented a joint many-task model io handle a variety of NLP tasks with growing depth of layers in a single end-to-end deep model. Our model is successively trained by considering linguistic hierarchies, directly connecting word representations to all layers, explicitly using predictions in lower tasks, and applying successive regularization. In our experiments on five different types of NLP tasks, our single model achieves ihe state-of-the-art results on chunking, dependency parsing, semantic relatedness, and textual entailment.

ACKNOWLEDGMENTS

We ihank the Salesforce Research team members for their fruitful comments and discussions.

REFERENCES

Chris Aiberti, David Weiss, Greg Coppola, and Slav Petrov. Improved Transition-Based Parsing and Tagging with Neural N etworks. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1354- 1359, 2015.

Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov, and Michael Collins. Globally Normalized Transition-Based Neural Networks. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2442-2452, 2016.

Giuseppe Attardi and Felice DellOrletta. Chunking and Dependency Parsing. In Proceedings of LREC 2008 Workshop on Partial Parsing, 2008.

Bernd Bonnet. Top Accuracy and Fast Dependency Parsing is not a Contradiction. In Proceedings of the 23rd international Con ference on Computational Linguistics, pp. 89-97, 2010.

11 6 /

Under review as a conference paper at ICLR 2017

Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, and Hui Jiang. Enhancing and Combining Sequential and Tree LSTM for Natural Language Inference. CoRR, abs/1609.06038, 2016.

Do Kook Choe and Eugene Charniak. Parsing as Language Modeling. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 2331-2336, 2016.

Ronan Collobert, Jason Weston, Leon Bottou, Michael Kar!en nad Kora Kavukcuoglu, and Pavel Kuksa. Natural Language Processing (Almost) from Scratch. Journal of Machine Learning Research, 12:2493-2537, 2011.

Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A. Smith. Transition- Based Dependency Parsing with Stack Long Short-Term Memory. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 334-343, 2015.

Akiko Eriguchi, Kazuma Hashimoto, and Yoshimasa Tsuruoka. Tree-to-Sequence Attentional Neural Machine Translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 823-833, 2016.

Ian J. Goodfellow, David Warde -Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Maxout Networks. In Proceedings of The 30th Internationa! Conference on Machine Learning, pp. 1319- 1327, 2013.

Alex Graves and Jurgen Schmidhuber. Framewise Phoneme Classification with Bidirectional LSTM ^" and Other Neural Network Architectures. Neural Networks, 18(5):602-610, 2005.

Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, liya Sutskever, and Ruslan Salakhut- dinov. Improving neural networks by preventing co-adaptation of feature detectors. CoRR, abs/1207.0580. 2012.

Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):

1735-1780, 1997.

Eliyahu Kiperwasser and Yoav Goldberg. Easy-First Dependency Parsing with Hierarchical Tree LSTMs. Transactions of the Association for Computational Linguistics, 4:445-461, 2016.

Taku Kudo and Yuji Matsumoto. Chunking with Support Vector Machines. In Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics, 2001.

Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit !yyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask Me Anything: Dynamic Memory Networks for Natural Language Processing. In Proceedings of The 33rd International Conference on Machine Learning, pp. 1378-1387, 2016.

Alice Lai and Julia Hockenmaier. Illinois-LH: A Denotational and Distributional Approach to Semantics. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pp. 329-334, 2014.

Zhizhong Li and Derek Hoiem. Learning without Forgetting. CoRR, abs/1606.09282, 2016.

Wang Ling, Chris Dyer, Alan W Black, Isabel Trancoso, Ramon Fermandez, Silvio Amir, Luis Marujo, and Tiago Luis. Finding Function in Form: Compositional Character Models for Open Vocabulary Word Representation. In Proceedings of the 20/5 Conference on Empirical Methods in Natural Language Processing, pp. 1520-1530, 2015.

Minh-Thang Luong, liya Sutskever, Quoc V. Le, Oriol Vinyals, and Lukasz Kaiser. Multi- task Sequence to Sequence Learning. In Proceedings of the 4th International Conference on Learning Representations, 2016.

Xuezhe Ma and Eduard Hovy. End-to-end Sequence Labeling via Bi-directional LSTM-CNNs- CRF. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1064-1074, 2016.

12 Under review as a conference paper at ICLR 2017

Marco Marelii, Luisa Beniivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zampareili, SemEval-2014 Task 1: Evaluation of Compositional Distributional Semantic Models on Full Sentences through Semantic Relatedness and Textual Entailment. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pp. 1-8, 2014.

Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed Representations of Words and Phrases and their Composiiionality. In Advances in Neural Information Processing Systems 26, pp. 311 1-31 19. 2013.

Ishan Misra, Abhinav Shrivastava, Abhinav Gupia, and Martial Hebert. Cross-stitch Networks for Multi-task. Learning. CoBB, abs/l604.03539, ^~ 2016.

Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global Vectors for Word Representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, pp. 1532-1543, 2014.

Vu Pham, Theodore B!ucbe, Christopher Kermorvant, and Jerome Louradour. Dropout improves Recurrent Neural Networks for Handwriting Recognition. CoRR, abs/ 1312.4569, 2014.

Andrei A, Rusu, Neil C. Rabinowitz, Guiilaume Desjardins, Hubert Soyer, James Kirkpatrick, Ko- ray Kavukcuoglu, Razvan Pascanu, and Raia Hadseli. Progressive Neural Networks. CoRR, abs/1606.04671, 2016.

Richard Socher, Brody Huval, Christopher D. Manning, and Andrew Y. Ng. Semantic Composi- tionaliiy through Recursive Matrix-Vector Spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1201-1211, 2012.

Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive Deep Models for Semantic Composiiionality Over a Sentiment Treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631-1642, 2013.

Anders S0gaard. Semi-supervised condensed nearest neighbor for part-of-speech tagging. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 48-52, 2011.

Anders S0gaard and Yoav Goldberg. Deep multi-task learning with low level tasks supervised at lower layers. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 231-235, 2016.

Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to Sequence Learning with Neural Networks. In Advances in Neural Information Processing Systems 27, pp. 3104-31 12. 2014.

Jun Suzuki and Hideki Isozaki. Semi -Supervised Sequential Labeling and Segmentation Using Giga-Word Scale Unlabeled Data , In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 665-673, 2008.

K. S. Tai, R. Socher, and C. D. Manning. Improved semantic representations from tree- structured long short-term memory networks. 2015a.

Kai Sheng Tai, Richard Socher, and Christopher D. Manning. Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pp. 1556-1566, 2015b.

ristina Toutanova, Dan Klein, Christopher D Manning, and Yoram Singer. Feature-Rich Part- of-Speech Tagging with a Cyclic Dependency Network. In Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pp. 173-180, 2003.

Yoshimasa Tsuruoka, Yusuke Miyao, and Jun' ichi Kazama. Learning with Lookahead: Can History- Based Models Rival Globally Optimized Models? In Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pp. 238-246, 2011.

13 David Weiss, Chris Aiberti, Michael Collins, and Slav Petrov. Siractured Training for Neural Network Transition-Based Parsing. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 323-333, 2015.

John Wieting, Mohit Bansal, Kevin Gimpel, and Karen Livescu, CHA AG RAM : Embedding Words and Sentences via Character n-grams. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. to appear, 2016.

Wenpeng Yin, Hinrich Schtze, Bing Xiang, and Bowen Zhou. ABCNN: Attention- Based Con- voluiional Neural Network for Modeling Sentence Pairs. Transactions of the Association for Computational Linguistics, 4:259-272, 2016.

Yao Zhou, Cong Liu, and Yart Pan. Modelling Sentence Pairs with Tree- tructured Attentive Encoder. In Proceedings of the 26th International Conference on Computational Linguistics, pp. to appear, 2016.

14 Under review as a conference paper at ICLR 2017

APPENDIX

A How Do SHARED EMBEDDINGS CHANGE

In our JMT model, the word and character rt-gram embedding matrices are shared across all of the five different tasks. To better qualitatively explain the importance of the shortcut connections shown in Table 7, we inspected how the shared embeddings change when fed into the different bi- LSTM layers. More concretely, we checked closest neighbors in terms of the cosine similarity for the word representations before and after fed into the forward LSTM layers. In particular, we used the corresponding part of W _u in Eq. ( 1) to perform linear transformation of the input embeddings, because u _t directly affects the hidden states of the LSTMs. Thus, this is a context-independent analysis.

Table 10 shows the examples of the word "standing". The row of "Embedding" shows ihe cases of using the shared embeddings, and the others show the results of using the linear-transformed embeddings. In the column of "Only word", the results of using only the word embeddings are shown. The closest neighbors in the case of "Embedding" capture the semantic similarity, but after fed into ihe POS layer, the semantic similarity is almost washed out. This is not surprising because it is sufficient to cluster the words of ihe same POS tags: here, M , VBG, etc. In the chunking layer, the similarity in terms of verbs is captured, and this is because it is sufficient to identify ihe coarse chunking tags: here, VP . In the dependency layer, ihe closes- neighbors are adverbs, gerunds of verbs, and nouns, and all of them can be child nodes of verbs in dependency trees. However, this information is not sufficient in further classifying the dependency labels. Then we can see that in the column of "Word and char", jointly using the character n-gram embeddi tsgs adds the morphological information, and as shown in Table 8, the LAS score is substantially improved.

In the case of semantic tasks, the projected embeddings capture not only syntactic, but also semantic similarities. These results show that different tasks need different aspects of the word similarities, and our JMT model efficiently transforms the shared embeddings for the different tasks by the simple linear transformation. Therefore, without ihe shortcut connections, the information about the word representations are fed into the semantic tasks after transformed in the lower layers where the semantic similarities are not always important. Indeed, the results of the semantic tasks are very poor without the shortcut connections.

15 Under review as a conference paper at ICLR 2017

Word and char Only word

leaning stood

kneeling stands

Embedding saluting sit

clinging pillar

railing cross- legged

warning ladder

waxing rc6280

POS dunking beihle

proving warning

tipping f-a-18

applauding fight

disdaining favor

Chunking pickin pick

readjusting rejoin

reclaiming answer

guaranteeing patiently

resting hugging

Dependency grounding anxiously

hanging resting

hugging disappointment

stood stood

stands unchallenged

Related tiess unchallenged stands

notwithstanding beside

judging exists

nudging beside

skirting stands

Entailment straddling pillar

contesting swung

footing ovation

Table 10: Closest neighbors of the word "standing" in the embedding space and the projected space in each forward LSTM.

16