TECHNICAL FIELD

The present disclosure relates generally to the field of search engines; and more specifically to a method for responding to a search query.

BACKGROUND

Generally, various search engines, such as Google, Petal Search, Yandex, Google Lens, and the like, play a significant role in modern internet services. Such search engines retrieve the most relevant information within the available web content, in response to various queries made by users. The available web content is a large-scale dataset, which may include a billion or trillion raw data points. For example, as of the year 2021, there are 1.2 billion websites and 56.5 billion web pages, 130 million different books in the world, and 800 million videos hosted on YouTube® alone. The creation of web (or the internet) content is still ongoing at an unprecedented rate. Despite being large scale, the web content is usually created without an explicit intention to serve a search purpose (or a search query). To make a search successful, the search engines are required to understand the relevance between query inputs and raw data points, and make an efficient search from a large number of raw data points. In order to deliver an efficient and successful search experience, conventionally, raw data points of a large-scale dataset are firstly represented as feature vectors by employing a feature vector extraction model, to capture the similarity or relevance between the raw data points and various queries made by users. Thereafter, the feature vectors are organised into a structured database, for example, a backend vector database, to facilitate fast information retrieval.

The search engines are usually in long-term and continuous development and are frequently updated to perform effective searches and to adapt to new and emerging application scenarios. Typically, for an updated model, the corresponding vector database is regenerated. Therefore, new feature vectors have to be extracted from each data point of the large-scale dataset and organised into a new vector database. Thereafter, queries of the user are processed against the new vector database and the old vector database is discarded. However, the regeneration of the new vector database is both costly and time-intensive, which may result in delay or even cancellation of a new service publication. For example, extracting feature vectors from one trillion images would take 2.8 million GPU hours, given a feature extraction speed estimation of 10 milliseconds per image, which is prohibitively slow and expensive.

Thus, there exists a technical problem of updating a feature extraction model in order to deliver a cost-effective and successful search experience.

SUMMARY

The present disclosure provides a method for responding to a search query. An aim of the present disclosure is to provide a solution that overcomes the problems encountered in the prior art and provides an improved method for responding to a search query.

According to an aspect of the present disclosure, there is provided a method for responding to a search query generated by a user. The method comprises extracting one or more feature vectors from the search query and comparing the extracted feature vectors with a set of corresponding feature vectors in a pre-existing vector database. The method further comprises generating a first set of results based on the comparison and translating the extracted feature vectors and the first set of results to a new vector domain. The method further comprises comparing the translated extracted feature vectors and the translated first set of results and generating a second set of results based on the comparison, where the number of results in the second set is less than or equal to the number of results in the first set.

The disclosed method enables enhanced search experience and takes less time to upgrade the search service using a new feature extraction model with less computational resources as compared to existing systems. The extraction of feature vectors from the search query and the translating of the extracted feature vectors and the first set of results to the new vector domain avoids generating a new vector database from scratch (i.e., from raw data points), which in turn makes the method computationally efficient.

In an implementation form, the method further comprises at least one of the following: providing the second set of results to the user, or generating a third set of results based on the second set of results, and providing the third set of results to the user. It is advantageous to provide either the second set of results or the third set of results generated based on the second set of results to the user based on the relevance to the query made by the user.

In a further implementation form, the method further comprises ranking the results in the first set after the comparison with the translated extracted feature vectors.

The ranking (more specifically, re-ranking) the results in the first set after comparison with the translated extracted feature vectors results in better search accuracy.

In a further implementation form, the ranking is based on feature vectors extracted using an updated feature extraction model.

By virtue of performing the ranking based on feature vectors extracted using the updated feature extraction model, an efficient and successful search experience can be achieved.

In a further implementation form, the method further comprises aggregating the ranked results and the first set of results to generate the second set of results.

The aggregation of the ranked results in the first set and the first set of results generates the second set of results, which is likely to be more relevant to the query made by the user.

In a further implementation form, the method further comprises omitting the first set of results or the ranked results during aggregation.

The omission of the first set of results or the ranked results during aggregation may provide a more relevant set of search results to the user.

In a further implementation form, comparing comprises determining the similarity or relevance between the extracted feature vectors from the search query and the corresponding feature vectors in the pre-existing vector database.

The use of the pre-existing vector database for comparison with the extracted feature vectors from the search query leads to faster and more cost-efficient updates in the search service.

In a further implementation form, the method further comprises obtaining raw data points of the first set of results.

The obtaining of the raw data points of the first set of results leads to improving accuracy. In a further implementation form, the method further comprises translating the pre-existing vector database in its entirety to construct a new vector database.

The translation of the pre-existing vector database to the new vector database reduces complexity and runtime overhead.

In a further implementation form, translating the feature vectors comprises training a machine learning model with a batch of training data.

The training of the machine learning model with the batch of training data imparts information obtained from the batch of training data to the translated feature vectors.

In another aspect, the present disclosure provides a computer program comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method.

In an implementation form, the computer program is stored on a non-transitory computer- readable medium.

It has to be noted that all devices, elements, circuitry, units, and means described in the present application could be implemented in the software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity which performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements, or any kind of combination thereof. It will be appreciated that features of the present disclosure are susceptible to being combined in various combinations without departing from the scope of the present disclosure as defined by the appended claims.

Additional aspects, advantages, features, and objects of the present disclosure would be made apparent from the drawings and the detailed description of the illustrative implementations construed in conjunction with the appended claims that follow. BRIEF DESCRIPTION OF THE DRAWINGS

The summary above, as well as the following detailed description of illustrative embodiments, is better understood when read in conjunction with the appended drawings. For the purpose of illustrating the present disclosure, exemplary constructions of the disclosure are shown in the drawings. However, the present disclosure is not limited to specific methods and instrumentalities disclosed herein. Moreover, those in the art will understand that the drawings are not to scale. Wherever possible, like elements have been indicated by identical numbers.

Embodiments of the present disclosure will now be described, by way of example only, with reference to the following diagrams wherein:

FIG. 1 is a flowchart of a method for responding to a search query generated by a user, in accordance with an embodiment of the present disclosure;

FIG. 2 illustrates a technical workflow for responding to a search query through feature translation, in accordance with an embodiment of the present disclosure;

FIG. 3 illustrates a technical workflow for responding to a search query through feature translation, in accordance with another embodiment of the present disclosure;

FIG. 4 illustrates a training process of a feature translation model, in accordance with an embodiment of the present disclosure; and

FIG. 5 illustrates a technical workflow of processing of an old vector database to a new vector database through feature translation, in accordance with an embodiment of the present disclosure.

In the accompanying drawings, an underlined number is employed to represent an item over which the underlined number is positioned or an item to which the underlined number is adjacent. When a number is non-underlined and accompanied by an associated arrow, the nonunderlined number is used to identify a general item at which the arrow is pointing. DETAILED DESCRIPTION OF EMBODIMENTS

The following detailed description illustrates embodiments of the present disclosure and ways in which they can be implemented. Although some modes of carrying out the present disclosure have been disclosed, those skilled in the art would recognize that other embodiments for carrying out or practicing the present disclosure are also possible.

FIG. 1 is a flowchart of a method for responding to a search query generated by a user, in accordance with an embodiment of the present disclosure. With reference to FIG. 1, there is shown a method 100 for responding to a search query generated by a user. The method 100 includes steps 102 to 112.

The method 100 introduces an efficient way of generating search results according to one or more new artificial intelligence (Al) models through feature translation. The method 100 is described in detail, in following steps.

At step 102, the method 100 comprises extracting one or more feature vectors from a search query. The search query corresponds to a query from a user for which the user desires to obtain relevant information by availing a search service provided by a search engine. In an implementation, the search query may be a textual input. In another implementation, the search query may be an audio input. After receiving the search query from the user, one or more feature vectors are extracted from the search query. The process of extracting the feature vectors from the search query may also be termed as query features extraction. Generally, the feature vectors can be defined as one or more vectors that stores the features for one or more particular observations in a specific order. The extracted feature vectors from the search query may also be termed as query feature vectors. It is to be understood by a person of ordinary skill in the art that the search query may be a query generated automatically from an Artificial intelligence (Al) system in an example, without limiting the scope of the disclosure.

At step 104, the method 100 further comprises comparing the extracted feature vectors with a set of corresponding feature vectors in a pre-existing vector database. The pre-existing vector database may also be termed as an old vector database. The pre-existing vector database holds an existing feature vector database, which is generated using an old version of a feature extraction model. In this way, the method 100 uses the pre-existing vector database in contrast to conventional methods, which discard the pre-existing vector database. In an implementation, the step of comparing comprises determining the similarity or relevance between the extracted feature vectors from the search query and the corresponding feature vectors in the pre-existing vector database. The comparison of the extracted feature vectors (or query feature vectors) with the set of corresponding feature vectors in the pre-existing vector database includes calculating a similarity or relevance index between the extracted feature vectors from the search query and the corresponding feature vectors in the pre-existing vector database. In an example, the similarity or relevance index may be measured by use of mathematical metrics, such as feature correlation, Euclidean distance, Hamming distance, and the like. In another example, the similarity or relevance index can be measured automatically by use of a metric learned by an Al model from certain data, for example, through metric learning.

At step 106, the method 100 further comprises generating a first set of results based on the comparison. After comparing the extracted feature vectors (or query feature vectors) from the search query with the set of corresponding feature vectors in the pre-existing vector database, the first set of results is generated. The first set of results corresponds to the closest matches to the search query made by the user and the corresponding feature vectors in the pre-existing vector database. The first set of results may also be referred to as top-N closest matches, where N is an integer with a value more than the size of the final output of search results. The first set of results may also be referred to as initial ranking results.

In an implementation, the method 100 further comprises obtaining raw data points of the first set of results. In an implementation, the raw data points of the first set of results (e.g., the top- N closest matches) can be obtained, depending on the design of a feature translation model, described in detail, for example, in FIG. 4.

At step 108, the method 100 further comprises translating the extracted feature vectors and the first set of results to a new vector domain. After generation of the first set of results, the extracted feature vectors from the search query and the first set of results are translated to the new vector domain by use of the feature translation model, described in detail, for example, in FIG. 4.

In an implementation, translating the feature vectors comprises training a machine learning model with a batch of training data. The translation of the feature vectors (or the extracted feature vectors) is performed by use of the feature translation model. The feature translation model includes training the machine learning model using the batch of training data, described in detail, for example, in FIG. 4.

In an implementation, the method 100 further comprises translating the pre-existing vector database in its entirety to construct a new vector database. The translation of the feature vectors (or the extracted feature vectors) may result into a translation of the pre-existing vector database entirely to the new vector database.

At step 110, the method 100 further comprises comparing the translated extracted feature vectors and the translated first set of results. After translation of the extracted feature vectors and the first set of results to the new vector domain, the translated extracted feature vectors and the translated first set of results are compared with each other. The comparison of the translated extracted feature vectors with the translated first set of results includes calculation of similarity or relevance between the translated extracted feature vectors and the translated first set of results.

In an implementation, the method 100 further comprises ranking the results in the first set after the comparison with the translated extracted feature vectors. Based on comparison of the translated extracted feature vectors and the translated first set of results, the results in the first set are ranked (i.e., re-ranked) accordingly. Alternatively stated, the top-N closest matches are re-ranked based on comparison of the translated extracted feature vectors and the translated first set of results.

In an implementation, the ranking is based on feature vectors extracted using an updated feature extraction model. In an implementation, the ranking (i.e., re-ranking) of the results in the first set can be done based on feature vectors which are extracted using the updated feature extraction model (or the new feature extraction model). The updated feature extraction model may be an artificial intelligence (Al) feature extraction model, trained by various machine learning algorithms.

At step 112, the method 100 further comprises generating a second set of results based on the comparison, where the number of results in the second set is less than or equal to the number of results in the first set. After comparison of the translated extracted feature vectors and the translated first set of results, the second set of results is generated based on the comparison. However, the number of results in the second set is less than or equal to the number of results in the first set. The second set of results may also be referred to as top-K closest matches, where K is an integer with a value less than or equal to the number of results in the first set (i.e., the top-N closest matches).

In an implementation, the method 100 further comprises aggregating the ranked results and the first set of results to generate the second set of results. The aggregation of the ranked results in the first set and the first set of results generates the second set of results, which may be more relevant to the query made by the user.

In an implementation, the method 100 further comprises omitting the first set of results or the ranked results during aggregation. The aggregation of the first set of results (i.e., top-N closest matches) and the ranked (i.e., re-ranked) results in the first set, can omit either the first set of results (i.e., top-N closest matches) or the ranked (i.e., re-ranked) results in the first set. The omission of the first set of results or the ranked results during aggregation may remove duplicate or non-relevant results and provide a more accurate set of results to the user.

In an implementation, the method 100 further comprises at least one of the following: providing the second set of results to the user, or generating a third set of results based on the second set of results, and providing the third set of results to the user. In an implementation, the second set of results may be provided to the user as the final output of the search query. In another implementation, the third set of results may be generated based on the second set of results and provided to the user as the final output of the search query.

Thus, the method 100 is based on updating the feature extraction model in the search engine at a fast rate and in a cost-effective manner. Typically, with an updated Al model, a new vector database is regenerated from the raw data points, which is very slow and cost-intensive. However, the method 100 updates the feature extraction model using the feature translation model which translates the pre-existing vector database into the new vector database instead of regenerating the new vector database from the raw data points.

In contrast to a conventional method of updating the feature extraction model, where the new vector database is considered and the pre-existing vector database (i.e., the old vector database) is discarded, the method 100 is based on using the pre-existing vector database (i.e., the old vector database) for generating the first set of results. This makes the method 100 computationally efficient and reliable as compared to existing methods and systems. Moreover, the method 100 is adaptive to migrations between the pre-existing vector database (i.e., the old vector database) and the updated feature extraction model (i.e., new feature extraction model). Furthermore, the method 100 can be used to verify the effectiveness of the updated feature extraction model (i.e., new feature extraction model) on a search service in real time by either switching on or switching off the feature translation model.

The steps 102 to 112 are only illustrative, and other alternatives can also be provided where one or more steps are added, one or more steps are removed, or one or more steps are provided in a different sequence without departing from the scope of the claims herein.

FIG. 2 illustrates a technical workflow for responding to a search query through feature translation, in accordance with an embodiment of the present disclosure. FIG. 2 is described in conjunction with elements from FIG. 1. With reference to FIG. 2, there is shown a block diagram 200 of a technical workflow for responding to a search query through feature translation. The technical workflow includes a sequence of operations, such as operations 202, 204, 206, 210, 212, 214, 216, 218, 220, and 222

At operation 202, a search query of user is received as an input. The received input may also be referred to as a system input.

At operation 204, features from the search query are extracted. Alternatively stated, feature vectors from the search query are extracted using a pre-existing version of a feature extraction model.

At operation 206, the extracted feature vectors from the search query and a set of corresponding feature vectors from a pre-existing vector database 208 are compared with each other. The preexisting vector database 208 refers to an old vector database.

At operation 210, a first set of results (i.e., top-N closest matches) is generated based on the comparison between the extracted feature vectors from the search query and the set of corresponding feature vectors in the pre-existing vector database 208.

At operation 212, the extracted feature vectors are translated to new vector domain, using a feature translation model, described in detail, for example, in FIG. 4. After translation, the translated extracted feature vectors are obtained.

At operation 214, the first set of results (i.e., the top-N closest matches) is translated to a new vector domain using the feature translation model. After translation, the translated first set of results is obtained. At operation 216, the translated extracted feature vectors and the translated first set of results are compared with each other.

At operation 218, based on comparison between the translated extracted feature vectors and the translated first set of results, the results in the first set are re-ranked.

At operation 220, the first set of results and the re-ranked results in the first set are aggregated using various aggregation algorithms. During aggregation, either the first set of results or the re-ranked results in the first set may be omitted.

At operation 222, a second set of results (i.e., top-K closest matches) is generated based on an aggregation of the first set of results and the re-ranked results in the first set. The second set of results is provided to the user as final outputs of the search query.

The method 100 (of FIG. 1) may be executed on or by a device that may include but not limited to, a computer, a server, a portable electronic device, a smart phone, a tablet, and the like. The device may include a processor that may be configured to execute operations of the method 100. In an implementation, the device may have either a single hardware server and/or a plurality of hardware servers operating in a parallel or distributed architecture.

FIG. 3 is a diagram that illustrates a technical workflow for responding to a search query through feature translation, in accordance with another embodiment of the present disclosure. FIG. 3 is described in conjunction with elements from FIGs. 1, and 2. With reference to FIG. 3, there is shown a block diagram 300 that illustrates a technical workflow for responding to a search query through feature translation.

The technical workflow of the block diagram 300 for responding to a search query is same as that of the technical workflow of the block diagram 200 (of FIG. 2) except that the operation 212 (of the FIG. 2) is replaced by an operation 302 in FIG. 3. Similar to the block diagram 200 (of FIG. 2), the block diagram 300 illustrates the update of the search service with the new Al model in an online mode.

At operation 302, feature vectors from the search query are extracted using an updated feature extraction model. The updated feature extraction model is obtained using a feature translation model, described in detail, for example, in FIG. 4. Thereafter, the feature vectors extracted using the updated feature extraction model are compared with the translated first set of results at the operation 216. Since, the query feature vectors are extracted using the updated feature extraction model, therefore, there is no need for translating the extracted feature vectors for carrying out the comparison with the translated first set of results. The translation of the extracted feature vectors may incorporate inaccuracies. Therefore, in the FIG. 3, the feature vectors extracted using the updated feature extraction model are used for the computation of the second set of results. Since, the second set of results (i.e., the top-K closest matches) are obtained using the updated feature extraction model hence, the second set of results (i.e., the top-K closest matches) are more accurate and reliable.

FIG. 4 illustrates a training process of a feature translation model, in accordance with an embodiment of the present disclosure. FIG. 4 is described in conjunction with elements from FIGs. 1, 2, and 3. With reference to FIG. 4, there is shown a block diagram 400 that illustrates a training process of a feature translation model. The block diagram 400 includes a training dataset 402, a first feature extraction model 404, a second feature extraction model 406, a first feature vector 408, a second feature vector 410, a feature translation model 412, a third feature vector 414 and a loss function 416.

The feature translation model 412 can be trained as follows. The old feature vector extraction model can be represented as the first feature extraction model 404 (also represented as feature vector extraction model #1) and an updated feature vector extraction model can be represented as the second feature extraction model 406 (also represented as feature vector extraction model #2). The training dataset 402, the first feature extraction model 404 (i.e., feature vector extraction model #1) and the second feature extraction model 406 (i.e., feature vector extraction model #2) are used to train the feature translation model 412.

In order to train the feature translation model 412, the feature vectors corresponding to the first feature extraction model 404 (i.e., feature vector extraction model #1) and the second feature extraction model 406 (i.e., feature vector extraction model #2) are extracted and are represented as the first feature vector 408 (also represented as feature vectors #1) and the second feature vector 410 (also represented as feature vectors #2), respectively. In this way, each training data point receives a pair of feature vectors that is the first feature vector 408 and the second feature vector 410. The first feature vector 408 is considered as an input to the feature translation model 412 and an output of the feature translation model 412 can be represented as the third feature vector 414 (also represented as feature vectors #3). However, the third feature vector 414 (i.e., the feature vectors #3) has the same feature dimensionality with the second feature vector 410 (i.e., the feature vectors #2).

Optionally, the training dataset 402 can be fused with the first feature vector 408 (i.e, the feature vectors #1) to act as the input to the feature translation model 412. The feature translation model 412 can be implemented using a machine learning model. Examples of implementation of the machine learning model may include but are not limited to, auto-encoders, generative adversarial networks (GAN), and the like. In order to converge the third feature vector 414 (i.e., the feature vectors #3) to the second feature vector 410 (i.e., the feature vectors #2), either the loss function 416 between the third feature vector 414 (i.e., the feature vectors #3) and the second feature vector 410 (i.e., the feature vectors #2) is enforced or relaxed criteria is set between the third feature vector 414 and the second feature vector 410 so that the third feature vector 414 generates same topological properties as the second feature vector 410 generates. For example, the third feature vector 414 may have the same closest neighbors for each data point within the training dataset 402, after converging to the second feature vector 410.

The feature translation model 412 has lower complexity and runtime overhead in comparison to the new feature extraction model. The feature translation model 412 can get full information from the training dataset 402 which is used to develop the new feature extraction model through the design of input. The feature translation model 412 considers the existing feature vectors (i.e., pre-existing vector database or the old vector database) as the input that is already in use for the search service, which means that the feature translation model 412 is not learning from scratch (i.e., from the raw data points) as the new feature extraction model is generated from scratch. Moreover, the training process of the feature translation model 412 can be seen as a model distillation procedure that learns directly from the new feature extraction model than from the training labels. The loss function 416 can be designed with relaxation to enforce the second feature vector 410 (i.e., the feature vectors #2) and the third feature vector 414 (i.e., the feature vectors #3) to produce the same search results, for example, ranking the results in the database inspite of strictly forcing the results to be the same representations through learning. The feature translation model 412 can be switched on or switched off to allow the A/B test without updating the vector database, depending on an application scenario.

FIG. 5 illustrates processing of an old vector database to a new vector database through feature translation, in accordance with an embodiment of the present disclosure. FIG. 5 is described in conjunction with elements from FIGs. 1, 2, 3, and 4. With reference to FIG. 5, there is shown a processing pipeline 500 that illustrates processing of a pre-existing vector database 502 (i.e., an old vector database) to a new vector database 508 through feature translation. The processing pipeline 500 includes operations 504 and 506.

Initially, the pre-existing vector database 502 (i.e., the old vector database) is configured. Thereafter, at operation 504, the pre-existing vector database 502 is translated to a new vector domain using the feature translation model 412 (of FIG. 4). At operation 506, a database is generated corresponding to the new vector domain of the pre-existing vector database 502, which further results in the generation of the new vector database 508.

The generated new vector database 508 replaces the existing one that is the pre-existing vector database 502 (i.e., the old vector database) and serves as the new backend vector database for search service. Thus, instead of processing the first set of results (i.e., the top-N closest matches) using the feature translation model 412 in the online mode as described in FIGs. 2 and 3, the old vector database 502 is translated and reorganized into the new vector database 508 in an offline mode. Since, the feature translation model 412 manifests low complexity and runtime overhead, the time and resource costs should be much lower than extracting new vector database 508 directly from the new feature extraction model. Thus, the processing of the old vector database 502 to the new vector database 508 using the feature translation model 412 in the offline mode provides a fast and cost-effective feature extraction model update in a search engine.

Modifications to embodiments of the present disclosure described in the foregoing are possible without departing from the scope of the present disclosure as defined by the accompanying claims. Expressions such as "including", "comprising", "incorporating", "have", "is" used to describe and claim the present disclosure are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural. The word "exemplary" is used herein to mean "serving as an example, instance or illustration". Any embodiment described as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments and/or to exclude the incorporation of features from other embodiments. The word "optionally" is used herein to mean "is provided in some embodiments and not provided in other embodiments". It is appreciated that certain features of the present disclosure, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the present disclosure, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable combination or as suitable in any other described embodiment of the disclosure.