The present invention relates to a directory assistant method and apparatus, and more particularly, to a directory assistant method and apparatus in an automatic dialogue telecommunications system.

The directory assistant (DA) system, providing telephone numbers to customers, is an important telecommunications business. For example, the Kelsey Group estimates that telecom companies worldwide collectively receive more than 516 millions DA calls per month, almost all of which are currently handled by operators.

Automating this service using speech recognition is a large market opportunity.

The conventional DA system is implemented by using a restricted dialogue. Traditionally, it first asks a user to say the name of the person to be reached, then uses a speech recognizer to locate several candidates from a directory database. If the candidates are too many, the DA system further asks the user to spell the name of the desired person or to provide extra information, for example, the name of the street where the desired person lives. In this way, the rang of the candidates can be further narrowed down. Finally, the DA system asks the user to choose the right one by answering a corresponding number or just"yes/no". This DA system works well for a small Western DA system. But it may not work well for a large-scale directory assistant system having, for example, 12,000, 000 entries used in a large city, since the above- mentioned input information is not sufficient to differentiate all possible candidates.

The same system does not work well for a large-scale Chinese DA system, either. The input information is not sufficient to differentiate all possible candidates due to the following specific features. First of all, Chinese is a monosyllabic language. Each word of Chinese contains exactly one syllable. There are more than 13000 commonly used words and only 1300 legal syllables. On average, there are about

10 homophones for each syllable. Secondly, a Chinese name is usually shorter than a Western name. The Chinese name usually has three syllables only. Moreover, there are about two hundred family names (surnames) frequently used by billions of Chinese.

Therefore, more information is needed to solve the ambiguities in a Chinese DA system.

Thirdly, Chinese is an ideographic language. Chinese usually introduce their names to other people by describing their name word by word and by some commonly used phrases. There is no easy and standard way to"spell"or"compose"Chinese words.

Therefore, the performance of current DA systems is not satisfactory, especially the Chinese DA systems.

It is an object of the present invention to provide a directory assistant method and apparatus for providing desired directory entry information. The directory assistant method and apparatus use natural language dialogue system to ask users to describe the desired directory entry and then use the relevance knowledge databases to parse and understand these descriptions and interpret their meanings. Finally, the directory assistant method and apparatus integrate all available information from several dialogues turns, directory database and relevant knowledge database, and then provide the users'desired directory entry information.

It is another object of the present invention to provide a computer program product residing on a computer readable medium having a plurality of instructions stored thereon which, when executed by a processor, cause the processor to provide desired directory entry information.

The invention will become more fully understandable from the detailed description given below and the accompanying drawings, which are given by way of illustration only and thus are not imitative, wherein: Fig. 1 illustrates a block diagram of the directory assistant apparatus of the present invention;

Fig. 2 illustrates the generation of name description grammar rules from frequently used templates and frequently used words of the present invention; and Fig. 3 illustrates a flow chart of the directory assistant method of the present invention.

Referring to Fig. 1, the directory assistant apparatus of the present invention, such as a Mandarin directory assistant apparatus, comprises a database 30 for storing directory entry information, grammar rules and concept sequences; an acoustic recognition unit 10 for receiving a speech signal describing the desired directory entry, recognizing the speech signal and generating recognized word sequences; a speech interpreting unit 20 for interpreting the recognized word sequences by using a predetermined grammar rule and relevant information thereof stored in the database 30 to form concept sequences and interpreting the concept sequences according to semantic meaning and relevant information thereof stored in the database and current system status thereof, thereby generating at least one candidate by using one of maximum a posterior probabilities and maximum likelihood criterion for the desired directory entry, in addition, the speech interpreting unit 20 further updates the system status; a look up unit 40 for looking up at least one directory entry information corresponding to the at least one candidate from the database 30; and an output unit 60 (such as a speech output unit) for outputting the at least one directory entry information located.

The directory assistant apparatus of the present invention further comprises a question generator 60 for generating a question to request more information, wherein the question is one of requesting a user to supply more information, listing-based confirmation and open-question confirmation. The listing- based confirmation is used when the potential candidates are in a limit of numbers, or the probability of the top one is far from those of the others. The open-question confirmation is used when the most popular description way in name database to ask users for confirmation, for example: you mean (the same Li3 as in Li3 Dengl Huel).

The acoustic recognition unit 10 further comprises a speech recognizer

11 for recognizing the input speech signal and generating recognized word sequences; a confusion analyzer 12 for expanding the recognized word sequences according to a confusion table 13, wherein the confusion table 13 is pre-trained and comprises all confusable words, their corresponding correct ones and occurring probabilities; and a confidence measurement unit 14 for filtering out confusable word pairs according to a confidence table 15.

The database 30 comprises a relevant knowledge database 31 and a directory database 32. The relevant knowledge database 31 comprises words and using frequencies thereof, ways to describe the words, grammar rules, attributes and corresponding using frequencies, communication concepts and their frequencies of usage, corresponding grammar rules, semantic meaning and frequencies of usage, while the directory database 32 comprises a plurality of entries, wherein each entry comprises name, telephone number, relevant information and frequencies of usage.

In the relevant knowledge database 31, popular words in names are stored with several popular descriptive ways, wherein the grammar rule and concept sequences are used to describe the desired directory entry comprising entry name or at least one word of entry name and relevant information thereof, and the grammar rule is generated by frequently used grammar templates and frequently used words. The grammar templates are generated by one of frequently used nouns, names of famous people, idioms, character strokes, letters, words, and character roots, etc. For examples, words in names can be described as follows: - famous family name description, like (same Li3 as in Li3 Dengl Huel).

- famous name description, like (same Li3 as in Li3 Dengl Huel).

- Common used word, phrase and especially four-word Chinese idioms, like (same Zhau4 as in Zhau4 Chien2 Sunl Li3). common used writing/strokes description, like (Wang2 that has three horizontal lines and one vertical line); or (Chen2 that has an ear and a east).

Fig. 2 illustrates the generation of name description grammar rules from frequently used templates and frequently used words of the present invention. The present invention first builds a database which collects the name description grammar rules and their corresponding semantic tags.

There are two ways to build the database. The first way is to collect as many names as possible and their corresponding character descriptions. From this database, we have found name description grammar rules and their probability statistics such as LN (descriptions of the last name) 84, FN1 (descriptions of the first word of the first name) 85 and FN2 (description of the second word of the first name) 86.

The second way is to find frequently used grammar templates from a small database of name descriptions (for example the database mentioned above). Then we use the found grammar templates and frequently used words to generate the necessary grammar rules. For example, we have found that the most popular ways to describe the words of names are: - Frequently used nouns (FNoun) 81 ; - Names of famous people (FName) 82; - Idioms (CI); - Character Strokes (CS); - Frequently used foreign words (FW); - Character roots (CR) 83; Other irregular way (OW).

We can then build the necessary grammar rules by combining these grammar templates and frequently used words (collected from dictionary, internet, newspaper, etc.).

Referring to Fig. 3, the directory assistant method of the present invention, such as a Mandarin directory assistant method, is described as follows: The method first prompts a question to ask the user to speech input the desired entry (100); then receives a speech signal describing the desired directory entry (110) ; recognizes the speech signal and generating recognized word sequences, expands

the recognized word sequences according to a confusion table and filters out confusable word pairs according to a confidence table (120); interprets the recognized word sequences by using a predetermined grammar rule and relevant information thereof stored in a database to form concept sequences (130) ; interprets the concept sequences according to semantic meaning and relevant information thereof stored in the database and current system status thereof, thereby generating at least one candidate by using one of maximum a posterior probabilities and maximum likelihood criterion for the desired directory entry and updates the system status; looks up at least one directory entry information corresponding to the at least one candidate from the database and generates a question to request more information if there are uncertainties (150) ; outputs the at least one directory entry information located (160); and confirms the located directory entry information and repeats the above steps until the desired directory entry information is located (170).

The above-mentioned method can be implemented by computer program instructions. The computer program instructions can be loaded into a computer or other programmable processing devices to perform the functions of the method illustrated in Fig. 3. The computer program instructions can be stored in a computer program product or computer readable medium. Examples of a computer program product or computer readable medium includes recordable-type medium such as a magnetic tape, a floppy disc, a hard disc drive, a RAM, a ROM and an optical disc and transmission-type medium such as digital and analog communication links.

The present invention is directed to understanding ways of describing words in names, building up a relevant knowledge database to store ways of describing words in names and using the database as grammar rules to parse input speech. By this new architecture, the present invention can use the natural language dialogues system to ask the user to describe the words of names when there are still uncertainties. The present invention then uses the relevance knowledge database to parse and understand these descriptions and interpret their meaning. Finally, the present invention combines all available information to narrow down the range of possible candidates and finally locates the correct directory entry. Although part of the present invention is described by using Chinese words as an example, the present invention can also be applied to other languages. For example, famous family name description, like (same Li3

as in Li3 Dengl Huel), can be changed to"Bush as in George Bush." Although the present invention and its advantages have been described in detail, it should be understood that various changes, substitutions and alternations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

PRIOR ART REFERENCES 1. Nick Wang and Leo Liao, Chinese name recognition, invention disclosure, submitted to Philips Corporate Intellectual Property, 11 Oct. 2000 2. Andreas Kellner, Bernhard Rueber, Frank Seide, and Bach-Hiep Tran., "PADIS--an Automatic Telephone Switchboard and Directory Information System", Speech Communication, 23: 95--111, Oct 1997 3. Bernd Souvignier, Andreas Kellner, Bernhard Rueber, Hauke Schramm, and Frank Seide. , "The Thoughtful Elephant: Strategies for Spoken Dialog Systems". IEEE Transactions on Speech and Audio Processing, 8 (1) : 51--62, January 2000 4. Georg Rose,"PADIS-XXL, Large Scale Directory Assistant, " PowerPoint Slide, Man-Machine Interface, 2000.

5. Andreas Meylahn, "SpeechFinder for SpeechPerl 2000-Developer's Guide v2.0", Philips Speech Processing-Aachen, Germany, 21 Sep. 2000 6. Yen-Ju Yang"Statistics-based spoken dialogue modeling <BR> and its applications, "PhD's Thesis, National Taiwan University, Taiwan, 1999.<BR> <P>7. Jan Kneissler and Dietrich Klakow, "Speech Recognition for Huge Vocabularies by Using Optimized Sub-Word Units" 8. Dietrich Klakow, Georg Rose and Xavier Aubert, "OOV-Detection in Large Vocabulary System Using Automatically Defined Word-Fragments as Fillers", Proc. EUROSPEECH, Vol. 1,1999, pp. 49-53.

9. Alex Weibel, Petra Geutner, Laura Mayfield Tomokiyo, Tanja Schultz and Monika Woszcyna, "Multilinguality in Speech and Spoken Language Systems,"Proceedings of the IEEE, Vol. 88, No. 8, August 2000, pp. 1297-1313.