JP4177195B2 - Recognition grammar creation system - Google Patents

Description

  The present invention relates to a telephone response service such as a voice portal using a voice recognition function and an automatic response service such as a car navigation system with a voice recognition function, and in particular, recognition for creating a recognition grammar used in such a service. It relates to a grammar creation system.

  Recent typical speech recognition methods include continuous word recognition and word spotting, but each method requires the recognition vocabulary to be described separately, and the one that describes the recognition vocabulary is called "recognition grammar". It is. The recognition grammar needs to cover the vocabulary spoken by the user. This is because only the vocabulary described in the recognition grammar can be recognized, and the accuracy of the recognition grammar (whether it covers the user utterance) directly leads to the recognition rate. Therefore, an accurate recognition grammar is required. However, the difficulty in creating a recognition grammar lies in the diversity of human utterances. Languages change with the times, for example, they are obsolete just like the abbreviations we have omitted in recent years. There are a number of phrases, including omissions, to express a concept, and the recognition grammar must cover this.

  For example, assume that a recognition grammar for company names in Japan is created. The official vocabulary (official name) has, for example, a database that collects the official names of company names, and based on this database, a recognition grammar can be created. However, there is no database that stores phrases often used in conversations, including abbreviations, so it is necessary to collect such phrases. Examples of combinations of formal vocabulary and phrases often used in conversational sentences include the phrase “JR West” for the official vocabulary “West Japan Railway Company” and “ANA” for the official vocabulary “All Nippon Airways”. (Ana) "and so on.

  In the case of such a well-known company, it is possible to come up with phrases, but in other cases, as a handy means, for example, a WWW search engine is used to collect phrases. In addition, there are certain rules based on empirical rules, and for example, they are often created by a method such as omitting or taking an initial letter.

  Here, FIG. 14 shows a conventional procedure for creating a recognition grammar. That is, conventionally, first, a phrase candidate such as an abbreviated name is created from the original vocabulary (step S91), and a large-scale text corpus is searched to check whether the created phrase candidate actually exists. (Step S92). As such a large-scale text corpus, for example, WWW (World Wide Web) is used. Then, for example, the search results are reflected in the vocabulary database, such as registering only the wording candidates that have been confirmed by the search of the large text corpus in the vocabulary database (step S93), and then the recognition grammar is created from the vocabulary database. (Step S94).

Conventionally, a speech recognition text input device that attempts to improve recognition performance in response to user-specific utterance variations by adding user-specific vocabulary and expressions to the user language model is also disclosed in Patent Document 1. ing.
JP 2002-229585 A

  However, according to the above-described conventional procedure, a large text corpus such as the WWW is searched in the dark cloud in order to check whether or not there is a wording candidate, so the wording candidate and the text are the same. Even when a word having a different meaning is a search hit, it is often erroneously determined that the wording candidate exists. Then, if the recognition grammar is generated as it is from the vocabulary registered in the vocabulary database based on such a misjudgment, there is a problem that only a grammar with low accuracy can be obtained.

  In order to solve the above-mentioned problem, the present invention can improve the accuracy of determining the suitability of wording candidates by selecting a text corpus to be searched, and can create a recognition grammar with high accuracy. It aims at providing a simple recognition grammar creation system.

  In order to achieve the above object, the recognition grammar creation system according to the present invention selects and selects a vocabulary database for storing formal vocabularies and their wordings, and a corpus to be searched for checking suitability of wording candidates. A search corpus selection unit that requests the corpus to search for the wording candidate, a search result analysis unit that analyzes a search result in the corpus selected by the search corpus selection unit, an analysis result by the search result analysis unit, and a predetermined result A database update unit that determines the suitability of the wording candidates based on the criteria, and stores the wording having relevance among the wording candidates in the vocabulary database, and reads out the vocabulary from the vocabulary database and reads the vocabulary according to the grammar specification The configuration includes a grammar creation unit for creating a recognition grammar.

  According to this configuration, the search corpus selection unit appropriately selects a corpus to be searched for checking the suitability of the phrase candidates, thereby preventing a phrase that is not actually used from being registered in the vocabulary database. it can. In this way, it is possible to create a recognition grammar with high accuracy by using a vocabulary database that contains phrases that are actually used.

  In the recognition grammar creation system according to the above configuration, the search corpus selection unit includes a candidate corpus database that stores information on a candidate corpus that is a candidate corpus to be searched, and a search keyword for the candidate corpus that is related to the formal vocabulary. A candidate corpus search unit that inputs at least one keyword to search for information on candidate corpora on the Internet according to the search keyword, and candidate corpus registration that registers information obtained by the candidate corpus search unit in the candidate corpus database It is preferable that the search corpus selection unit refers to information registered in the candidate corpus database to select a corpus to be searched.

  In the above aspect, “information on candidate corpus” includes access information to the candidate corpus. According to the above aspect, information on candidate corpora that are candidates for selection of the search corpus selection unit is registered in the candidate corpus database, and the search corpus selection unit refers to this to check the suitability of the wording candidate. Therefore, it is possible to appropriately select a corpus to be searched for.

  In the recognition grammar creation system according to the above configuration, the search result analysis unit includes an appearance information generation unit that generates the appearance information of the wording candidate in the corpus, and the generation result of the appearance information by the appearance information generation unit is It is also preferable that the analysis result is passed to the database update unit. The appearance information of the vocabulary in the corpus can be regarded as the degree to which the vocabulary is used. Therefore, by determining whether each wording candidate is stored in the vocabulary database according to the degree to which the vocabulary is used, a highly accurate vocabulary is obtained. This is because a database can be created, and thus a highly accurate recognition grammar can be created.

  In the recognition grammar creation system according to the above configuration, when the database update unit stores the wording in the vocabulary database, the appearance information generated by the appearance information generation unit for the wording is associated with the wording and the wording is generated. The recognition grammar creation system further includes a grammar vocabulary determination unit that selects a vocabulary for which a recognition grammar should be created from the vocabulary database and passes the grammar creation unit to the grammar creation unit. It is also preferable to adopt an embodiment. This is because the appearance information of the vocabulary in the corpus can be regarded as the degree to which the vocabulary is used, so that it is useful as a criterion for determining whether to add the vocabulary to the grammar when creating the grammar.

  In the recognition grammar creation system according to the above configuration, the grammar creation unit may create a recognition grammar having information indicating a word priority based on the appearance information. This is because the accuracy of speech recognition using the recognition grammar can be improved by providing the recognition grammar with information indicating the degree to which the vocabulary is used.

  According to the present invention, by selecting a text corpus to be searched, a recognition grammar creation system capable of improving the accuracy of determining the suitability of wording candidates and creating a highly accurate recognition grammar is created. it can.

  FIG. 1 is a block diagram showing a schematic configuration of a recognition grammar creation system according to an embodiment of the present invention. As shown in FIG. 1, the recognition grammar creation system according to the present embodiment includes a vocabulary database 1, a vocabulary candidate creation unit 2, a grammar creation unit 3, a database editing unit 4, a synonym database 5, a database update unit 6, and a search corpus selection. Unit 7, search result analysis unit 9, and grammar storage unit 10.

  In this recognition grammar creation system, the vocabulary candidate creation unit 2 creates a wording candidate from the formal vocabulary, and requests the large-scale corpus 8 outside the system to search for the wording candidate. As the large-scale corpus 8, WWW connected locally to the recognition grammar creation system, WWW connected via an external network such as the Internet, or the like can be used. Of course, it is also possible to use a corpus other than WWW.

  The large-scale corpus 8 executes the requested search and sends the search result to the search result analysis unit 9. The search result analysis unit 9 determines the suitability of the wording candidate based on the search result sent from the large-scale corpus 8 and passes the determination result to the database update unit 6. Then, the database update unit 6 registers only the wording candidates determined to be suitable by the search result analysis unit 9 in the vocabulary database 1.

  Hereinafter, the configuration and operation of each unit will be described in more detail.

  The vocabulary database 1 stores a formal vocabulary and a so-called “phrase” used in a conversation sentence as a modification of the formal vocabulary in association with each other. For example, FIG. 2 shows an example of data registered in the vocabulary database 1. In this example, in the vocabulary database 1, one entry is an ID (identification number), a vocabulary number indicating the number of vocabularies included in the entry, and vocabularies corresponding to the number of vocabularies (first vocabulary, second vocabulary,... ) And. The original vocabulary (formal vocabulary) is registered in the first vocabulary, and the wording of the first vocabulary is registered after the second vocabulary. For example, in the example of FIG. 2, the vocabulary “general affairs”, which is an abbreviation thereof, is registered as the phrase of the first vocabulary (official vocabulary) “general affairs department”. In this example, each entry has an ID and a vocabulary number, but these are not essential.

  The database editing unit 4 includes an input device such as a keyboard and a mouse or an output device such as a display (both not shown), and the user performs editing operations on the entries of the vocabulary database 1 and the contents of the vocabulary database 1 Can be confirmed by the user. For example, when creating a new vocabulary entry in the vocabulary database 1, the user inputs a formal vocabulary (first vocabulary) using the database editing unit 4. In addition, if there is a word that comes to mind, it can be added after the second vocabulary using the database editing unit 4 in the same manner.

  The vocabulary candidate creation unit 2 generates a candidate (phrase candidate) to be registered in the vocabulary database 1 as a wording of the formal vocabulary from the formal vocabulary registered in the first vocabulary of the entry according to a predetermined rule. In this embodiment, the vocabulary candidate creation unit 2 registers the generated wording candidates in the vocabulary database 1 after the second vocabulary. The vocabulary candidate creation unit 2 may refer to the synonym database 5 when creating the wording candidate.

  Here, a procedure for generating wording candidates by the vocabulary candidate creating unit 2 will be described with a specific example.

  As described above, the vocabulary candidate creation unit 2 generates wording candidates according to a predetermined rule. As the predetermined rule, for example, the following empirical rule (heuristic rule) can be considered. Note that the following four are merely examples, and any other rule can be applied on condition that an appropriate wording can be generated.

  (1) The name “XXX” is called “XXX” by omitting “part”.

  (2) “XX Research Department” is called “XX Lab” by omitting the back of “ken”.

  (3) When composed of a plurality of words, it is called by taking the first letter of the word.

  (4) When composed of a plurality of words, it is called by taking the first letter of the word.

  For example, if the user uses the database editing unit 4 to create an entry as shown in FIG. 2 in the vocabulary database 1, the vocabulary candidate creation unit 2 generates a wording candidate using the above empirical rules, Register after the second vocabulary of each entry. For example, the result is as shown in FIG. In the example shown in FIG. 3, the vocabulary candidate creating unit 2 performs “Sojihon” and “SJH” for the formal vocabulary “Software Business Headquarters”. The wording candidate “Essieieichi)” is generated and registered in the second vocabulary and the third vocabulary of the entry. In addition, for the official vocabulary “Media Solutions Research Department”, the Vocabulary Candidate Creation Department 2 has created “Media Solutions Laboratory”. ”,“ Mesoken ”,“ MS Research Department ”, and“ MS Lab ”are generated, and the second to fifth entries are created. Registered in the vocabulary.

  Here, some examples of the configuration of the vocabulary candidate creation unit 2 for creating wording candidates by omission will be specifically described with reference to FIGS.

  The configuration example shown in FIG. 4 is a configuration for creating a wording candidate using a plurality of initial words such as “Software Business Headquarters” to “Soviet Books”. In this case, the vocabulary candidate creation unit 2 may be provided with a morphological analysis unit 21, an initial acquisition unit 22, and an initial synthesis unit 23. The morpheme analysis unit 21 performs morpheme analysis on the input formal vocabulary and divides it into parts of speech. For example, the formal vocabulary “software business headquarters” is divided into three nouns “software”, “business”, and “headquarters”. The initial letter acquisition unit 22 takes the first letter of the divided part of speech. That is, in the above case, “So”, “Thing”, and “Book” are acquired. Next, the initial character synthesizing unit 23 combines the characters acquired by the initial character acquiring unit 22 according to the word order, so that “Sojihon” is obtained as a wording candidate.

  Further, the configuration example shown in FIG. 5 is a configuration for creating a wording candidate “MS Research Department” from “Media Solution Research Department”, for example. In this case, the vocabulary candidate creation unit 2 includes the abbreviation rule application unit 24, the abbreviated creation rule storage unit 25, and the alphabet conversion unit 26 in addition to the morpheme analysis unit 21, the initial acquisition unit 22, and the initial synthesis unit 23 described above. May be provided.

  In the configuration example shown in FIG. 5, for example, when the above-mentioned “media solution research department” is input as a formal vocabulary, the morpheme analysis section 21 converts this into “media”, “solution”, “research”, “part”. And is passed to the omission rule application unit 24. The omission rule application unit 24 refers to the omission creation rule storage unit 25 and creates an abbreviation by applying the omission creation rule. The omission creation rule storage unit 25 stores in advance omission rules obtained from empirical rules and the like. The omission rules stored here are, for example, as follows.

  (1) “Research Institute” becomes “Research”.

  (2) “Research Department” becomes “Lab”.

  (3) “Research Center” becomes “Lab”.

  (4) “XXX part” becomes “XX”.

  By applying such abbreviated rules, the abbreviated rule applying unit 24 creates an abbreviation “ken” from the two nouns “research” and “part” among the above four nouns and omits them. Along with “Media” and “Solution” that have not been performed, the data is transferred to the alphabet conversion unit 26. The alphabet conversion unit 26 converts katakana characters into alphabets. Next, the initial acquisition unit 22 takes the initial of the word. As a result, “M”, “S”, and “ken” are acquired. Finally, the acronym synthesizer 23 combines the acronyms according to the word order to obtain a wording candidate “MS Lab”.

  In the configuration example shown in FIG. 4 or 5, the initial acquisition unit 22 acquires only one initial character. For example, the initial 2 is set so that “personal computer” becomes “personal computer”. Sometimes get characters.

  The configuration example illustrated in FIG. 6 is a configuration for creating a wording candidate “Media Solution Lab” from “Media Solution Research Department”, for example. In this case, the vocabulary candidate creation unit 2 may be provided with a word synthesis unit 27 in addition to the morphological analysis unit 21 and the omission rule application unit 24 described above.

  In this configuration example, for example, when the above-mentioned “media solution research department” is input, the morphological analysis section 21 divides this into four nouns “media”, “solution”, “research”, and “part”. Then, it is passed to the omission rule application unit 24. The omission rule application unit 24 refers to the omission creation rule storage unit 25 to generate the words “media”, “solution”, and “ken”. The word synthesizing unit 27 combines these words in the order of words to obtain a wording candidate “Media Solution Lab”.

  Alternatively, in the configuration example illustrated in FIG. 7, the vocabulary candidate creation unit 2 does not perform morphological analysis, creates an abbreviation according to only the abbreviation rules stored in the abbreviation creation rule storage unit 25 in advance, It is the structure to do. In this case, for example, if the omission rule “genetic factor” becomes “gene” is stored in advance in the omission creation rule storage unit 25, when the formal vocabulary “genetic factor” is input, Is created.

As described above, when the vocabulary candidate creation unit 2 creates a wording candidate from the official vocabulary and registers it in the vocabulary database 1, the wording candidate generated and registered by the vocabulary candidate creation unit 2 in the recognition grammar creation system of this embodiment. Is requested to search the large corpus 8. As shown in FIG. 1, for example, the large-scale corpus 8 includes a large number of text corpuses 8b ₁ , 8b ₂ ,... And search units 8a ₁ , 8a ₂ ,. . For example, a WWW site can be used as the text corpus 8b, and a search engine for WWW can be used as the search unit 8a. In FIG. 1, only two combinations of the search unit 8a and the text corpus 8b are illustrated, but the number is arbitrary.

  By the way, if the large-scale corpus 8 is searched in the dark clouds as in the conventional case, there is a possibility that words having the same text as the wording candidate but having a different meaning may hit the search. For example, if “SJH” is used as an abbreviation for “Swedish Junior High School” and “MS Lab” is used as an abbreviation for “Multi-Style Research and Development Center”, There is a risk of misjudging the suitability of candidates.

  Therefore, in the recognition grammar creation system of this embodiment, the search corpus selection unit 7 selects a text corpus to be searched from a large number of text corpuses 8b in the large-scale corpus 8. For example, a specific site on the WWW can be selected as a search target.

  Therefore, the search corpus selection unit 7 includes a candidate corpus registration unit 72 and a candidate corpus database 73 as shown in FIG. The candidate corpus registration unit 72 includes an input / output device (not shown), and allows the user to input a candidate corpus. For example, when the user inputs a URL or the like of a WWW site deemed appropriate as a search target using the input / output device, the candidate corpus registration unit 72 registers the WWW site in the candidate corpus database 73.

As described above, when a text corpus to be searched is registered in the candidate corpus database 73 of the search corpus selection unit 7, thereafter, the search corpus selection unit 7 applies to the text corpus registered in the candidate corpus database 73. And request a search to check the suitability of the wording candidates. That is, for example, if the text corpus 8b ₁ shown in FIG. 2 is registered in the candidate corpus database 73, the search corpus selection unit 7 matches the wording candidates to the search engine 8a ₁ corresponding to the text corpus 8b ₁ . Request a search to find out.

  In the above example, the user determines which text corpus is the search target. However, the search corpus selection unit 7 can also select the text corpus to be searched. . In this case, as shown in FIG. 9, the search corpus selection unit 7 is provided with a candidate corpus search unit 71 in addition to the candidate corpus registration unit 72 and the candidate corpus database 73 described above. The candidate corpus search unit 71 searches the large corpus 8 for a WWW site including at least one of a formal vocabulary (first vocabulary) and a keyword given by the user as a keyword related to the formal vocabulary. Then, the WWW site found as a result of this search is registered in the candidate corpus database 73.

  Note that the search corpus selection unit 7 automatically extracts keywords for selecting a search corpus using a concept hierarchy database (not shown), instead of giving keywords by the user as described above. May be. By using the concept hierarchy database, for example, when the formal vocabulary is “personal computer”, “mainframe”, “workstation”, etc., it is possible to extract “computer” which is a superordinate concept thereof. Then, using this as a keyword, a search target corpus may be selected from the large-scale corpus 8.

Alternatively, when there are a plurality of wording candidates, a text corpus to be searched for examining the suitability of one wording candidate is searched using a combination of the formal vocabulary (first vocabulary) and other wording candidates as a keyword. It is also effective to select by. For example, taking the ID2 entry shown in FIG. 3 as an example, a text corpus for examining the suitability of “Media Solutions Lab” registered as the second vocabulary,
(1) "Media Solution Research Department"
(2) AND search between “Media Solution Research Department” and “Meso Lab” (3) AND search between “Media Solution Research Department” and “MS Research Department” (4) “Media Solution Research Department” and “MS Lab” Is selected from the large-scale corpus 8 by an AND search with “”.

  As described above, the search corpus selection unit 7 selects, for each of the wording candidates, the search target text corpus for determining whether the wording candidate is actually used. One or more are selected from the corpus 8b, and the search engine 8a corresponding to the selected text corpus 8b is requested to search for the wording candidate. The search result is sent to the search result analysis unit 9.

  When the wording candidate exists in the search target text corpus (when a search hit occurs), the search result analysis unit 9 determines that the wording candidate has relevance (that is, the wording candidate is actually used). On the other hand, if the wording candidate does not exist in the search target text corpus, it is determined that the wording candidate is not compatible (not actually used). The determination result is sent to the database update unit 6, and the wording candidate determined to be incompatible is deleted from the entry of the vocabulary database 1 by the data update unit 6.

  For example, as a result of searching for the wording candidate as shown in FIG. 3 using the text corpus 8b selected by the search corpus selection unit 7, the third vocabulary “SJH” of the second entry and the third entry Suppose that the third vocabulary of “Meso Ken” did not make a search hit. In this case, when the data updating unit 6 deletes these wording candidates from the entry, the stored contents of the vocabulary database 1 are as shown in FIG.

  In the above description, the search result analysis unit 9 determines the suitability of the wording candidates according to the presence or absence of the search hit, and the database update unit 6 deletes the wording candidates that are determined not to match from the entry. However, the method for determining the suitability of wording candidates and the method of reflecting the determination result in the vocabulary database in the present invention are not limited to this.

  For example, it is also effective for the search result analysis unit 9 to determine the suitability of the wording candidate according to the appearance information of the wording candidate. Vocabulary appearance information in a text corpus can be viewed as the degree to which the vocabulary is used. Even in this sense, the appearance information is useful as a criterion for determining whether or not to add the vocabulary to the grammar when creating the grammar. Generally, since the recognition rate tends to decrease as the number of vocabulary to be recognized increases, the grammatical vocabulary may be limited, and in such a case, appearance information can be used.

  Further, the phrasing candidate determined to be incompatible (or low in fitness) is left without being deleted from the vocabulary database 1, and the grammar creation unit 3 creates a grammar according to the appearance information when creating the grammar. You may make it determine the wording candidate registered into.

  It is also possible to describe the appearance information as it is in the grammar. When appearance information (generally called a probability in many cases) is added to a word and described in the grammar, it is used by the recognition engine at the time of matching the recognized word. Such appearance information is preferably registered in the vocabulary database 1 together with the vocabulary. In addition, at least the following three types of ways of expressing the appearance information of each wording candidate can be considered, but are not limited to these.

(1) Ratio of the number of appearances of each wording candidate with respect to the total number of search hit words (2) Whether or not the number of appearances of each wording candidate has exceeded a predetermined threshold (binary expression)
(3) The number of occurrences of each wording candidate itself For example, the number of hits when searching the “media solution research department” and the wording candidate using the text corpus 8b selected by the search corpus selecting part 7 is shown in FIG. It shall be as shown.

  As shown in FIG. 11 (b), the search result analyzing unit 9 uses the hit counts shown in FIG. 11 (a) as relative values for the appearance information of each vocabulary with respect to the total number of vocabulary hits (153). calculate. By calculating as a relative value in this way, even when there is a large variation in the number of hits depending on the vocabulary, handling of the calculated value becomes easy. For example, the search result analysis unit 9 determines that the phrase candidates whose appearance information calculated in this way exceeds a predetermined threshold (for example, 10% or more in the case of FIG. 11B) are compatible.

Next, a procedure in which the grammar creation unit 3 creates a recognition grammar from the vocabulary database 1 created as described above will be described. The grammar creation unit 3 reads vocabulary data from the vocabulary database 1 and creates a grammar according to a given grammar specification. For example, in the SRGS specification,
<Rule id = "group name">
<One-of>
<Item> Somubu </ item>
<Item> Some </ item>
</ One-of>
<One-of>
<Item> Soft Ajigyo Honbu </ item>
<Item> Sojihon </ item>
</ One-of>
<One-of>
<Item> Measy Assuryukenkyubu </ item>
<Item> Measy Assuryuken </ item>
<Item> Emues Kenkyubu / item
<Item> Emusuke </ item>
</ One-of>
</ Rule>
</ Grammar>
Create a recognition grammar like

  The recognition grammar created in this way is sent from the grammar creation unit 3 to the grammar storage unit 10 and stored therein. The grammar storage unit 10 is realized by an arbitrary storage device or storage medium. The recognition grammar stored in the grammar storage unit 10 is copied to a storage device or storage medium in a telephone response service system such as a voice portal or an automatic response service system such as a car navigation system with a voice recognition function. Used for voice recognition. Alternatively, the grammar storage unit 10 itself can be removed from the speech recognition grammar creation system of this embodiment and incorporated in these systems.

Note that appearance information may be used when creating a recognition grammar. In this case, for example, if the appearance frequency of each vocabulary is as shown in FIG. 12, the ratio of the appearance frequency 60 of “SOMUBU” to the appearance frequency 40 of “SOMU” is 1 for the first entry. : 1.5, so
<Item weight = "1.5"> Somu </ </ item>
<Item weight = "1"> Som </ item>
become that way. That is, the recognition grammar includes information indicating the priority of each vocabulary obtained from the appearance information.

  When appearance information is also registered in the vocabulary database 1, in order to further improve the accuracy of the recognition grammar, it is preferable to provide a threshold value and create the recognition grammar from words having a threshold value or higher. In this case, as shown in FIG. 13, the grammar vocabulary for selecting the vocabulary to be passed to the grammar creation unit 3 by comparing the appearance information of the vocabulary database 1 and the threshold between the vocabulary database 1 and the grammar creation unit 3. A determination unit 11 may be provided.

  As described above, according to the present embodiment, by selecting the text corpus to be searched, it is possible to improve the accuracy of determining the suitability of wording candidates and to create a highly accurate recognition grammar. Become. Furthermore, by using the appearance information of each wording candidate calculated from the search result, a recognition grammar with higher accuracy can be created.

  In the embodiment described above, the wording candidate generated by the vocabulary candidate creation unit 2 is temporarily registered in the entry of the vocabulary database 1, and the search result analysis unit 9 determines the suitability according to the search result in the large-scale corpus 8. The database update unit 6 deletes the wording candidate determined to be absent from the entry in the vocabulary database 1. However, the present invention is not limited to this, and the wording candidate generated by the vocabulary candidate creation unit 2 is temporarily stored in a memory in the system, and the search result is determined according to the search result in the large-scale corpus 8. Only the wording candidates that are determined to be compatible by the analysis unit 9 may be extracted from the memory by the database update unit 6 and registered in the entry of the vocabulary database 1.

  In addition, the recognition grammar creation system according to the present invention can be configured as a single piece of hardware or a combination of a plurality of pieces of hardware. In the latter case, the communication line for connecting each component may be either wired or wireless, and each component may be located at a remote location.

  In addition to the recognition grammar creation system described in each of Supplementary Notes 1 to 6 below, the recognition grammar creation method and the recognition grammar creation program described in Supplementary Notes 7 and 8, respectively, are also an embodiment of the present invention.

(Supplementary note 1) A vocabulary database that memorizes formal vocabularies and their phrases,
A search corpus selection unit that selects a corpus to be searched for checking suitability of the wording candidate and requests the selected corpus to search for the wording candidate;
A search result analysis unit for analyzing a search result in the corpus selected by the search corpus selection unit;
A database update unit that determines the suitability of the wording candidate based on the analysis result by the search result analysis unit and a predetermined criterion, and stores the wording having relevance among the wording candidates in the vocabulary database;
A recognition grammar creation system comprising: a grammar creation unit that reads a vocabulary from the vocabulary database and creates a recognition grammar of the vocabulary according to grammar specifications.

(Supplementary Note 2) The search corpus selection unit
A candidate corpus database that stores information about candidate corpora that are corpora to be searched;
A candidate corpus search unit that inputs at least one keyword related to the formal vocabulary as a search keyword for the candidate corpus and searches for information on the candidate corpus on the Internet according to the search keyword;
A candidate corpus registration unit that registers information obtained by the candidate corpus search unit in the candidate corpus database;
The recognition grammar creation system according to appendix 1, wherein the search corpus selection unit refers to information registered in the candidate corpus database and selects a corpus to be searched.

  (Additional remark 3) The said search result analysis part is provided with the appearance information generation part which produces | generates the appearance information of the said wording candidate in the said corpus, and the said database update is made into the said analysis result by using the generation result of the appearance information by the said appearance information generation part The recognition grammar creation system according to appendix 1 or 2, which is passed to the department.

(Supplementary Note 4) When the database update unit stores the wording in the vocabulary database, the appearance information generated by the appearance information generation unit for the wording is stored in the vocabulary database in association with the wording,
The recognition grammar creation system further includes:
The recognition grammar creation system according to appendix 3, further comprising a grammar vocabulary determination unit that selects a vocabulary for which a recognition grammar is to be created based on the appearance information from the vocabulary database and passes the selected vocabulary to the grammar creation unit.

  (Additional remark 5) The recognition grammar preparation system of Additional remark 3 or 4 with which the said grammar preparation part produces the recognition grammar which has the information which shows the priority of a word based on the said appearance information.

(Supplementary Note 6) A vocabulary database for storing formal vocabulary words and phrases,
A vocabulary candidate creation unit that creates wording candidates from formal vocabulary;
A search corpus selection unit that selects a corpus to be searched for checking suitability of the wording candidate and requests the selected corpus to search for the wording candidate;
A search result analysis unit for analyzing a search result in the corpus selected by the search corpus selection unit;
A database update unit that determines the suitability of the wording candidate based on the analysis result by the search result analysis unit, and stores the wording having relevance among the wording candidates in the vocabulary database;
A recognition grammar creation system comprising: a grammar creation unit that reads a vocabulary from the vocabulary database and creates a recognition grammar according to grammar specifications.

(Appendix 7) Create wording candidates from formal vocabulary,
Select a corpus to be searched for checking the suitability of the wording candidate, request the selected corpus to search for the wording candidate,
Receive and analyze search results in the corpus selected for search,
Determining the suitability of the wording candidates based on the result of the analysis, and storing the wording having relevance among the wording candidates in the vocabulary database;
A recognition grammar creation method for reading a vocabulary from the vocabulary database and creating a recognition grammar according to a grammar specification.

(Appendix 8) Create wording candidates from formal vocabulary,
Select a corpus to be searched for checking the suitability of the wording candidate, request the selected corpus to search for the wording candidate,
Receive and analyze search results in the corpus selected for search,
Determining the suitability of the wording candidates based on the result of the analysis, and storing the wording having relevance among the wording candidates in the vocabulary database;
A computer program comprising instructions for causing a computer to execute processing for reading a vocabulary from the vocabulary database and creating a recognition grammar according to a grammar specification.

  As described above, according to the present invention, it is possible to improve the accuracy of determining the suitability of wording candidates by selecting a text corpus to be searched, and to create a highly accurate recognition grammar. A creation system can be provided.

The block diagram which shows schematic structure of the recognition grammar preparation system concerning one Embodiment of this invention. Explanatory diagram showing an example of data stored in the vocabulary database Explanatory diagram showing an example of data stored in the vocabulary database Block diagram showing a configuration example of the vocabulary candidate creation unit Block diagram showing a configuration example of the vocabulary candidate creation unit Block diagram showing a configuration example of the vocabulary candidate creation unit Block diagram showing a configuration example of the vocabulary candidate creation unit Block diagram showing a configuration example of the search corpus selection unit Block diagram showing a configuration example of the search corpus selection unit Explanatory diagram showing an example of data stored in the vocabulary database Explanatory diagram showing the relationship between the number of hits and appearance information for each vocabulary Explanatory diagram showing an example of data stored in the vocabulary database A block diagram showing a configuration example for creating a grammar in consideration of appearance information of each vocabulary Flow chart showing conventional recognition grammar creation procedure

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Vocabulary database 2 Vocabulary candidate creation part 3 Grammar creation part 4 Database editing part 5 Synonym database 6 Database update part 7 Search corpus selection part 8 Large-scale corpus 8a Search part 8b Text corpus 9 Search result analysis part 10 Grammar storage part 11 Grammar vocabulary Discriminator 21 Morphological analyzer 22 Initial acquisition unit 23 Initial synthesis unit 24 Omission rule application unit 25 Omission creation rule storage unit 26 Alphabet conversion unit

Claims (8)

A vocabulary candidate creation unit that creates a wording candidate from a formal vocabulary based on a predetermined rule;
A vocabulary database that memorizes the official vocabulary and its phrases,
A search corpus selection unit that selects a corpus to be searched for checking suitability of the wording candidate and requests the selected corpus to search for the wording candidate;
A search result analysis unit for analyzing a search result in the corpus selected by the search corpus selection unit;
A database update unit that determines the suitability of the wording candidate based on the analysis result by the search result analysis unit and a predetermined criterion, and stores the wording having relevance among the wording candidates in the vocabulary database;
A recognition grammar creation system comprising: a grammar creation unit that reads a vocabulary from the vocabulary database and creates a recognition grammar of the vocabulary according to grammar specifications. The search corpus selection unit
A candidate corpus database that stores information about candidate corpora that are corpora to be searched;
A candidate corpus search unit that inputs at least one keyword related to the formal vocabulary as a search keyword for the candidate corpus and searches for information on the candidate corpus on the Internet according to the search keyword;
A candidate corpus registration unit that registers information obtained by the candidate corpus search unit in the candidate corpus database;
The recognition grammar creation system according to claim 1, wherein the search corpus selection unit refers to information registered in the candidate corpus database and selects a corpus to be searched.   The search result analysis unit includes an appearance information generation unit that generates the appearance information of the wording candidate in the corpus, and passes the generation result of the appearance information by the appearance information generation unit to the database update unit as the analysis result. The recognition grammar creation system according to claim 1 or 2. The recognition grammar creation system according to any one of claims 1 to 3, wherein the predetermined rule includes at least one of an empirical rule and an omission rule. When the database update unit stores the wording in the vocabulary database, the appearance information generated by the appearance information generation unit for the wording is stored in the vocabulary database in association with the wording,
The recognition grammar creation system further includes:
Wherein the lexical database, with a grammar vocabulary determination unit selects the vocabulary to create a recognition grammar based on the occurrence information passed to the grammar creation unit, recognition grammar creation system according to claim 3 or 4. The recognition grammar creation system according to any one of claims 3 to 5, wherein the grammar creation unit creates a recognition grammar having information indicating a word priority based on the appearance information. A recognition grammar creation method that creates a recognition grammar using a computer that can access a vocabulary database that stores formal vocabulary words and phrases,
  The vocabulary candidate creation unit of the computer creates a wording candidate from a formal vocabulary based on a predetermined rule;
  A search corpus selection unit of the computer selects a corpus to be searched for checking suitability of the wording candidates, and requests the selected corpus to search for the wording candidates;
  The search result analysis unit of the computer receives and analyzes the search result in the corpus selected as the search target;
  The database update unit of the computer determines the suitability of the wording candidates based on the result of the analysis, and stores the wording having relevance among the wording candidates in the vocabulary database;
  A recognition grammar creation method comprising: a grammar creation unit of the computer reading a vocabulary from the vocabulary database and creating a recognition grammar according to a grammar specification. A computer program for causing a computer capable of accessing a vocabulary database for storing formal vocabulary and phrases to execute a recognition grammar creation process,
  A procedure for creating wording candidates from a formal vocabulary based on predetermined rules,
  A procedure for selecting a corpus to be searched for checking the suitability of a wording candidate and requesting the selected corpus to search for the wording candidate;
  A procedure for receiving and analyzing search results in a corpus selected for search;
  Determining the suitability of the wording candidate based on the result of the analysis, and storing the wording having relevance among the wording candidates in the vocabulary database;
  A computer program for causing a computer to execute a procedure for reading a vocabulary from the vocabulary database and creating a recognition grammar according to a grammar specification.

Images