JP5807891B2 - Language model learning apparatus and computer program - Google Patents

Description

  The present invention relates to a language model learning apparatus that learns a statistical language model by obtaining a large number of natural language word strings from natural language sentences included in a corpus, and more particularly, from a natural language sentence included in a corpus, to a predetermined language language learning apparatus. The present invention relates to a language model learning apparatus capable of automatically generating a language model suitable for the purpose.

  In speech recognition technology, a statistical language model is used. The statistical language model models appearance frequency information of words (or word strings; hereinafter simply referred to as “words”) in a corpus composed of a large amount of natural language sentences. For each word appearing in the corpus, the statistical language model is a list of the words together with their appearance frequency. A language model for a word string (called N-gram) in which N words are connected in a predetermined order is called an N-gram language model. Usually, 1-gram to 3-gram language models are often used.

  In the speech recognition technology, the validity of a phoneme string obtained as a result of speech recognition is verified using a language model. Of phoneme strings with high acoustic likelihood, only hypotheses with high likelihood calculated by the language model are adopted as candidates for speech recognition results.

  Such a statistical language model is used not only for speech recognition technology but also for machine translation and the like to calculate the validity of translation results.

  In order to create a language model, a large amount of natural language sentences are required. However, there is no need for a large amount of natural language sentences. It is desirable to construct a language model from an appropriate natural language sentence corresponding to an application in which the language model is used and a target area to which the language model is applied. For example, if a field to which speech recognition technology is applied is clearly understood, it is desirable to create a language model from natural language sentences related to the field.

  Currently, a large number of machine-readable documents are available. Techniques for selecting natural language sentences necessary for creating a language model from such machine-readable documents have been developed.

  However, as long as a natural language sentence actually written by a human is used, there is no guarantee that the expression desired to be included in the language model is included therein. Conversely, it is highly possible that a sentence extracted from a large amount of machine-readable documents includes a language expression that is not related to an application to which the target field or language model is applied. Therefore, for example, even if a certain field or application is clearly conscious, there is a problem that it is difficult to actively build a language model suitable for the field or application.

  After all, in the conventional technology, given a corpus, the language model modeling technology itself is devised to ensure the performance of the language model.

  For example, Patent Document 1 listed below discloses smoothing in a language model having appearance frequency information of words 3-gram, 2-gram, and 1-gram, for example. In the 3-gram language model, there is a high possibility that 3-grams having an appearance frequency of 0 appear due to lack of learning data. If such a language model is used as it is, the likelihood of the word string of the speech recognition result cannot be correctly evaluated. Smoothing is a technique to alleviate such problems.

  In the technique of Patent Document 1, the types of language models that can be used for smoothing and their dependency relationships (both are referred to as “dependency relationships”) are stored in a storage device in advance. For example, if the total number of words appearing in the corpus used for learning is smaller than a predetermined threshold value, another language model can be used due to its dependency, for example, to smooth the 3-gram language model .

JP 2009-145775 A

  However, the technique disclosed in Patent Literature 1 described above cannot improve the performance of the language model if the original learning data itself is biased or if there are a shortage of samples related to the target field. That is, a more essential solution is to solve the problem of how to prepare learning data suitable for the field or application to which the language model is applied.

  Ideally, it is only necessary to prepare learning data including all sentences that may be issued in the target field or application, and create a language model based on the learning data. At present, it seems that the total amount of data on the Web is closest to such learning data. However, as described above, even if it is data on the Web, as long as it is created by humans, the total number is limited, and it includes all sentences that may be issued in the target field or application. It can never be. Thus, the problem is how to efficiently collect natural language sentences that contain as much text as possible in the target field or application.

  Therefore, an object of the present invention is to provide a language model learning apparatus capable of efficiently generating a language model that has a relatively high probability of being assigned to a natural language word string that may be issued in a target field or application. Is to provide.

  The language model learning device according to the first aspect of the present invention is used together with a machine-readable corpus storage unit that stores a corpus including a plurality of natural language sentences, and learns a language model suitable for a specific application from the corpus. It is a language model learning device for This apparatus includes a template storage means for storing a word string template prepared in advance for a specific application, and a word for extracting a word string pattern matching the word string template stored in the template storage means from the corpus A word string pattern extracted by the word string extraction means based on a string extraction means and a word string transformation rule prepared in advance so that a word string in a natural language in a format according to a preselected purpose is generated; Deformation means for deforming, and learning means for learning a language model using the word string output by the deformation means as learning data.

  A word string template is prepared in advance in the template storage means, and a word string pattern that matches the word string template is extracted from the corpus. The word string transformation rules are applied to these word string patterns so that natural language word strings having a format in accordance with a preselected purpose are generated. As a result, a new expression that does not exist in the corpus is generated. As a result, regardless of the limitation on the number of word strings included in the corpus, it is possible to generate a natural language word string including many word strings in accordance with the purpose, including expressions that are not included in the corpus.

  Preferably, the template storage means includes a seed template storage means for storing a seed template that is a basic word string template to be extracted from a machine-readable sentence, and each of the seed templates stored in the seed template storage means. In contrast, an extension template generation means for generating an extension template by applying a template extension rule prepared in advance, an extension template generated by the extension template generation means, and a seed template stored in the seed template storage means And an extended template storage means for providing the word string extraction means as a word string template.

  By expanding the seed template with the template expansion rule, more templates than the initially prepared seed template can be generated. The number of extracted word string patterns can be increased from the word string patterns included in the corpus. As a result, it is possible to generate more natural language word strings composed of word strings in accordance with the purpose.

  More preferably, each of the word string templates stored in the template storage means has an array of one or a plurality of variables representing arbitrary word strings satisfying predetermined constraints and text data representing other word strings. Including.

  The predetermined constraint condition may be a word class to which a word represented by each variable belongs. A word string extraction unit that performs morphological analysis on each of a plurality of natural language sentences stored in the corpus, and adds a tag of a word class to which the morpheme belongs to each morpheme and outputs the morpheme string; Each of the word string templates stored in the template storage means is compared with the morpheme string output by the morpheme analyzing means, and the word string template and the morpheme string match except for the variables included in the word string template. And means for extracting from the corpus that the word class of the morpheme at the position corresponding to the variable in the word string template in the morpheme string matches the word class of the variable.

  Preferably, each of the word string templates stored in the template storage means includes one or a plurality of variables representing arbitrary words satisfying a predetermined constraint condition, other word strings, and the interval between these variables and the word strings. Syntax information indicating grammatical relationships.

  More preferably, the predetermined constraint condition is a word class to which a word represented by each variable belongs. A word string extraction unit that performs morphological analysis on each of a plurality of natural language sentences stored in the corpus, and adds a tag of a word class to which the morpheme belongs to each morpheme and outputs the morpheme string; Parses the morpheme sequence output by the morpheme analysis unit, outputs a word string pattern composed of syntax information of a natural language sentence, and a word string template stored in the template storage unit. Each of which is compared with the word string pattern output by the syntax analysis means, and is a portion having a structure that matches the word string template except for variables in the word string pattern output by the syntax analysis means, and If the word class of the word in the position corresponding to the variable in the word string template matches the word class of the variable And means for extracting from the path.

  More preferably, the language model learning device further includes a frequency storage means for storing the appearance frequency of the word string pattern appearing in a predetermined corpus for each word string pattern, and the modified word output from the deformation means For each of the columns, the transformation means by determining and copying the number of times of copying the word string based on the appearance frequency stored in the frequency storage means for the word string pattern that generated the word constituting the word string Frequency adjusting means for adjusting the frequency of appearance of words in the word string output from.

  The language model learning apparatus further configures the word string for each of the frequency storage means for storing the appearance frequency of words appearing in a predetermined corpus and the transformed word string output from the deformation means. Based on the appearance frequency of each word stored in the frequency storage means for the word, the appearance frequency of the word in the word string output from the deformation means is adjusted by determining and copying the number of times the word string is copied. Frequency adjusting means.

A weight may be assigned in advance to each of the seed templates stored in the template storage unit. Each of the extension templates stored in the extension template storage means is assigned a weight smaller than the weight of the seed template that is the basis of the extension template. The language model learning device further copies the word string according to the weight assigned to the word string template used in the word string extraction unit for each of the transformed word strings output from the deformation unit. The frequency adjustment means for adjusting the appearance frequency of the word included in the word string after the deformation outputted from the deformation means.

  The computer program according to the second aspect of the present invention causes a computer to function as each unit of any of the language model learning devices described above.

  A speech recognition device according to a third aspect of the present invention includes any of the language model learning devices described above, a language model storage unit for storing a language model learned by the language model learning device, and a language model storage unit. Speech recognition means for performing speech recognition of input speech by using the stored language model.

1 is a block diagram of a language model learning device according to an embodiment of the present invention. It is a figure which shows the example of a seed template set. It is a figure which shows the example of a template expansion rule. It is a figure which shows the example of a word string deformation | transformation rule. It is a flowchart which shows the control structure of the program for extending a seed template. It is a flowchart which shows the control structure of the program for extracting the word sequence pattern which matches a template from a web corpus. It is a flowchart which shows the control structure of the program for producing | generating and outputting the word string of a natural language of a predetermined form by applying a conversion rule to the extracted word string pattern. 1 is an external view of a computer system that implements a language model learning apparatus according to an embodiment of the present invention. It is a block diagram which shows the hardware constitutions of the computer shown in FIG.

  In the following description and drawings, the same parts are denoted by the same reference numerals. Therefore, detailed description thereof will not be repeated.

[Constitution]
Referring to FIG. 1, a language model learning device 30 according to an embodiment of the present invention has a specific form related to a specific field from a natural language sentence included in a Web corpus 32 including sentences collected from the Web. A learning corpus 34 composed of natural language word strings is generated, and a language model learning module 36 learns a language model using the learning corpus 34 as learning data. With this language model learning device 30, it is possible to construct a language model 38 suitable for speech recognition of a specific form of sentence relating to a specific field. In the present embodiment, as will be described later, a language model is constructed for the speech recognition apparatus 40 that performs speech recognition of the input speech 42 that is a question sentence for a disease and outputs a speech recognition result 44. Note that the web corpus 32, the learning corpus 34, and the language model 38 are all stored in a nonvolatile storage medium such as a hard disk in the present embodiment.

  The language model learning device 30 includes a seed template set storage unit 50 that stores a seed template set that includes a seed template that describes a word string pattern to be filled by a word string to be extracted from the Web corpus 32, a seed template, A template expansion rule storage unit 54 that stores template expansion rules that are referred to in order to generate different types of expansion templates, and a template expansion rule storage unit 54 for each seed template stored in the seed template set storage unit 50 A template extension processing unit 52 for applying an applicable one of the stored template extension rules and outputting an extension template set, and an extension for storing the extension template set output by the template extension processing unit 52 And a PLATES set storage unit 56.

  In this specification, the word string pattern is a syntax describing a grammatical relationship between words constituting a natural language sentence or a natural language word string (hereinafter referred to as a “natural language word string or the like”). Information. For example, a parse tree. Each leaf of the parse tree is associated with a word string included in a natural language word string or the like. The template has a structure similar to that of the word string pattern and is compared with the word string pattern.

  With reference to FIG. 2, the seed template stored in the seed template set storage unit 50 describes a basic syntax structure to be satisfied by the word string pattern extracted from the Web corpus 32. For example, the seed template includes a parse tree specified in advance by the user and a word string corresponding to each node of the parse tree. The seed template can also be automatically generated from a natural language sentence. In the present embodiment, it is assumed that the seed template is manually prepared by a human. In this embodiment, the seed template is described using a so-called regular expression. Various regular expressions are known, but any of them may be used here.

  In the present embodiment, words are arranged at positions corresponding to the leaves of the parse tree of the word string template including the seed template. These words include a variable representing a word that should satisfy a predetermined constraint condition and text data representing a word string other than the variable. The template may not contain any variables. As a template, a symbol indicating that a certain word comes to the beginning or end of a sentence can be described.

  All of the templates (excluding syntax information) shown in FIG. 2 include variables (A, B). These variables are also defined by the above regular expression. For these variables, an attribute such as a word class of a word at a position corresponding to the variable, or a combination thereof is designated. The word class specifies, for example, a disease name, a drug name, a symptom name, a substance name, a place name, a person name, a part-of-speech, a verb utilization form, and a set to which a word belongs when the words are classified by their attributes. A plurality of attributes may be specified for one variable. In that case, information specifying whether the plurality of attributes are in an AND relationship or an OR relationship is also attached to the variable.

  In the example of FIG. 2, if there is a word string pattern “A is caused by B”, the word string pattern (or a sentence including this word string pattern; hereinafter simply referred to as “word string pattern”) is extracted. If the word class “disease name” is designated as the variable A, a word string pattern in which “A is the cause of B” and “A” is the disease name is extracted. The same applies to “B”. Thus, the word string pattern extracted by matching with a specific rule is referred to as an “instance” here.

  The numerical value on the right of each seed template shown in FIG. 2 is the weight assigned to each seed template. Although there may be an embodiment in which no weight is used, in the present embodiment, the appearance frequency (number of times of output to the corpus) of the finally obtained word string is adjusted using this weight. In the present embodiment, this weight range is greater than 0 and less than or equal to 1.

  In the example shown in FIG. 2, all seed templates include two variables, but the seed templates are not limited to this. For example, an expression that does not include any variable, an expression that includes only one variable, an expression that specifies the beginning or end of a sentence, and the like can be used. When representing the beginning or end of a sentence, a character string (tag) representing them is attached to the template.

  Only a specific word can be designated as a seed template, or a part of speech string can be designated.

  Referring to FIG. 3, the template extension rule stored in template extension rule storage unit 54 is a rule for extending the seed template stored in seed template set storage unit 50 to generate an extension template. Template extension rules can also be described using regular expressions.

  For example, consider “A <disease name reason is B” shown in FIG. “<Disease Name>” is a word tag assigned to the variable A, and indicates that a tag “Disease Name” is assigned to the word at this position.

  According to the template expansion rule shown in FIG. 3, from the template “A is the reason for <name of disease> is B”, “the reason for A is B”, “A is caused by B”, “A is generated by B”, A template such as “A is because of B” can be generated. Here, the description of the word class assigned to the variable is not repeated.

  As described above, a large number of template expansion rules are prepared in advance, and new templates (these are referred to as “expansion templates”) can be generated by applying these template expansion rules to the seed template. If a large number of template expansion rules are prepared, a large number of expansion templates can be generated from one seed template, and the number of instances extracted from the Web corpus 32 can be increased.

  Although not shown in FIG. 3, when a thesaurus in which words are arranged according to a semantic structure is used, the number of templates that can be generated by the template expansion rule can be increased. For example, for the word class “drug name”, if the higher word class of this class is “substance name”, if the variable whose word class is “drug name” is in the template, It is also possible to generate a template that replaces the “substance name”. Whether or not such replacement is possible is a matter that depends on the design of the system and also a matter that depends on the setting during operation of the system.

  In the example shown in FIG. 3, the weight of each extended rule is attached to the right side of each extended rule. In the present embodiment, the appearance frequency of the word string pattern finally extracted or generated is adjusted according to the product of the weight assigned to each template expansion rule and the weight of the seed template that is the basis. The template extended by the extension rule is not designated by the user as a seed template. Therefore, in the language model finally obtained, it is better to set the appearance frequency of the instance extracted based on the extended template to be lower than that extracted based on the seed template. Conceivable. Therefore, in the present embodiment, the weight assigned to each template expansion rule is a value greater than 0 and less than 1.

  Referring again to FIG. 1, the language model learning device 30 further uses the extension template stored in the extension template set storage unit 56 to select an instance that matches one of the extension templates from the sentence included in the Web corpus 32. A filter 60 to be extracted, a syntax analysis dictionary 58 to which the filter 60 refers when performing syntax analysis of each sentence in the Web corpus 32, and an extracted sentence corpus comprising instance sentences extracted from the Web corpus 32 by the filter 60 An extracted sentence corpus storage device 62 for storing the extracted sentence, and a word string transformation rule for converting the extracted sentence stored in the extracted sentence corpus storage device 62 into a sentence type corresponding to the target field and application of the finally obtained language model Stored in the word string transformation rule storage unit 64 and the extracted sentence corpus storage device 62. To each sentence, an applicable one of the word string modification rules stored in the word string modification rule storage unit 64 is applied, and a modification module 66 for outputting the transformed sentence and an output from the modification module 66 And a modified word string set storage unit 68 for storing a modified sentence set composed of the modified sentences. The word string deformation rules stored in the word string deformation rule storage unit 64 are also described by regular expressions in this embodiment.

  With reference to FIG. 4, a simple example of a word string transformation rule is shown. Here, as described above, the language model learning device 30 is used to create a language model for recognizing a voice input of a question about a disease. The word string transformation rule shown in (1) of FIG. 4 is a rule for generating a question sentence “Tell me about what causes A” from a word string pattern of the form “B causing A”. . Again, “A” and “B” are variables. As with the template, attributes such as word classes may be specified for the variables. If the variable has an attribute, the word string pattern that includes the attribute and matches the left side of the word string modification rule is transformed into the word string shown on the right side of the word string modification rule.

  In the example shown in FIG. 4, only one each of the left side and the right side of the word string transformation rule is shown. However, the present invention is not limited to such an embodiment. An implementation may be performed in which a plurality of word string modification rules having the same left side and different right sides are combined into one modification rule.

  Referring again to FIG. 1, the language model learning device 30 further includes a frequency calculation module 70 for calculating the appearance frequency of each word string including the structure appearing in the Web corpus 32, and the frequency calculation module 70. A frequency data storage unit 72 that stores frequency data composed of appearance frequencies calculated for a word string, and for each modified sentence stored in the modified word string set storage unit 68, a weight attached to the modified sentence; Based on the frequency data of the word string stored in the frequency data storage unit 72, the number of times of output of the modified sentence is determined, and the word in the corpus finally obtained is output by repeatedly outputting the modified sentence. A frequency adjustment module 74 for adjusting the appearance frequency of the columns. A set of modified sentences output by the frequency adjustment module 74 forms the learning corpus 34.

  In the present embodiment, the frequency data stored in the frequency data storage unit 72 is an appearance probability for each word string by structure.

  Referring to FIG. 5, a program for realizing template extension processing unit 52 in FIG. 1 by a computer has the following control structure. This program executes step 100 for reading all the template expansion rules stored in the template expansion rule storage unit 54 into the main storage device of the computer, and step 104 described below for each rule read in step 100. And a step 106 of outputting the merged seed template and extension template obtained in the process of step 102 as an extension template set and completing the process after the process of step 102 is completed. .

  Step 104 is applied to all seed templates stored in the seed template set storage unit 50, if the extended rule currently being processed is applicable, and creates a new template (extended template) 110. In step 112, the product of the weight of the seed template that is the base and the weight of the extension rule is calculated and added as a weight to the newly created template, and the extension template created in step 112 is Merging with the calculated weights into the seed template 114. If the new template created in step 110 has already been merged with the seed template, the template is not added to the seed template.

  Referring to FIG. 6, the program for realizing filter 60 shown in FIG. 1 includes a step 130 for executing the following step 132 for each sentence stored in Web corpus 32.

  Steps 130 and 141 perform morphological analysis and syntax analysis, respectively, with reference to the syntax analysis dictionary 58 with respect to the sentence to be processed, and word classes, usage forms, etc. by morphological analysis and syntax analysis processing. Step 142 for receiving a word string pattern consisting of syntax information having a word string (morpheme string) with a tag indicating, and executing the following step 144 for each template stored in the extended template set storage unit 56. Including. Here, since the target language is Japanese, in step 140, syntax analysis including morphological analysis is performed. In the case where the target language is a language that separates words with spaces such as English, morphological analysis is not necessary here, and syntax analysis processing including analysis processing such as part-of-speech analysis may be executed. For the morphological analysis, an existing morphological analysis program may be used. As a morphological analysis program, for example, JUMAN (URL = http: //nlp.kuee.kyoto-u.ac.jp/nl-resource/juman.html) or ChaSen (URL = http: //chasen-legacy.sourceforge .jp /) can be used. There are two types of syntax analysis processing: dependency analysis and phrase structure analysis. Any of these may be used, but in the present embodiment, dependency analysis is used. An existing Japanese syntax analysis system KNP (URL = http: //nlp.kuee.kyoto-u.ac.jp/nl-resource/knp.html) may be used.

  Step 144 is a step 150 for determining whether or not there is a portion that matches the template to be processed in the word string pattern to be processed, and when the determination in step 150 is affirmative, And a step 152 for outputting the template with the weight of the template to be processed added to the matching part, and moving the process to the next template. If the determination in step 150 is negative, nothing is done and the process moves to the next template. The criterion for determining the match here is that the syntax information constituting the word string pattern includes a part that matches the syntax information of the word string template except for variables, and the word string template is included in the part. Whether or not the word class of the word at the position corresponding to the variable in the syntax information matches the word class of the variable. In this determination, if even a part of the word string pattern to be processed matches the word string pattern, it is output. In addition, when not using syntax information and using only a morpheme string (word string), the word string template and the morpheme string match except for the variables included in the word string template, and the word string in the morpheme string It may be based on whether or not the word class of the morpheme at the position corresponding to the variable in the template matches the word class of the variable.

  Referring to FIG. 7, the program for realizing the transformation module 66 and the frequency adjustment module 74 of FIG. 1 performs the following step 182 for each word string included in the sentence stored in the extracted sentence corpus storage device 62. Step 180 is executed.

  Step 182 includes a step 200 of executing the following step 202 for each word string modification rule stored in the word string modification rule storage unit 64.

  Step 202 determines whether or not the word string pattern to be processed matches the left side of the modification rule to be processed. If there is no match, step 210 proceeds to the next modification rule, and the determination in step 210 is performed. If the determination is affirmative, step 212 for generating a new word string by modifying the word string pattern to be processed according to this deformation rule, and the pattern of the word included in the word string pattern to be processed following step 212 Appearance frequency (the product of words when there are multiple words) in (word string structure), the weight assigned to the word string pattern to be processed, and the number of sentences included in the learning corpus 34 finally obtained Step 214 for calculating the product of a predetermined constant to adjust the value, and the number of times determined by the integer part of the value calculated in step 214. Tsu repeated a word string obtained by the flop 212 and a step 216 to be output. When the value calculated in step 214 is less than 1, in the present embodiment, in step 216, the transformed word string is output only once.

[Operation]
The language model learning device 30 shown in FIGS. 1 to 7 operates as follows. A large number of sentences are collected in advance from the Web and stored in the Web corpus 32 in advance. The frequency calculation module 70 performs morphological analysis and syntax analysis for each sentence included in the Web corpus 32 in advance, calculates the appearance frequency for each word structure for each word, and stores it as frequency data in the frequency data storage unit 72. Let This process is almost the same as learning a normal language model.

  In the present embodiment, a user creates a seed template, a template expansion rule, and a word string modification rule in advance, and stores them in the seed template set storage unit 50, the template expansion rule storage unit 54, and the word string modification rule storage unit 64, respectively. Remember. All of these use regular expressions. In these fields, the language model 38 that is finally obtained is applied to which field, and the creation policy is determined depending on what application is used. However, what template and rule are finally created is determined by the user's selection.

  When the seed template and the template expansion rule are prepared, the template expansion processing unit 52 operates, and the template expansion rule stored in the template expansion rule storage unit 54 is added to each of the seed templates stored in the seed template set storage unit 50. Apply to extend the template. A number of templates are generated by this extension and stored in the extension template set storage unit 56.

  Further, a word string deformation rule is generated and stored in the word string deformation rule storage unit 64 in advance. The syntax analysis dictionary 58 is prepared in a format suitable for the morphological analysis and syntax analysis program used in the filter 60.

  All the extension templates stored in the extension template set storage unit 56 are read in advance and stored in a main storage unit (not shown). The filter 60 reads a sentence from the Web corpus 32 and performs morphological analysis and syntax analysis on each of them (step 140 in FIG. 6). Further, the filter 60 stores the word string pattern (syntax information consisting of morpheme strings with tags such as word classes and semantic classes and the parse tree having them as leaves) obtained by the syntax analysis in the main storage unit. It is determined whether or not there is a portion that matches the stored extension template (step 150). If there is a portion that matches any of the extended templates (Yes at step 150), the filter 60 outputs the word string pattern to the extracted sentence corpus storage device 62 together with the weight assigned to the matched template (step 152). ). The extracted sentence corpus storage device 62 stores these word string patterns together with tags and weights attached to the words. The filter 60 repeats this for all sentences stored in the web corpus 32.

  The deformation module 66 applies the word string deformation rule storage unit 64 to each word string pattern stored in the extracted sentence corpus storage device 62. That is, the deformation module 66 calls a deformation rule for each word string pattern to be processed, and determines whether or not the word string pattern matches the left side of the deformation rule (step 210). When the word string pattern matches the left side of the deformation rule (Yes in Step 210), the deformation module 66 deforms the word string pattern according to the right side of the deformation rule, and generates a word string from the word string pattern ( Step 212). The modified word string is stored in the modified word string set storage unit 68. The frequency adjustment module 74 uses the product of the appearance frequency of words appearing in the word string, the weight of the template applied in the filter 60 applied to the word string, and a predetermined weight as the weight of the modified word string. Calculated as a product with a constant (step 214). The frequency adjustment module 74 repeatedly outputs the modified word string by the number of times of the integer part of the weight thus calculated (step 216). All of the output modified sentences are stored in the learning corpus 34. When the process of step 216 is completed, the deformation module 66 executes a process according to the next deformation rule. If the determination in step 210 is negative, the modification module 66 does nothing with the modification rule, and executes the process according to the next modification rule.

  In this way, when the modification module 66 and the modified word string set storage unit 68 apply all the modification rules for a certain word string pattern, the same process is executed for the next word string pattern.

  When all the transformation rules are applied by the transformation module 66 and the frequency adjustment module 74 for all the word string patterns, the processing ends.

  In this way, the learning corpus 34 is created. The learning corpus 34 is a sentence that is deformed by a word string modification rule prepared in advance from a sentence that matches the initially prepared seed template and a sentence that is extended from the seed template and that matches the extended template related to the seed template. Consists of. The extension template includes synonyms of words included in the seed template, paraphrasing expression of the seed template, and the like. The word string transformation rule assumes a sentence pattern often used in an application in which a language model as a final target is used. Accordingly, the learning corpus 34 includes many words or their synonyms that often appear in utterances related to a specific field, and phrases often used in a specific application. Moreover, since the seed template is expanded by the template expansion rule, the extended template set storage unit 56 includes a very large number of templates. Moreover, since a regular expression is used for this template, a very large number of word string patterns (morpheme strings) are extracted from the web corpus 32 by matching the template with sentences included in the web corpus 32. Here, the word “extraction” is used, but paraphrasing is also accepted as a template, so that expressions not included in the Web corpus 32 are also extracted by the processing of the filter 60.

  The Web corpus 32 is considered to contain the largest number of expressions as an available corpus. However, the expressions included in the Web corpus 32 are created by humans, and therefore the number is inevitably limited. On the other hand, as in the present embodiment, the deformation module 66 is manually created by extending the template, matching with the Web corpus 32 using various extended templates, and changing the expression variously using the extended templates. A much wider representation will be stored than what has been done. Therefore, the language model 38 learned by using the learning corpus 34 generated using these expressions as learning data is adapted to the originally intended field or application, and includes expressions that are not included in the Web corpus 32. Appearance probabilities can be calculated for a very wide range of expressions. As a result, the speech recognition using the language model 38 can realize a highly accurate recognition rate for the field and application intended when the seed template and the word string transformation rule are created.

  However, it will be apparent to those skilled in the art that the Web corpus 32 used in the present invention is not limited to those containing only sentences collected from the Web. As the web corpus 32, a web corpus collected from the web and a sentence obtained from another source may be used, or a corpus that does not include a sentence collected from the web may be used.

  In the above-described embodiment, the template expansion by the template expansion processing unit 52 is only for the seed template. However, the present invention is not limited to such. You may make it increase the number of templates further by applying a template expansion rule to the expansion template obtained by applying a template expansion rule with respect to a seed template. In this case, the template expansion process may be performed a predetermined number of times, or the template expansion process may be repeatedly performed until no new expansion template appears.

  In the example shown in FIG. 4, only a deformation rule for deforming one word string pattern to generate one new word string is shown. However, the present invention is not limited to such an embodiment. For example, a rule that creates a new word string from two word string patterns by including a description referring to another word string pattern in the rule may be used.

  For example, when there are a word string pattern ending with a certain word and a word string pattern beginning with the same word in the transformed set of word string patterns, the two word string patterns are A new word string can be created by connecting each other around a common word. For example, consider a case where a word string pattern “capital of Bolivia” is extracted as an instance of a template “B of A” and an instance “where is the capital” is extracted for a template “where is X”. The last word of the former and the first word of the latter are both “capital”. In such a case, a new word string “where is the capital city of Bolivia” can be generated by connecting both of them with the capital city as the center.

  As another example, when a variable does not exist in the template, simply connecting the templates to each other is also performed as a kind of modification. For example, let us consider a case in which something like “It is (end of sentence)” exists in the template, and there is an expression “It is a capital” in the transformed word string. In this case, the instance “capital of Bolivia” is directly connected to the expression “I am”, and the expression “I am the capital of Bolivia” is also generated as a modified word string.

  For such processing, it is necessary to prepare a modification rule for that purpose separately from the modification rule stored in the word string modification rule storage unit 64. After the process shown in FIG. 7 is completed, the modified word string may be further processed according to these rules.

  In the above embodiment, the frequency adjustment module 74 adjusts the number of copies of the modified sentence by the product of the weight assigned to the modified sentence and the product of the appearance probabilities of the words included in the modified sentence. However, the present invention is not limited to such an embodiment. For example, all the weights assigned to the deformed sentences may be equal values. Further, instead of the product of the appearance probabilities of all the words included in the modified sentence, for example, the appearance probability of only a noun may be used.

  In the above embodiment, various weights are given to the seed template in advance. However, the present invention is not limited to such an embodiment. The weight assigned to the seed template may be constant, and the template weight may be determined only by which template expansion rule is used. Alternatively, when there are a plurality of applicable templates for the word string included in the web corpus 32, a weight that increases according to the number of templates may be given. When a template is created by applying the template expansion rule not only to the seed template but also to the expansion template, it is desirable that the weight of the template is reduced each time the expansion template is applied.

  Furthermore, when filtering by the filter 60, the template weight applied at the time of extraction is only attached to the extracted word string in the above embodiment. However, the present invention is not limited to such an embodiment. For example, the weight may be changed depending on how large a portion of the sentence to be processed matches a template. In this case, it is desirable that the weight is not changed when the whole sentence matches one template, and the weight becomes smaller as the ratio of the matched portion to the whole sentence becomes smaller.

  In the above embodiment, no weight is given to the word string deformation rules stored in the word string deformation rule storage unit 64. However, the present invention is not limited to such an embodiment. For example, a weight greater than 0 and less than or equal to 1 may be assigned to the word string modification rule in advance, and the weight assigned to the matched sentence may be multiplied by this weight to obtain the weight of the sentence after modification.

  In the above embodiment, each rule is described using regular expressions. However, the present invention is not limited to such an embodiment. As long as the rules can be accurately described according to the purpose, any description method may be used.

  Further, in the above-described embodiment, syntax analysis is performed on each sentence of the corpus. However, the present invention is not limited to such an embodiment, and only morphological analysis may be performed. In this case, a morpheme string arranged one-dimensionally is obtained, but if this is also regarded as a kind of structure, the process of the above embodiment can be applied as it is as the subsequent process.

[Realization by computer]
The language model learning device 30 according to this embodiment can be realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware.

  Referring to FIG. 8, this computer system 330 includes a computer 340 having an FD (flexible disk) drive 352 and a CD-ROM (compact disk read only memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. including.

Referring to FIG. 9, in addition to FD drive 352 and CD-ROM drive 350, computer 340 includes CPU (central processing unit) 356 and bus 366 connected to CPU 356, FD drive 352 and CD-ROM drive 350. And a read only memory (ROM) 358 for storing a boot-up program and the like, and a random access memory (RAM) 360 connected to the bus 366 for storing a program command, a system program, work data, and the like. The computer system 330 further includes a network interface (I / F) 344 that provides a connection to the Internet. Although not shown, the computer 340 is connected to the mobile phone network via the network I / F 344 and can perform data communication with the mobile phone 300.

  A computer program for causing the computer system 330 to operate as the language model learning device 30 is stored in the CD-ROM 362 or FD 364 inserted in the CD-ROM drive 350 or FD drive 352 and further transferred to the hard disk 354. . Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, from the FD 364, or via a network.

  This program includes a plurality of instructions for causing the computer 340 to operate as the language model learning device 30 of this embodiment. Some of the basic functions required to perform this operation are provided by operating system (OS) or third party programs running on the computer 340 or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary for realizing the system and method of this embodiment. This program includes only instructions that execute the operation as the language model learning apparatus 30 by calling an appropriate function or “tool” in a controlled manner so as to obtain a desired result. Just go out.

  The web corpus 32, the seed template set storage unit 50, the template extension rule storage unit 54, the extension template set storage unit 56, the syntax analysis dictionary 58, the extracted sentence corpus storage unit 62, and the word string modification rule storage unit shown in FIG. 64, the modified word string set storage unit 68, the frequency data storage unit 72, the learning corpus 34, the language model 38, and the like are all realized by the hard disk 354 or the RAM 360 shown in FIG. In particular, for example, areas such as the Web corpus 32, the seed template set storage unit 50, the template expansion rule storage unit 54, the extracted sentence corpus storage unit 62, and the word string modification rule storage unit 64 are usually reserved in the hard disk 354. When the program is executed, necessary information is read from these areas and loaded into the RAM 360 as necessary. Data stored in the extended template set storage unit 56, the extracted sentence corpus storage unit 62, the modified word string set storage unit 68, etc. has a work file-like character. Therefore, it is generated in the RAM 360 at the time of generation, and stored in the hard disk 354 if it is necessary to store it. The same applies to the learning corpus 34 and the language model 38.

  The operation of computer system 330 is well known and will not be repeated here.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are included. Including.

30 Language Model Learning Device 32 Web Corpus 34 Learning Corpus 36 Language Model Learning Module 38 Language Model 50 Seed Template Set Storage Unit 52 Template Extension Processing Unit 54 Template Extension Rule Storage Unit 56 Extended Template Set Storage Unit 58 Syntax Analysis Dictionary 60 Filter 64 Word string transformation rule storage unit 66 transformation module 74 frequency adjustment module

Claims (10)

A language model learning device that is used together with a machine-readable corpus storage means for storing a corpus including a plurality of natural language sentences, and for learning a language model suitable for a specific application from the corpus,
Template storage means for storing a word string template prepared in advance for the specific application;
A word string extraction means for extracting from the corpus a word string pattern that matches a word string template stored in the template storage means;
Deformation means for deforming the word string pattern extracted by the word string extraction means based on a deformation rule prepared in advance so that a natural language word string in a format according to a preselected purpose is generated When,
Look including a learning means for performing learning of the language model word train output by said deformation means as learning data,
The template storage means includes
Seed template storage means for storing a seed template that is a basic word string template to be extracted from a machine-readable sentence;
Extension template generation means for generating an extension template by applying a template extension rule prepared in advance for each of the seed templates stored in the seed template storage means;
An extension template storage means for storing the extension template generated by the extension template generation means and the seed template stored in the seed template storage means and providing the word string extraction means as the word string template , language model learning device. The language model learning device according to claim 1 ,
Each of the word string templates stored in the template storage means includes an array of one or a plurality of variables representing arbitrary words satisfying predetermined constraint conditions and text data representing other word string patterns. Model learning device. The language model learning device according to claim 2 ,
The predetermined constraint condition is a word class to which a word represented by each variable belongs,
The word string extraction means includes
Morphological analysis of each of the plurality of natural language sentences stored in the corpus, morpheme analysis means for attaching a tag of a word class to which the morpheme belongs to each morpheme and outputting it as a morpheme string;
Each of the word string templates stored in the template storage unit is compared with the morpheme string output by the morpheme analyzing unit, and the word string template and the morpheme string are identical except for variables included in the word string template. And a means for extracting, from the corpus, a word class of a morpheme in a position corresponding to a variable in a word string template in the morpheme string, which matches the word class of the variable. Model learning device. The language model learning device according to claim 1 ,
Each of the word string templates stored in the template storage means includes one or a plurality of variables representing arbitrary words satisfying predetermined constraint conditions, other word strings, and grammatical relationships between these variables and the word strings. A language model learning apparatus including a word string pattern including syntax information indicating a relationship. The language model learning device according to claim 4 ,
The predetermined constraint condition is a word class to which a word represented by each variable belongs,
The word string extraction means includes
Morphological analysis of each of the plurality of natural language sentences stored in the corpus, morpheme analysis means for attaching a tag of a word class to which the morpheme belongs to each morpheme and outputting it as a morpheme string;
A syntax analysis means for performing a syntax analysis on the morpheme sequence output by the morpheme analysis means, and outputting a word string pattern composed of syntax information of the natural language sentence;
Each of the word string templates stored in the template storage means is compared with the word string pattern output by the syntax analysis means, and the variables in the word string patterns output by the syntax analysis means are excluded. A portion having a structure that matches the word string template, and the word class of the word at a position corresponding to the variable of the word string template in the portion matches the word class of the variable A language model learning device including means for extracting from the corpus. The language model learning device according to any one of claims 4 to 5 , further comprising:
A frequency storage means for storing the appearance frequency of a word string pattern appearing in a predetermined corpus for each word string pattern;
Provided between the deformation means and the learning means, receives the transformed word string output from the deformation means, and stores the frequency for the word string pattern that generated the word string for each of the word strings Based on the appearance frequency stored in the means, the number of times of copying the word string is determined, copied, and repeatedly output to the learning means, thereby adjusting the appearance frequency of the word in the word string output from the deformation means A language model learning device including frequency adjusting means for performing the operation. The language model learning device according to any one of claims 1 to 5 , further comprising:
A frequency storage means for storing the appearance frequency of words appearing in a predetermined corpus;
It is provided between the deforming means and the learning means, receives the transformed word string output from the deforming means, and stores the word constituting the word string in the frequency storage means for each of the word strings. Based on the appearance frequency of each stored word, the number of times of copying of the word string is determined, copied, and repeatedly output to the learning means, whereby the appearance frequency of words in the word string output from the deformation means is determined. A language model learning device including frequency adjusting means for adjusting. The language model learning device according to claim 1 ,
A weight is assigned in advance to each of the seed templates stored in the template storage means,
Each of the extension templates stored in the template storage means is assigned a weight smaller than the weight of the seed template that is the basis of the extension template,
The language model learning device further includes:
A word string template provided between the deforming means and the learning means, receiving the transformed word string output from the deforming means, and using the word string template used in the word string extracting means for each of the word strings. Frequency adjusting means for adjusting the appearance frequency of the word included in the modified word string output from the deforming means by copying the word string and repeatedly outputting it to the learning means according to the assigned weight A language model learning device. A computer connected to a machine-readable corpus storage means for storing a corpus including a plurality of natural language sentences;
Template storage means for storing a word string template;
A word string extraction means for extracting from the corpus a word string pattern that matches a word string template stored in the template storage means;
Deformation means for deforming the word string pattern extracted by the word string extraction means based on a deformation rule prepared in advance so that a natural language word string in a format according to a preselected purpose is generated When,
Function as learning means for learning a statistical language model using a set of word strings output by the deformation means as learning data ;
The template storage means includes
Seed template storage means for storing a seed template that is a basic word string template to be extracted from a machine-readable sentence;
Extension template generation means for generating an extension template by applying a template extension rule prepared in advance for each of the seed templates stored in the seed template storage means;
An extension template storage means for storing the extension template generated by the extension template generation means and the seed template stored in the seed template storage means and providing the word string extraction means as the word string template , Computer program. The language model learning device according to any one of claims 1 to 8 ,
Language model storage means for storing a language model learned by the language model learning device;
A speech recognition apparatus including speech recognition means for performing speech recognition of input speech by using a language model stored in the language model storage means.

Images