JP5193798B2 - Dictionary creating device, dictionary creating method, dictionary creating program, and recording medium recording dictionary creating program - Google Patents

Description

  The present invention provides a dictionary for use in a question answering system in which a computer answers a question expressed in a natural language, an information retrieval system, an information extraction system, an automatic summarization system, an automatic translation system, an automatic paraphrasing system, a speech recognition system, etc. The present invention relates to a device, a dictionary creation method, a dictionary creation program, and a recording medium on which the dictionary creation program is recorded.

  In a conventional language processing device, a probabilistic model is assumed for the case frame of the predicate, machine learning is performed using corpus data in which the correct answer of the case frame is manually assigned, and the parameters of the probability model are estimated and determined. An apparatus that outputs the most likely term structure using a probabilistic model has been proposed (see, for example, Non-Patent Document 1). In this method, when a predicate is a word in a sentence and a word for a term for the predicate is given, the case frame with the level of the semantic attribute of the word is expressed in information theory. It is a method to check whether the rule is powerful. It does not deal with term recognition or zero pronoun analysis for a given text.

  Further, as disclosed in Non-Patent Document 2, a method of learning a probability model using a large amount of text corpus, performing case analysis, and determining a predicate term structure has been proposed. , We do not deal with zero pronouns whose terms appear in a sentence different from the sentence containing the predicate.

A method for identifying zero pronouns is described in Non-Patent Document 3.
Lee Wang, Naoki Abe, “Learning Dependencies Between Case Slots”, Information Processing Society of Japan Research Report, Natural Language Processing Research Report, Vol. 96, no. 114, PP. 93-99 Daisuke Kawahara and Ikuo Kurohashi, “Integrated probabilistic model of syntax and case analysis based on large-scale case frames acquired from the Web”, Proc. Of the 12th Annual Conference of the Language Processing Society, PP. 1111-1114, 2006 Ryu Iida, Kentaro Inui, “Identification of antecedents of Japanese zero pronouns by machine learning considering contextual cues”, IPSJ Journal, Vol. 45, no. 3, PP. 906-918 Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiroaki Nakaiwa, Kentaro Ogura, Yoshifumi Oyama, Yoshihiko Hayashi, Japanese Vocabulary System, Iwanami Shoten (1997) Taku Kudo, Yuji Matsumoto, “Japanese Dependency Analysis by Chunking Stage Application”, IPSJ Journal, Vol. 43, no. 6, pp. 1834-1842, 2002 Vladimir Vapnik, “The Nature of Statistical Learning Theory”, 2nd Edition, Springer (1999)

  In a conventional language processing device that outputs a predicate term structure, even if case frame information is registered in the dictionary, if there are multiple verb and noun usages, which usage case information is used for analysis There is no clear standard, and manual adjustment is necessary. The adjustment is very labor intensive and it is difficult to find an adjustment method that improves the analysis accuracy by adjustment. .

  Therefore, Non-Patent Document 2 proposes a method for automatically constructing a predicate term structure probability model from a large-scale text corpus. However, this method does not deal with zero pronouns in which terms appear in a sentence different from the sentence containing the predicate, and it is difficult to perform a highly accurate predicate term structure analysis when multiple sentences are given. there were.

  Non-Patent Document 3 deals with a zero pronoun identification method, but does not deal with predicate term structure analysis. Also, the case where the predicate is converted into a noun and exists in a compound noun is not handled. The problems of zero pronouns and compound nouns affect each other in the predicate term structure analysis, and even if they are handled in order, the overall predicate term structure analysis accuracy may be lowered.

  There has been no method to handle predicate term structure analysis in a unified manner including zero pronouns and compound nouns.

  The present invention solves the above-mentioned problems, and its purpose is to automatically learn a term decision rule for judging a term with high accuracy and to perform predicate term structure analysis including zero pronouns and compound nouns. An object is to provide a dictionary creation device, a dictionary creation method, a dictionary creation program, and a recording medium on which the dictionary creation program is recorded, which can be handled in a unified manner.

  In order to solve the above-mentioned problems, the present invention provides a predicate or a behavioral noun by using a machine learning method based on a text in which a correct term structure is manually tagged for a predicate and a behavioral noun. Predicate or operability based on the basic form, part of speech, semantic category, whether it is a functional word, whether it is a symbol, whether it is a symbol, dependency between clauses, predicate state, etc. A term decision rule for judging a term for a noun (hereinafter collectively referred to as a “predicate”) is automatically learned and output.

That is, the dictionary creation device according to claim 1, for each word included in the text to be analyzed described in a natural language, the notation of the word and the basic form of the word in order from the beginning of the text to be analyzed. The clause to which the word belongs, the state of the predicate in a dependency relationship with the word, information indicating whether the word is a predicate or a behavioral noun, and the word is a pre-description If there is a term relationship to a word or behavioral noun, the correct term structure is stored in the syntax / semantic analysis result table, and each predicate or behavioral noun in the syntax / semantic analysis result table is stored. In addition, when a word other than the predicate or behavioral noun in the syntax / semantic analysis result table that satisfies the constraint condition starting from the predicate or behavioral noun is searched for, the word is the most from the predicate or behavioral noun. Appears nearby The constraint condition that becomes a word is extracted as an attribute to create a training attribute index table, and for each word other than the predicate or action noun, a value indicating whether the word applies to each of the constraint conditions Training to create a training vector table composed of training vectors configured, and teacher variables for identifying the correct term structure and other term structures for each word other than the predescription word or the action noun With respect to the training vector and the teacher variable described in the training vector table, data creation means, and data consisting of a set of a teacher variable and a training vector indicating a correct term structure for each type of term structure is a positive example The distance between two parallel hyperplanes that divide the positive example and the negative example is the maximum, with data consisting of pairs of training variables and training vectors indicating term structures other than the correct answer as negative examples. A weight representing the importance of the attribute is learned using a machine learning method based on the obtained hyperplane, and the weight obtained by the learning is added to the training attribute index table. A weight learning means for creating a weight table, and a list of rules with reference to the weight table, in which attributes are rearranged in descending order of importance, as a term determination rule for determining a term structure for a basic form of a predicate or a behavioral noun The information indicating the correct term structure is information indicating the correct term structure for the basic form of the predicate or action noun included in the text to be analyzed, and the constraint condition (A) the search range or search direction of the word based on the phrase containing the predicate or behavioral noun as the origin, and (b) the predicate or behavior name as the origin. And the presence or absence of modification relationship between the phrase and the predicate or clause that contains the word other than the operation of nouns including, (c) state predicate in the predicate or receiving relates the word other than the operation of a noun related Whether or not the above information is taken into account, and the output term determination rule is a dictionary.

In addition, the dictionary creation method according to claim 2 includes: for each word included in the text to be analyzed described in a natural language, the notation of the word and the basic form of the word in order from the beginning of the text to be analyzed. The clause to which the word belongs, the state of the predicate in a dependency relationship with the word, information indicating whether the word is a predicate or a behavioral noun, and the word is a pre-description A dictionary creation method in a device having a syntax / semantic analysis result table storing a correct answer term structure and a term structure for a word or action noun, wherein training data creation means For each predicate or behavioral noun in the syntax / semantic analysis result table, a word other than the predicate or behavioral noun in the syntax / semantic analysis result table that satisfies the constraint condition starting from the predicate or behavioral noun When searching for The constraint condition such that the word becomes a word appearing closest to the predicate or behavioral noun is extracted as an attribute to create a training attribute index table, and for each word other than the predicate or behavioral noun, Identify training vectors composed of values indicating whether or not a word meets each of the above-mentioned constraints, and the correct term structure and other term structures for each word other than the predescription word or action noun A training vector table consisting of a teacher variable for performing the training, and a weight learning means for each of the types of term structures for the training vector and the teacher variable described in the training vector table. Data consisting of a combination of a teacher variable indicating a term structure and a training vector is a positive example, and data consisting of a combination of a teacher variable indicating a term structure other than the correct answer and a training vector is used. As an example, a hyperplane that maximizes the distance between two parallel hyperplanes that divide a positive example and a negative example is obtained, and the importance of the attribute is determined using a machine learning method based on the obtained hyperplane. A weight learning step of learning a weight to represent, adding a weight obtained by the learning to the training attribute index table to create a weight table, and a term determination rule creating means, referring to the weight table, Outputting a list in which attributes are sorted in descending order of attributes as a term determination rule for determining a term structure for a basic form of a predicate or a behavioral noun, and the information indicating the correct term structure is analyzed This is the information indicating the correct term structure of the basic form of the predicate or behavioral noun contained in the target text, and the constraint conditions are based on the phrase containing the starting predicate or behavioral noun (a) And when the, the search range or search directions of the single word, (b) modification relationship between clauses containing the word other than the phrase containing a predicate or operability nouns serving as the starting point the predicate or operability nouns And (c) whether or not to consider information on the state of predicates that are in a dependency relationship with a word other than the preceding descriptive word or action noun, and the output term determination rule is a dictionary It is characterized by.

Also, the dictionary creation program according to claim 3 is a dictionary creation program to function as the respective means described computer to claim 1.

A recording medium according to a fourth aspect is a computer-readable recording medium in which the dictionary creating program according to the third aspect is recorded.

  According to the above configuration, for text that is tagged manually with the correct term structure, the predicate or behavioral noun, the basic form of the word included in the text, the part of speech, the semantic category, and the dependency relationship between clauses, predicate There is an advantage that a term determination rule for determining a term for a predicate or a behavioral noun can be automatically learned and output from information such as the state of.

According to the present invention, the following excellent effects can be obtained.
(1) A term determination rule for determining a term with high accuracy can be automatically learned.
(2) The term decision rule can be automatically learned including zero pronouns and compound nouns in which terms appear in a sentence different from the sentence containing the predicate. Therefore, by using the dictionary created in the present invention, predicate term structure analysis including zero pronouns and compound nouns can be handled uniformly.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments. FIG. 1 is a block diagram showing a configuration of a dictionary creation device 1 according to an embodiment of the present invention, and FIG. 2 is a flowchart showing an operation of the device of FIG.

  In FIG. 1, 2 is a training text (text to be analyzed) written in a natural language in which a correct term structure is manually added to a predicate or an action noun, and the text is subjected to syntax and semantic analysis. The result is a syntax that contains the basic form of the word contained in the text, part of speech, semantic category, whether it is a functional word, whether it is a symbol, the dependency between clauses, and the state of the predicate. It is a semantic analysis result table.

  Information regarding whether or not it is a function word and whether or not it is a symbol may be stored in the syntax / semantic analysis result table 2 in advance, or not stored in the syntax / semantic analysis result table 2. It is also possible to dynamically calculate and use the basic form of the word and the part of speech information as necessary.

  3 refers to the syntax / semantic analysis result table 2 and refers to the basic form, part of speech, semantic category, whether it is a functional word, whether it is a function word, whether it is a symbol, and the dependency between phrases. It is a training data creation unit as a training data creation means for extracting a training attribute index table 4 by extracting attributes for learning from the relationship and predicate state, and creating a training vector table 5 at the same time by creating a training vector. .

  6 is a weight learning unit that learns a weight representing the importance of an attribute using the training vector table 5 and creates a weight table 7 by adding the weight obtained by the learning to the training attribute index table 4. Is a weight learning unit.

  Reference numeral 8 denotes a term determination rule creating unit serving as a term determination rule creating unit that refers to the weight table 7 and outputs a list in which attributes are rearranged in the order of importance, as a term determination rule (dictionary).

  Next, the operation of the dictionary creating apparatus 1 configured as described above will be described with reference to FIG.

  First, the user inputs the training text with correct term structure and the syntax / semantic analysis result of the text into the syntax / semantic analysis result table 2 to the dictionary creating apparatus 1 (step S1).

  Next, the training data creation unit 3 extracts the attributes constituting the predicate and action noun term determination rules from the syntax / semantic analysis result table 2, creates the training attribute index table 4, and uses the attributes Then, a training vector is created and stored in the training vector table 5 (step S2).

  Next, the weight learning unit 6 learns the weight representing the importance of the attribute from the training vector table 5 using machine learning technology, and adds the weight obtained by the learning to the training attribute index table 4; It is stored in the weight table 7 (step S3).

  Next, the term determination rule creation unit 8 rearranges the information in the weight table 7 in the order of the attribute weights, creates and outputs a term determination rule, and ends (step S4).

  The syntax / semantic analysis result table 2, the training data creation unit 3, the training attribute index table 4, the training vector table 5, the weight learning unit 6, the weight table 7, and the term determination rule creation unit 8 of FIG. Regarding the specific configuration and operation, the syntax / semantic analysis result table 120, the training data creation unit 102, the training attribute index table 121, the training vector table 122, the weight learning unit 103, and the weight table 123 shown in FIG. Since they are the same as the term determination rule creation unit 104, the description thereof is omitted here.

  FIG. 3 is a block diagram showing the configuration of the dictionary creation device 10 according to another embodiment of the present invention, and FIG. 4 is a flowchart showing the operation of the device of FIG.

  In FIG. 3, reference numeral 101 denotes a training text (text to be analyzed) described in a natural language in which a correct term structure is given to a predicate or an action noun, and a basic form of a word included in the text. , Part-of-speech, semantic category, whether it is a functional word, whether it is a symbol, dependency relation between clauses, predicate state, syntax / semantic analysis to create a syntax / semantic analysis result table 120 It is a syntax / semantic analysis unit as a means.

  102, referring to the syntax / semantic analysis result table 120, the basic form of a word included in the text, the part of speech, the semantic category, whether it is a function word, whether it is a symbol, and the dependency between phrases It is a training data creation unit as a training data creation unit that extracts attributes for learning from the state of relations and predicates, creates a training attribute index table 121, and creates a training vector at the same time to create a training vector table 122. .

  103 is a weight learning unit that learns the weight representing the importance of the attribute using the training vector table 122 and adds the weight obtained by the learning to the training attribute index table 121 to create the weight table 123. Is a weight learning unit.

  Reference numeral 104 denotes a term determination rule creating unit serving as a term determination rule creating unit that refers to the weight table 123 and outputs, as a term determination rule (dictionary), a list in which attributes are rearranged in order of importance.

  The syntax / semantic analysis unit 101, the syntax / semantic analysis result table 120, the training data creation unit 102, the training attribute index table 121, the training vector table 122, the weight learning unit 103, the weight table 123, and the term determination rule creation unit 104, Each function to be described later is achieved by, for example, a computer.

  Next, the operation of the dictionary creating apparatus 10 configured as described above will be described with reference to FIG. First, a training text in which a correct term structure is manually added to a predicate or a behavioral noun written in a natural language is input to the syntax / semantic analysis unit 101 to the dictionary creation device 10 ( Step S11). For example, assume that a sentence as shown in FIG. 5 is input.

  Here, the part surrounded by the tags <NP ID = number> and </ NP> in FIG. 5 represents a term, and the part enclosed by <PRED ~> and </ PRED> represents a predicate. . Although it does not appear in this text, a behavioral noun is enclosed by tags <EVENT ~> and </ EVENT>. In addition, the description such as “NOM =“ 1 ”ACC =“ 2 ”” described in the “˜” part of the <PRED ~> tag and the <EVENT ~> tag is “in the basic form of this predicate or action noun. On the other hand, the correct answer of the term taking the main case represents "I" whose ID number is 1, and the correct answer of the term taking the main case represents "green pepper" whose ID number is 2. Here, for the sake of simplicity, it is assumed that only the main term and the opposite term are handled as the types of terms, but the same processing can be performed for the types of other terms.

  Next, the syntax / semantic analysis unit 101 performs syntax analysis and semantic analysis on the training text, thereby indicating whether or not the basic form, part of speech, semantic category, and function word of the word included in the text are symbols. And the dependency relationship between clauses and the state of the predicate are specified, and the analysis result is stored in the syntax / semantic analysis result table 120 (step S12). For example, if you parse the sentence “I hate bell peppers, but my mother forcibly eaten yesterday” and perform semantic analysis, as shown in FIG. , Sentence number, clause number, dependency clause number, word number in the clause, basic form of the word, part of speech, whether it is a function word or symbol, and the meaning category of the word. In addition, from the correct term structure given to the text, which word is the predicate or behavioral noun to be analyzed in the text, or the terms constituting the term structure for the predicate or behavioral noun are obtained. . In this case, suppose that “eating” is the predicate that is of interest now, and the terms for “eating” are “I” and “green pepper”, and in particular, “I” is the main term, and “green pepper” is the counter case. .

  Here, the sentence numbers in FIG. 6 are integers of 0 or more, and 0, 1, 2,. . . And given. The clause number is an integer greater than or equal to 0. 0, 1, 2,. . . And given. Further, the related clause number is the number of the clause that is determined to be related as a result of syntax analysis for each word. However, when there is no destination at the end of the sentence, the destination clause number is set to -1. The word number is an integer greater than or equal to 0, and is 0, 1, 2,. . . And given. The word basic form is the basic form of each word. The part of speech is the part of speech of each word. Whether the functional word / symbol is a functional word that represents the function of the content word attached to the content word, such as “ha” and “ga”, instead of the content word that represents the semantic content in the clause including the word. Or a symbol. The semantic category is a semantic category assigned to each word using a thesaurus such as a Japanese vocabulary system (see Reference Document: Non-Patent Document 4).

  The syntax / semantic analysis unit 101 can use a single syntax analyzer such as CaboCha (reference: see Non-Patent Document 5).

  Next, the training data creation unit 102 assumes that, in the syntax / semantic analysis result table 120, a word that appears in the nearest place from the predicate under some restriction is a term for the predicate, starting from the predicate. Then, the attribute that becomes the restriction is extracted, and the training attribute index table 121 is created (step S13). For example, in the example of FIG. 6, when “I”, which is the main term for the predicate “eat”, “searches for a word before the predicate”, the meaning category is “person” and Among words whose function word is “ha”, it is the word in the nearest place. For each word appearing in the input text other than the predicate, it is examined what kind of constraint is closest to the predicate, and the constraint is used as an attribute. For example, the constraint is a combination of (target word search method, part of speech, semantic category, predicate state in dependency relationship).

  Here, the search method of the target word is, for example, searching for a word ahead of the predicate, and in a state where the clause including the word is related to the clause including the target predicate, the functional word / symbol of the clause including the word and Ic when considering the state of the predicate that the word is related to, search for the word after the predicate, and the functional word of the clause that includes the word from the clause containing the target predicate to the clause containing the word / Oc to consider the state of the predicate with which the symbol and the word are involved, sc if the predicate and the target word are in the same clause, and a clause between the clause containing the predicate and the clause containing the target word If there is no receiving relationship and the same clause, the function word / symbol of the clause containing the word and the predicate state in the dependency relationship are taken into consideration first, but such a word is still If not, how to search behind the predicate And c. Further, let fw be a method for searching forward from a predicate without considering the predicate state in which the function word / symbol and the word are related, and bw be a method for searching backward similarly. For example, in the example of FIG. 6, the word “mother” is closest to the predicate “eat” due to restrictions such as (search method = ic, word semantic category = person, function word =, state = passive). is there. In this way, the training attribute index table 121 as shown in FIG. 7 is created. Many attributes are created, of which only 10 are shown here.

  Next, the training data creation unit 102 creates training vectors for each word other than predicates in the syntax / semantic analysis result table 120 based on the training attribute index table 121 and the syntax / semantic analysis result table 120, It stores in the vector table 122 (step S14). For example, in FIG. 5, it is assumed that training vectors are created in order from the top for words excluding predicates, and these vectors are represented by x_1, x_2, x_3,. . . And In this case, a training vector as shown in FIG. 8 is generated. It is checked whether or not each attribute of the training attribute index table 121 as shown in FIG. 7 meets the condition of each word. If so, the attribute value is 1, and if not, the attribute value is 0. For example, the fourth element in x_1 is 1, which means that the word “I” is the condition of attribute number 4 in FIG. 7, that is, the search type is nc, that is, the word that is not dependent on the predicate. When searching forward from a predicate in a word, the semantic category is “people”, the phrase containing the word is “ha”, and the predicate in a dependency relationship with the word is active This is because the position is closest to the predicate. On the other hand, the ninth element of x_1 is 0, which means that when searching for the word in front of the predicate without any function or condition conditions from the predicate “eat”, If the category is “People”, sentence number 1, phrase number 2, and word number 0 “mother” are the closest, and “I” is the second closest, so the condition is not met It is determined to be 0. Thus, a training vector is generated and stored in the training vector table 122.

  Further, for each word excluding the predicate, a teacher variable is stored in the training vector table 122 for each type of term (step S15). For example, y_ {1, NOM} = 1, y_ {2, NOM} = 0, y_ {3, NOM} = 0,. . . , Y_ {1, ACC} = 0, y_ {2, ACC} = 0, y_ {3, ACC} = 1,. . . , Is stored. Here, y_ {i, NOM} is a scalar variable that stores 1 if the i-th word is a correct principal term, and -1 otherwise. Similarly, y_ {i, ACC} is a scalar variable storing 1 if the i-th word is a correct case term, and -1 otherwise. Note that NOM represents English nominal corresponding to the main case, and ACC represents English accumulative equivalent to the main case. Thus, the training vector and the teacher variable as shown in FIG.

  Next, the weight learning unit 103 reads the training vector and the teacher variable from the training vector table 122 for each term type, and calculates the weight of the attribute importance using the machine learning method (step S16).

  As an example, a machine learning method called SVM (Support Vector Machine), in particular, a method of learning using a machine learning method called linear SVM in Non-Patent Document 6 will be described below.

For linear SVMs, (x ₁ , y ₁ ),. . . , (X _m , y _m ) ... For the m pieces of training data represented by (1), the positive example side (part where y> 0) and the negative example side (part where y <0) are divided. Find two parallel hyperplanes. At this time, a hyperplane in which the distance between the two hyperplanes is maximized is obtained. The hyperplane obtained here is expressed by equation (2).

  Here, w is a weight vector, b is a bias, and both w and b can be obtained by solving the optimization problem expressed by Equation (3) from the training data.

  In the case of the linear SVM, the element of the weight vector w obtained in this way means that the greater the value, the higher the importance of the corresponding attribute. For example, it is assumed that learning is performed on the contents of the training vector table 122 as shown in FIG. Here, in the linear SVM for the main term, (x_1, y_ {1, NOM}), (x_2, y_ {2, NOM}), (x_3, y_ {3, NOM}),. . . , (X_1, y_ {1, ACC}), (x_2, y_ {2, ACC}), (x_3, y_ {3, ACC}),. . . The vector and scalar variable pairs are used as training data.

As a result, it is assumed that the weight vector w _NOM related to the main case term and the weight vector w _ACC related to the case term as shown in FIG. 10 are obtained. Then, the weight learning unit 103 adds the weight vector obtained as shown in FIG. 10 to the table of the training attribute index table 121 (table of FIG. 7), thereby obtaining a weight table (FIG. 11) related to the main item. A weight table (FIG. 12) regarding the relative terms is obtained and stored in the weight table 123.

  Next, the term determination rule creation unit 104 collects a plurality of weight tables in the weight table 123, sorts them by weights, and outputs them as a decision list (step S17). For example, the tables of FIGS. 11 and 12 are combined to obtain the table of FIG. FIG. 14 shows the top 10 items sorted by weight. This table is output as a decision list format to obtain a term determination rule table (dictionary).

  In the above embodiment, a specific example is described in which the output term structure is a term structure related to the mandatory surface case for the predicate basic form. However, even if the output term structure is related to any surface case or deep case, the syntax -It can be realized by the same means if the semantic analysis result table is made corresponding to them.

  Also, in the weight learning method, if a weight table is obtained, a learning method different from the machine learning method described above can be used.

  In addition, even if the target is a foreign language text such as English, instead of the dependency function word, the part structure other than the verb / noun and the phrase structure such as the noun phrase / verb phrase composed of words are used. This learning method can be used by using the relationship as a dependency relationship.

  Further, a part or all of the functions of each means in the dictionary creation device of the present embodiment can be configured by a computer program, and the program can be executed using the computer to realize the present invention. It goes without saying that the procedure in the dictionary creation method can be constituted by a computer program, and the program can be executed by the computer, and the program for realizing the function by the computer can be read by a computer-readable recording medium such as an FD. (Floppy (registered trademark) Disk), MO (Magneto-Optical disk), ROM (Read Only Memory), memory card, CD (Compact Disk) -ROM, DVD (Digital Versatile D) sk) -ROM, CD-R, CD-RW, HDD, and recorded in a removable disk, or stored, it is possible or distribute. It is also possible to provide the above program through a network such as the Internet or electronic mail.

The block diagram which shows the system configuration | structure of the dictionary preparation apparatus regarding one Example of this invention. The flowchart for demonstrating operation | movement of the dictionary creation apparatus regarding one Example of this invention. The block diagram which shows the system configuration | structure of the dictionary creation apparatus regarding the other embodiment of this invention. The flowchart for demonstrating operation | movement of the dictionary creation apparatus regarding the other embodiment of this invention. Explanatory drawing which shows the example of training data with a correct answer term structure input into the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of a syntax and a semantic analysis result table in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the training attribute index table in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the training vector produced within the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the training vector table in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the weight vector obtained by machine learning in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the weight table regarding the main item in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the weight table regarding the relative term in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the table which put together the weight table regarding a main item, and the weight table regarding a relative item in the dictionary creation apparatus of this invention. Explanatory drawing which shows the example of the term determination rule output from the dictionary creation apparatus of this invention.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1,10 ... Dictionary creation apparatus, 2,120 ... Syntax and semantic analysis result table, 3,102 ... Training data creation part, 4,121 ... Training attribute index table, 5,122 ... Training vector table, 6,103 ... Weight Learning unit, 7, 123 ... weight table, 8, 104 ... term determination rule creation unit, 101 ... syntax / semantic analysis unit.

Claims (4)

For each word included in the text to be analyzed described in natural language, in order from the beginning of the text to be analyzed, the notation of the word, the basic form of the word, the clause to which the clause to which the word belongs , The predicate state that is in a dependency relationship with the word, information indicating whether the word is a predicate or a behavioral noun, and the word is in a term relationship with the predescription word or the behavioral noun In this case, the correct answer term structure and a syntax / semantic analysis result table storing
For each predicate or behavioral noun in the syntax / semantic analysis result table, a word other than the predicate or behavioral noun in the syntax / semantic analysis result table that satisfies a constraint condition starting from the predicate or behavioral noun When the search is performed, the constraint condition is extracted as an attribute so that the word becomes a word appearing closest to the predicate or the action noun, and a training attribute index table is created, and a word other than the predicate or the action noun is created. A training vector composed of values indicating whether or not the word meets each of the constraints, a correct term structure for each of the words other than the predescription word or the behavioral noun, and other terms Training data creating means for creating a training vector table comprising a teacher variable for identifying a structure;
With respect to the training vectors and teacher variables described in the training vector table, for each type of term structure, data consisting of a combination of a teacher variable and a training vector indicating a correct term structure is used as a positive example. Using a data consisting of a set of a teacher variable indicating a term structure and a training vector as a negative example, a hyperplane that maximizes the distance between two parallel hyperplanes that divide the positive example and the negative example is obtained. Weight learning means for learning a weight representing the importance of the attribute using a machine learning method based on a hyperplane, and adding a weight obtained by the learning to the training attribute index table to create a weight table;
A term determination rule creating means for referring to the weight table and outputting a list in which attributes are sorted in the order of high importance attributes as a term determination rule for determining a term structure for a basic form of a predicate or an action noun,
The information indicating the correct term structure is information indicating the correct term structure for the basic form of the predicate or action noun included in the text to be analyzed,
The constraints are
(a) the search range or search direction of the word, based on the phrase including the predicate or behavioral noun as the starting point;
(b) the presence or absence of a dependency relationship between a clause including the predicate or behavioral noun as a starting point and a clause including a word other than the previous description word or behavior noun;
(c) whether or not to consider predicate state information that is in a dependency relationship with words other than predescription words or action nouns;
Including
A dictionary creation device, wherein the output term determination rule is a dictionary. For each word included in the text to be analyzed described in natural language, in order from the beginning of the text to be analyzed, the notation of the word, the basic form of the word, the clause to which the clause to which the word belongs , The predicate state that is in a dependency relationship with the word, information indicating whether the word is a predicate or a behavioral noun, and the word is in a term relationship with the predescription word or the behavioral noun In the case, a dictionary creation method in a device having a syntax / semantic analysis result table storing the correct answer term structure,
For each predicate or behavioral noun in the syntax / semantic analysis result table, the training data creation means uses the predicate or behavior in the syntax / semantic analysis result table that satisfies the constraint condition starting from the predicate or behavioral noun. When searching for a word other than a sex noun, the constraint condition such that the word becomes a word appearing in the nearest place from the predicate or action noun is extracted as an attribute to create a training attribute index table, and the predicate or For each word other than a behavioral noun, a training vector composed of a value indicating whether or not the word is applicable to each of the constraint conditions, and a correct answer term for each of the words other than the previous description word or the behavioral noun Creating a training vector table comprising teacher variables for identifying structures and other term structures;
The weight learning means, for the training vector and the teacher variable described in the training vector table, is a positive example of data consisting of a set of a teacher variable and a training vector indicating a correct term structure for each type of term structure. And a hyperplane that maximizes the distance between two parallel hyperplanes that divide the positive and negative examples, using the data consisting of a set of teacher variables and training vectors indicating the term structure other than the correct answer as a negative example. Based on the obtained hyperplane, a weight representing the importance of the attribute is learned using a machine learning method, and the weight obtained by the learning is added to the training attribute index table to create a weight table A weight learning step;
A step for generating a term determination rule with reference to the weight table, and outputting a list in which attributes are rearranged in the order of high importance as a term determination rule for determining a term structure for a basic form of a predicate or an action noun; Run,
The information indicating the correct term structure is information indicating the correct term structure for the basic form of the predicate or action noun included in the text to be analyzed,
The constraints are
(a) the search range or search direction of the word, based on the phrase including the predicate or behavioral noun as the starting point;
(b) the presence or absence of a dependency relationship between a clause including the predicate or behavioral noun as a starting point and a clause including a word other than the previous description word or behavior noun;
(c) whether or not to consider predicate state information that is in a dependency relationship with words other than predescription words or action nouns;
Including
A dictionary creation method, wherein the output term determination rule is a dictionary. A dictionary creation program for causing a computer to function as each means according to claim 1 . A computer-readable recording medium on which the dictionary creating program according to claim 3 is recorded.

Images