JP5356197B2 - Word semantic relation extraction device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To highly precisely extract a word semantic relation from text data by using an existing dictionary such as a synonym dictionary. <P>SOLUTION: A device for extraction of word semantic relation is configured to calculate various types of similarities as for arbitrary pairs of words in a text, and to generate identity vectors with each similarity as an element, and to give each pair of words a label indicating whether or not they are synonyms based on a synonym dictionary, and to learn a classifier from the identity vector and the label, and to identify whether or not the two words are synonyms by the learnt classifier. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a technique for extracting a semantic relationship between words from a text, and more particularly to a technique for extracting a word semantic relationship such as a synonym, a higher / lower term, a sibling, a parallel translation, and the like.

  With the spread of personal computers and the Internet, the amount of electronic documents accessible to users is increasing. One of techniques for efficiently finding a desired document from such large-scale document information is a document search technique. According to the document search technique, a user can efficiently obtain a desired document by finding a document including an input keyword. However, in many cases, a simple character string search is not sufficient. One problem that remains unsolved is the problem of synonyms. That is, since there are a plurality of words expressing the same meaning, a document expressing the same meaning cannot be found by a simple character string search, and a search omission may occur. In order to deal with such a problem of synonyms, it has been conventionally performed to provide a synonym dictionary in a search system.

  Since manual creation of a synonym dictionary requires a large cost, attempts have been made to automatically create a synonym dictionary from text data. One method for creating a synonym dictionary is to focus on the appearance context of words, that is, on words and character strings that appear in the vicinity of the word of interest. Non-Patent Document 1 discloses a context-based synonym extraction technique based on appearance context. There are also methods for dealing with notation fluctuations among synonyms. Non-Patent Document 2 discloses a notation-based synonym extraction technique for detecting katakana notation fluctuation based on pronunciation rules. With the recent spread of Web and Web document search engines, word semantic relationship extraction techniques using search engines have been proposed. In the approach using a search engine, the appearance context of words cannot be calculated in advance. Therefore, a method has been proposed in which a co-occurrence frequency is acquired by inputting a query with AND in a search expression, and a synonym is extracted by a statistic based on the co-occurrence frequency. Non-Patent Document 3 discloses a co-occurrence-based synonym extraction technique based on a search engine. There is also a synonym extraction technique that uses a synonym such as “C such as A or B” or a synonym pattern that explicitly indicates a broader or lowerer term. Non-Patent Document 4 discloses a pattern-based synonym extraction technique using a word pattern. Moreover, there is a bilingual relationship as one of the semantic relationships between words. The bilingual relationship can be regarded as an extension of the synonym relationship to multiple languages. Non-Patent Document 5 discloses a technique for automatically extracting a bilingual relationship. This technology is an extension of the context-based synonym extraction technology to multiple languages.

  The above synonym extraction technology is based on unsupervised learning, that is, a type of learning technology that does not use a correct answer given manually. Since unsupervised learning does not require the creation of correct answers, it is an advantage that the cost of manpower is low. However, the following problems exist.

  Currently, large dictionaries created manually are widely available. Existing synonym dictionaries, thesaurus dictionaries, and bilingual dictionaries are valuable resources that have been developed at high costs and need to be used as effectively as possible. The word semantic relationship extraction technique based on unsupervised learning does not assume the existence of such a manually created dictionary, and even if a manually created dictionary exists, it cannot be used to improve accuracy.

  As a method for solving the above problems, Non-Patent Document 6 discloses a synonym extraction method by supervised learning. In Non-Patent Document 6, synonym extraction is performed by supervised learning using a synonym dictionary created manually as a correct answer. Specifically, the meaning of a word is expressed based on the context of the word, which will be described later, and learning is performed by using a synonym dictionary that is a correct answer, and synonyms are extracted.

Aizawa: “Study on word similarity calculation using large-scale text corpus”, IPSJ Journal, vol. 49-3, pp. 1426-1436 (2008). Kubota et al: Unification method of katakana notation Preliminary classification and fluctuation detection method of katakana notation by graph comparison, IPSJ Natural Language Processing Study Group Report, NL97-16, pp.111-117,1993. P. Turney. 2001. Mining the web for synonyms: PMI-IR versus LSA on TOEFL. ECML 2001, 491-502. M. Hearst. Automatic acquisition of hyponyms from large text corpora.In Proceedings of the 14th International Conference on Computational Linguistics (COLING-92), pp. 539-545, 1992. Hiroyuki Kaji and Toshiko Aizono, “Extracting word correspondences from bilingual corpora based on word co-occurrence information,” Proceedings of the 16th International Conference on Computational Linguistics, pp.23-28, 1996. Masato Hagiwara: A Supervised Learning Approach to Automatic Synonym Identification based on Distributional Features, Proc. Of ACL 2008 Student Research Workshop, pp. 1-6, 2008.

  An object of the present invention is to realize a word semantic relationship extraction technique with higher accuracy than the prior art. In the supervised learning approach, while the above problems are solved, there are problems unique to supervised learning. The biggest problem is that the knowledge accumulated in previous research on unsupervised learning is not utilized. For example, in Non-Patent Document 6, all arbitrary words co-occurring with a word pair are used as features, and the similarity itself regarding the entire context distribution is to be learned from the teacher data. However, various proposals and improvements as disclosed in Non-Patent Document 1 have been made regarding the similarity of the distribution of contexts. It is necessary to apply supervised learning while incorporating such knowledge.

  Non-Patent Document 6 discloses a synonym extraction technique based on context-based similarity using the result of syntactic analysis, but various approaches that have been studied a lot in synonym extraction techniques based on unsupervised learning. Has not been studied. Past approaches to unsupervised learning have their own strengths and weaknesses. For example, the notation-based method disclosed in Non-Patent Document 3 can extract only specific types of synonyms such as Katakana variant notation. The pattern method disclosed in Non-Patent Document 4 can extract any type of synonym with relatively high accuracy, but has low coverage and it is difficult to extract all necessary synonyms. The context-based similarity is almighty with respect to the types of synonyms that can be extracted, and can cover a wide range of synonyms, but has a lower precision than the notation-based and pattern-based methods. Integration of these methods is essential for improving accuracy.

  The present invention has been made in order to solve the above-described problems. The word semantic relationship is capable of integrating a plurality of approaches and setting an appropriate threshold while utilizing an existing synonym dictionary and thesaurus dictionary. The purpose is to provide an extraction method.

  The word semantic relationship extraction device of the present invention refers to a means for generating a feature vector having a plurality of types of similarities as elements for a set of words extracted from a text, and a known dictionary. A means for providing a label indicating a word semantic relationship, a means for learning a word semantic relationship determination rule based on a plurality of feature vectors to which the label is assigned, and an arbitrary word based on the learned word semantic relationship determination rule Means for determining a word semantic relationship with respect to the set.

  An example of the word semantic relationship is a relationship as to whether or not two words in a set of words are synonyms. At this time, as a known dictionary, a synonym dictionary storing headwords and their synonyms is used.

  Other examples of word semantic relationships are whether two words in a set of words are synonyms, upper / lower relationships, sibling relationships, or neither, known at this time The dictionary uses a thesaurus that stores headwords and their synonyms, upper and lower terms, or siblings.

  Another example of the word semantic relationship is a bilingual relationship between two words in a set of words. At this time, a bilingual dictionary storing headwords and their translations is used as a known dictionary.

  The word meaning relationship extraction apparatus of the present invention can be realized by a computer system including a processor, a memory, and an interface.

  The similarity of a set of words that are elements of a feature vector can be obtained by various methods. As an example, a method of calculating a context-based similarity using a context matrix obtained by extracting a set of a word (processing target word) and a context word (context word) from a text and aggregating the extracted results It is. Another example is a method of calculating the character duplication degree based on the character duplication degree of an arbitrary word set in the text and calculating the similarity of the word set based on the character duplication degree. Or you may calculate the similarity of a word set based on the similarity of the character of the arbitrary word set in a text. Yet another example is a method of extracting a co-occurrence frequency indicating the frequency of simultaneous appearance for an arbitrary set of words in a text and calculating a co-occurrence similarity based on the extracted result.

  According to a typical embodiment of the present invention, additional information sources such as manually created synonym dictionaries, thesaurus dictionaries, bilingual dictionaries are used as teacher data, and at the same time, different types of similarities obtained by multiple approaches are integrated. By doing so, it is possible to perform word semantic relationship extraction with higher accuracy than in the past.

It is a block diagram which shows the structural example of the computer system by this invention. It is the figure which showed the relationship between a word meaning relationship extraction program, a dictionary, various tables, and a file. It is the sequence diagram which showed the flow of the process in the computer system of this invention. It is explanatory drawing of a similarity matrix. It is a flowchart of a word meaning relationship extraction process. It is explanatory drawing of a synonym dictionary. It is explanatory drawing of a thesaurus dictionary. It is a conceptual explanatory drawing of synonym identification. It is explanatory drawing of the screen shown to a user. It is explanatory drawing of a context matrix. It is explanatory drawing of a context matrix. It is a flowchart of a context extraction process. It is explanatory drawing of a morphological analysis result. It is explanatory drawing of a context pattern. It is a flowchart of a character duplication degree calculation process. It is a flowchart of a character similarity calculation process. It is explanatory drawing of a character similarity table. It is explanatory drawing of a co-occurrence frequency table. It is explanatory drawing of a word frequency table. It is explanatory drawing of a co-occurrence similarity table. It is explanatory drawing of the experimental result which shows the effect of the word meaning relationship extraction apparatus of this invention. It is explanatory drawing of a similarity matrix. It is explanatory drawing of the screen shown to a user. It is explanatory drawing of a bilingual dictionary. It is explanatory drawing of a similarity matrix. It is explanatory drawing of a context matrix. It is explanatory drawing of a context matrix.

Embodiments of the present invention will be described below with reference to the drawings.
[First Embodiment]
As a first embodiment, a synonym extraction device for extracting a word pair having a synonym relationship as a word semantic relationship will be described. FIG. 1 is a block diagram illustrating a configuration example of a computer system that implements the present invention. The computer system shown in FIG. 1 is used in the first embodiment of the present invention and is commonly used in the second and third embodiments of the present invention. Note that functions that are not used in some embodiments are also included.

  The word meaning relationship extraction device 100 includes a CPU 101, a main memory 102, an input / output device 103, and a disk device 110. The CPU 101 performs various processes by executing programs stored in the main memory 102. Specifically, the CPU 101 calls a program stored in the disk device 110 on the main memory 102 and executes it. The main memory 102 stores programs executed by the CPU 101, information required by the CPU 101, and the like. Information is input to the input / output device 103 from the user. The input / output device 103 outputs information in response to an instruction from the CPU 101. For example, the input / output device 103 includes at least one of a keyboard, a mouse, and a display.

  The disk device 110 stores various information. Specifically, the disk device 110 includes an OS 111, a word semantic relationship extraction program 112, a text 113, a manual creation dictionary 114, a similarity matrix 115, a context matrix 116, a part of speech pattern 117, a co-occurrence similarity table 118, and an identification model 119. The character similarity table 120 is stored.

  The OS 111 controls the entire processing of the word meaning relationship extraction apparatus 100. The manually created dictionary 114 is a variety of manually created dictionaries, and includes a synonym dictionary 1141, a thesaurus dictionary 1142, and a bilingual dictionary 1143. The synonym dictionary 114 is a dictionary in which synonyms created manually are stored. The thesaurus dictionary 115 is a dictionary in which synonyms and broader / lowerer words created manually are stored.

  The word meaning relationship extraction program 112 is a program that extracts word meaning relationships from the text 113 and the synonym dictionary 1141 or thesaurus dictionary 1142. The feature vector extraction subprogram 1121, the correct label setting subprogram 1122, and the identification model learning subprogram 1123. And an identification model application subprogram 1124.

  The text 113 is text to be input to the word meaning relationship extraction program 112 and does not need to be in a special format. In the case of a document including a tag, such as an HTML document or an XML document, it is desirable to perform preprocessing to remove the tag, but the processing is possible even in a state where the tag is included.

  The similarity matrix 115 is a matrix that stores a feature vector related to a word pair extracted from a text and a synonym dictionary, a label indicating whether or not it is a synonym, and the like. The context matrix 116 is a matrix that stores context information of words necessary for calculating context-based similarity. The part-of-speech pattern 117 is data used to extract context information of words necessary for calculating the context-based similarity from the text. The co-occurrence similarity table 118 is a table that stores co-occurrence-based similarity calculated based on word co-occurrence. The identification model 119 is a model for identifying whether a word pair is a synonym learned from a similarity matrix. The character similarity table 120 is a table that stores relationships between characters having similar meanings.

  FIG. 2A is a diagram illustrating relationships among the word meaning relationship extraction program, the dictionary, various tables, and files illustrated in FIG. The feature vector extraction subprogram 1121 reads the text 113, extracts all the words in the text, calculates various similarities for an arbitrary set of words, and outputs it as a similarity matrix 115. Information such as a context matrix 116 and a co-occurrence similarity table 118, which are necessary information, are created in advance. In the first embodiment, it is assumed that the text is composed of documents in the same language, for example, Japanese documents. However, even if some English documents are included, there is no problem other than unnecessary processing. The part-of-speech pattern 117 is used to create the context matrix 116. The correct answer label setting subprogram 1122 reads the synonym dictionary 1141, thesaurus dictionary 1142, and the bilingual dictionary 1143 as correct answer data, and sets a label indicating whether each word pair in the similarity matrix 115 is a correct answer, that is, a synonym. To do. The identification model learning subprogram 1123 reads the similarity matrix 115 and learns an identification model 119 for identifying whether a word pair is a synonym. The identification model application subprogram 1124 reads the identification model 119 and gives a determination result as to whether or not the word pair in the similarity matrix 115 is a synonym.

  FIG. 2B is a sequence diagram showing the flow of processing in the computer system of the present invention. First, the OS is loaded from the disk device to the main memory and waits for user input. The process is started by an instruction to execute the word meaning relationship extraction program by the user. First, a feature vector extraction subprogram is loaded into the main memory, and the feature vector extraction subprogram reads text, extracts all words in the text, and creates a context matrix using the part-of-speech pattern. Next, a character similarity table is created from words obtained from morphological analysis results and a manual creation dictionary. Next, a co-occurrence similarity table is created from the morphological analysis results. Then, a similarity matrix composed of various similarities is created. In the first embodiment, it is assumed that the text is a document in the same language, for example, a Japanese document. However, even if some English documents are included, there is no problem other than unnecessary processing.

  The correct answer label setting subprogram reads the manually created dictionary as correct answer data, and sets a label indicating whether each word pair in the similarity matrix is a correct answer, that is, a synonym. The identification model learning subprogram reads the similarity matrix and learns an identification model for identifying whether or not it is a word pair. The identification model application subprogram reads the identification model and gives a determination result as to whether or not the word pair in the similarity matrix is a synonym.

Hereinafter, the basic concept of the present invention will be described using the example of the similarity matrix shown in FIG.
Consider an arbitrary word pair included in text data. For example, the word pair is <computer, computer>. At this time, various scales for determining whether a word pair is a synonym can be assumed.

  For example, as disclosed in Non-Patent Document 1, there is a method using the similarity between appearance contexts of words (hereinafter referred to as context-based similarity). Further, similarity based on notation such as focusing on the number of overlapping characters as disclosed in Non-Patent Document 2 (hereinafter referred to as notation-based similarity) can be considered. Furthermore, as disclosed in Non-Patent Document 3, it is possible to use a similarity based on the frequency with which word pairs co-occur (hereinafter referred to as a co-occurrence-based similarity). Furthermore, there are various variations in each method. For example, in context-based similarity, there are variations depending on how the word appearance context is defined or how the distance calculation method is defined. Also, in the co-occurrence-based similarity, different statistics such as mutual information and Dice coefficient can be used as the similarity calculated from the co-occurrence frequency. In the present invention, such various similarities are considered to be the feature of the word pair, and the word pair is represented by a feature vector composed of values for each feature. In the example of FIG. 3, for example, the word pair <computer, computer> has a feature 1 dimension value of 0.3, a feature 2 dimension value of 0.2, and a feature N dimension value of 0.8. It is expressed as a vector.

  Further, it is determined whether or not the word pair is a synonym using a manually created dictionary such as a synonym dictionary or a thesaurus dictionary, and labeling is performed. That is, if <computer, computer> is included in the synonym dictionary, the label <computer, computer> is given as a correct answer. A line representing a correct answer, that is, a word pair is called a positive example. In the example of FIG. 3, since <computer, computer> and <computer, computer> are synonyms, “1” representing the correct answer is given as a label. If the word pair is not included in the synonym dictionary, a label of incorrect answer is given. A line representing an incorrect answer is called a negative example. In the example of FIG. 3, since <program, computer> is not a synonym, “−1” representing an incorrect answer is given as a label. In this way, by expressing a word pair with a vector of feature values and adding correct data, a classifier based on supervised learning such as a support vector machine can be applied. The above is the basic idea of the present invention.

  Here, when giving a label, caution is required if the word pair is not included in the manually created dictionary. Since manual dictionaries are not perfect, there are cases where they are synonyms even if they are not included in the synonym dictionary. A method for dealing with this problem will be described later.

  FIG. 4 is a flowchart of word semantic relationship extraction processing executed by the synonym extraction device according to the first embodiment of this invention.

  In step 11, it is determined whether or not all word pairs have been processed. If completed, go to Step 17. If there is an unprocessed word pair, the process proceeds to step 12. In step 12, it is determined whether or not the processing has been completed for all types of features. If completed, go to step 16. If there is an unprocessed feature, the process proceeds to step 13.

  In step 13, the i-th word pair is acquired. For example, word pairs can be acquired by, for example, preparing a whole word list by morphological analysis of text and acquiring a combination of two arbitrary words from the list. In step 14, the j-th feature is calculated for the acquired i-th word pair. Details of the processing in step 14 will be described later. Next, the process proceeds to step 15, and the feature calculation result is stored in the similarity matrix. An example of the similarity matrix is as described in FIG.

  In step 16, a label is set in the similarity matrix. The label is set by referring to the synonym dictionary or thesaurus dictionary. In the first embodiment, since a document in the same language is assumed, a bilingual dictionary is not normally used. However, in the case of a technical document, there is a case where an English word is included in a Japanese document. . In order to cope with such a case, a bilingual dictionary may be used.

  An example of the synonym dictionary is shown in FIG. 5, and an example of the thesaurus dictionary is shown in FIG. The synonym dictionary is data in which one is stored in the headword column and the other is stored in the synonym column for word pairs that are synonyms. For the sake of dictionary lookup, it is assumed that data is retained redundantly. That is, for the synonymous pair <computer, computer>, it is assumed that both a line having “computer” as a headword and a line having “computer” as a headword are held. Thereby, all synonym pairs can be acquired by confirming only the headword column.

  The thesaurus dictionary stores one word in the headword column and the other in the related word column for word pairs that are synonyms and upper / lower terms, and the type column indicates the type of related word for the headword. Stored data. For example, in the case of the example of FIG. 6, “computer” is found, “device” is a related word, and “device” is “computer” for a word pair having a higher / lower term relationship such as <computer, device>. "Is a broader term" (a more abstract word). It is assumed that the thesaurus dictionary also holds data redundantly for the sake of dictionary lookup. That is, for the word pair <computer, device>, it is assumed that both a row having “computer” as a headword and a row having “device” as a headword are held. Here, it is necessary to note that the type of the pair whose order is reversed is similarly reversed particularly when the word pair is in the upper / lower word relationship. For example, “computer” is a subordinate term of “device”.

  When setting a label in the similarity matrix, if the word pair matches a certain line in the synonym dictionary, that is, is a synonym, “1” is assigned as the correct label. Otherwise, process as follows. If the word pair is not a synonym, that is, there is no line containing the word pair in the synonym dictionary, but each word is included in another line of the synonym dictionary, the label “−” 1 ”is given. If at least one word of the word set is not included in the synonym dictionary, “0” is assigned as an unknown label.

  In the example of FIG. 3, since <computer, computer> and <computer, computer> are synonyms, “1” is given as a label. Also, <program, computer> is not a synonym, that is, “program” and “computer” are included in the synonym dictionary, but there is no line that includes both. 1 "is given. For <computer, virtualization technology>, “0” is assigned as a label on the assumption that “virtualization technology” is not included in the synonym dictionary. When referring to the thesaurus dictionary, the type column is referred to, and the same processing is performed only for the line whose type is a synonym.

  Returning to FIG. 4, in step 17, an identification model is learned. From the similarity matrix, a binary identification model is learned only for the rows whose labels are “correct” or “incorrect”. As an identification model, any model can be used. For example, CJCBurges, “A Tutorial on Support Vector Machines for Pattern Recognition” Data Mining and Knowledge Discovery, vol.2, pp.121-168 (1998) The support vector machine disclosed in can be used.

  FIG. 7 shows a conceptual diagram of synonym identification. The feature vector of each word pair corresponds to a certain point on the N-dimensional space represented by the features 1 to N, and is represented by a black square in FIG. At this time, the learning of the identification model is to find a boundary between a region where a word pair which is a synonym is arranged and a region where a word pair which is not a synonym is arranged. When a word pair whose unknown point, that is, whether it is a synonym, is given, it is an application of the identification model to determine whether it is a synonym depending on which region it belongs to. The support vector machine is characterized in that a non-linear identification model, that is, a boundary other than a straight line, a plane, or a hyperplane (a plane in a space of four or more dimensions) can be used.

  In step 18, word semantic relationship extraction is performed from the value of the similarity matrix according to the model. For every word pair in the matrix, the feature vector is input to the learned classifier to identify whether it is a synonym. The determination result of the discriminator is stored in the determination result column of the similarity matrix. Thereby, it is determined whether or not the word pair whose label is “unknown”, that is, “0”, is a synonym. It can also be used for error checking of a synonym dictionary manually. For a word pair to which a label other than “unknown” has already been assigned, only those having a different label and determination result are extracted and checked manually to check the synonym dictionary efficiently.

  FIG. 8 shows a screen example of the synonym dictionary editor. Although the label is a synonym but the determination result is a word pair that is not a synonym is displayed at the top of the screen, the label is changed depending on the result of the manual check. Similarly, although a label is not a synonym, a word pair that is a synonym is displayed in the lower part of the screen in the determination result, and the label is changed according to a human check result. With such an editor, the synonym dictionary can be checked. Of course, on the assumption that the data in the synonym dictionary is correct, it is also possible to target only “unknown” word pairs.

  Hereinafter, the process of step 14 in FIG. 4 will be described in detail. In step 14, various similarities are calculated as features for expressing word pairs. Hereinafter, description will be made for each type of similarity.

(1) Context-Based Similarity Hereinafter, a method for calculating the context-based similarity will be described. The context of a word indicates a word “near” or a word string at a location where the word appears in the text. Various contexts can be defined depending on what is defined as “neighbor”. In the following, an example is described in which the following verb and the immediately preceding adjective / adjective verb are used as the context as the context, but other occurrence contexts may be used instead, or added / combined. Is possible. There are also various methods for calculating the similarity between contexts.

  The context-based similarity is calculated based on the context matrix. FIG. 9 shows an example of the context matrix. The context matrix includes a heading field and a context information field, and stores context information including a repetition of a combination of a context word string and its frequency for words in the heading field. The example of FIG. 9 shows the case where the particle + predicate following the focused word is used as the context. For example, in “Computer”, “Start up” appears 15 times and “Connect” appears four times. For such a context matrix, context information of a row corresponding to any two words is acquired, and the similarity is calculated based on the frequency vector of the context word string. As the context-based similarity, a method used for document search by a term vector model can be used, and is disclosed in, for example, Kita, Tsuda, and Tsurugi-min "Information Search Algorithm" Kyoritsu Publishing (2002). The method can be used. In this embodiment, as an example, the similarity s is calculated by the similarity calculation method of the following equation.

  In addition, the description of the parameters in the formula is an explanation when applied to document search. In the case of synonym extraction, the input document is the target input word for synonym extraction, the target document is the synonym candidate word, and the input document. The word inside is replaced with the context word of the input word.

  There are various variations on what words are extracted as contexts. For example, “fast” can be extracted from an expression such as “fast computer” as the context of “computer”, and “calculate (calculate) average value” as the context of “calculate”. It is also possible to extract “average value” from such an expression. Such various variations of contexts may be handled together, or each context may be treated as a distinct feature. In the present embodiment, an example will be described in which two different types of contexts are treated as distinctive features. FIG. 10 shows an example of the result of extracting an adjective and an adjective verb that appear before the word of interest as a different type of context from FIG.

  Hereinafter, a method for creating a context matrix executed by the feature vector extraction subprogram 1121 will be described with reference to the flowchart of FIG.

  First, in step 1401, text is read and morphological analysis processing is performed. An example of the morphological analysis result is shown in FIG. The morphological analysis result is obtained by adding the part of speech to the result of dividing the text into words. The morphological analysis result is assumed to be temporarily stored in the memory, but may be temporarily stored in a file or the like. Note that the processing after step 1402 may be performed while performing morphological analysis in sentence units, paragraphs, files, or the like.

  In step 1402, it is determined whether or not all words in the morphological analysis result have been processed. If all processing has been completed, the entire processing ends. If there is an unprocessed word, the process proceeds to step 1403. The determination may be made by sequentially processing the first word and the second word among all the words.

  In step 1403, focusing on the i-th word, the part-of-speech string of a nearby word is checked against a predetermined part-of-speech pattern. An example of a part of speech pattern is shown in FIG. Pattern 1 is a pattern for extracting a subsequent verb as a context for a word of interest, and represents extracting a part-of-speech sequence in which a noun is followed by a particle and further a verb. Pattern 2 is a pattern for extracting the adjective / adjective verb that appears immediately before the word of interest as the context, and indicates that the adjective or a noun is followed by an adjective verb. ing. In the figure, (T) after the part of speech indicates that the word is an attention word, and (C) indicates that it is a context word (sequence).

  If the pattern matches the morpheme analysis result, the process proceeds to step 1404. Based on the matching result, the morpheme analysis result that matches the attention word of the pattern and the morpheme analysis result that matches the context word (sequence) are extracted and stored in the context matrix. To do. A context matrix is created for each pattern.

  For the morphological analysis result of FIG. 12, when i is 1, the word string “computer”, “on”, “activate”, and when i is 6, “window”, “ga”, “ The word string “appears” is extracted by pattern 1. Further, the word strings “new” and “window” are extracted by the pattern 2. In addition, by distinguishing the attention word and the context word in the pattern, “activate” is extracted as the context for the attention word “computer” from each extraction result. In addition, “appears” is extracted as the context for the attention word “window”. Similarly, “new” is extracted as the context for the attention word “window”.

  The context matrix can be created by the above processing. Since the context matrix is created for each pattern, the similarity obtained from each context matrix has a different feature. In addition, equation (1) includes a constant for document length normalization, but this constant cannot be determined automatically. Therefore, this value is changed to an appropriate value between 0 and 1, and the similarity is calculated. For example, the constant is calculated with four types of values of 0.1, 0.3, 0.5, and 0.7, and the context matrix includes two context matrices corresponding to the two types of patterns shown in FIG. Assume that the similarity is calculated using. In that case, 4 × 2 = 8 types of features are obtained.

(2) Notation Base Similarity Hereinafter, a method for calculating the notation base similarity will be described. The notation-based similarity is calculated for a set of words based on character information. In the case where synonyms are particularly different notations such as “computer” and “computer”, as disclosed in Non-Patent Document 2, since many characters are duplicated, the ratio of overlapping characters Can be used as similarity. Different notation words are often in katakana, but there are cases where the same characters are included, such as “analysis” and “analysis”, “trust” and “trust”, in addition to different notation words consisting of kanji. Therefore, the degree of similarity is calculated based on the overlapping degree of characters, not limited to Katakana. Hereinafter, the similarity based on the overlapping ratio of characters is referred to as a character overlapping degree. In the case of a word composed of Kanji characters, especially in the case of a word with a short number of characters such as a two-character word, there are many words having different meanings even if they include the same character, such as “analysis” and “dialysis”. In the present invention, the character overlap degree works effectively by combining with different kinds of similarities such as context-based similarity.

  Further, in the case of kanji, there are characters that have similar meanings even if they are different characters. For example, characters such as “U” and “Yu” have similar meanings. If such character similarity can be learned from the teacher data, the notation base similarity between words can be calculated even if the characters do not completely match. Word similarity based on character similarity is called similar character overlap.

(A) Character overlap The character overlap can be calculated by various methods. Here, as an example, the number of characters included in common between two words is counted. A method of calculation by normalizing the character string length of the shorter word will be described. When a plurality of the same characters are included, m corresponds to one, and when n is included in the other word, there is an m-to-n correspondence. In such a case, it is assumed that the smaller number of characters m or n overlaps.

  Below, the calculation method of the notation base similarity of two words i and j is demonstrated using FIG.

  In step 1411, it is checked whether or not all characters of word i have been processed. If so, go to Step 1415. If there is an unprocessed character, the process proceeds to step 1412. In step 1412, it is checked whether all characters of word j have been processed. If so, the process proceeds to step 1411. If there is an unprocessed character, the process proceeds to step 1413.

  In step 1413, the m-th character of word i and the n-th character of word j are compared to determine whether they match. If they match, the process proceeds to step 1414. If not, the process proceeds to step 1412. In step 1414, a flag is set for each of the mth character of word i and the nth character of word j. Thereafter, the process proceeds to Step 1412.

  In step 1415, the number of characters with the flags of word i and word j are counted, and the smaller one is set as the number of matching characters. For example, assuming that “window” and “window” are to be processed, the three characters “c”, “n”, and “do” match. As for “c”, two characters are included in the “window”, so that 4 characters are flagged in the “window”, and 3 characters are flagged in the “window”. Therefore, it is assumed that the three characters match.

  In addition to the above method, a character string to be normalized which has a common partial character string length from the beginning of two words as a degree of duplication and a common partial character string length from the end of two words as a degree of duplication Variations such as taking the length as the average of the both and the longer are considered. Further, as a more precise method, for example, it is possible to match two words by DP matching or the like and calculate the notation base similarity based on the number of matched characters, depending on the available calculation resources, A larger number of notation-based similarities can also be calculated. Also, the weight when the characters match can be changed based on the frequency of the characters. In document search, IDF (Inversed Document Frequency) is known as a method for calculating word weights, but it is considered that characters included in many words are less important in the same way. Can calculate the weight of the character.

(B) Similar Character Duplication Degree The character similarity is learned from the synonym dictionary, and the character duplication degree is calculated including similar characters. A method of calculating the similarity of characters will be described with reference to the flowchart shown in FIG.

  In step 1421, word pairs that are synonyms are acquired from the synonym dictionary. Next, in step 1422, character pairs made up of characters extracted from one word of the word pair and characters extracted from the other word are acquired for all combinations. For example, in the case of a word pair in which “respect” and “respect” are synonyms, “respect” / “reward”, “respect” / “reel”, “rear” / “reward”, “reel” / “reel” 4 types of character pairs are acquired.

  Next, proceeding to step 1423, the frequency of characters included in all words in the synonym dictionary is calculated. Next, proceeding to step 1424, character similarity is calculated for all character pairs. The character similarity is obtained by dividing the frequency of a character pair by the frequency of two characters constituting the character pair (Dice coefficient). Self-mutual information amount or the like may be used as the similarity.

  In step 1425, with respect to the similarity calculated in step 1424, the similarity for the character different from the similarity for the same character is normalized. Specifically, the average AS of similarity for the same character and the average AD of similarity for different characters are respectively calculated. For the same character, 1.0 is set regardless of the calculated similarity. For different characters, the value obtained by multiplying the value calculated in step 1424 by AD / AS is used as the final similarity. An example of the character similarity table is shown in FIG.

  It is possible to calculate the similar character overlap degree using the character similarity table. The similar character overlap degree may be calculated in the same manner as the character overlap degree. In the case of different characters, the number of characters is added by 1 when the characters match in the character overlap, whereas in the case of the similar character overlap, the similar character table is referred to. It is a point to add character similarity. When the characters match, 1.0 is stored in the similar character table, and thus the character overlap is the same.

(3) Co-occurrence-based similarity The co-occurrence-based similarity indicates a high possibility of appearing simultaneously in the text. Usually, synonyms are said to be difficult to appear at the same time. For example, it is recommended to use one of the different notations such as “computer” and “computer”, and it is rare that both notations appear at the same time in the same document. However, abbreviations such as “European Union” and “EU” are often used simultaneously in the same text. Therefore, the co-occurrence frequency can be a clue for extracting synonyms.

  In the morpheme analysis result, paying attention to the i-th word, all the co-occurrence of the word and the attention word appearing at positions within the predetermined N words from the attention word are extracted and stored in the co-occurrence frequency table. An example of the co-occurrence frequency table is shown in FIG. In addition, the appearance frequencies of the individual words that have appeared are simultaneously calculated and stored in the word frequency table. An example of the word frequency table is shown in FIG. For example, a Dice coefficient is calculated as the co-occurrence base similarity from the values of the word frequency table and the co-occurrence frequency table. The Dice coefficient is F (A, B) / (f (A) + f (), where the frequencies of the words A and B are f (A) and f (B), respectively, and the co-occurrence frequency is F (A, B). B)). In addition, other measures such as the amount of self-mutual information can be used, or a plurality of types can be used. FIG. 19 shows an example of the co-occurrence similarity table 118.

  Through the above processing, synonyms can be extracted with higher accuracy than in the prior art. An example of the result is shown in FIG. FIG. 20 shows a comparison result between the conventional method (unsupervised learning based on context-based similarity and supervised learning using context words) and the method of the present invention. About 10 GB Japanese text collected from the Web was used. The average accuracy rate, which is an evaluation index, is a measure that is usually used in evaluating document search accuracy, and it is a comprehensive measure of accuracy rate (a measure that indicates low noise) and recall rate (a measure that indicates low leakage). It is a scale for judging. The precision ratio and the recall ratio are usually in a trade-off relationship, and when the parameter is changed in the same method, one becomes better and the other becomes worse. For example, in a certain synonym extraction method, when the number of synonym candidates to be extracted is increased, the recall is improved (leakage is reduced), but the precision is deteriorated (noise is increased). Therefore, in comparison between methods, it is meaningless to simply compare only the precision. In the average precision, the precision can be accurately compared by acquiring the precision at each recall and changing the recall as 10%, 20%, and 30%.

  # 1 corresponds to the method disclosed in Non-Patent Document 1, and # 2 corresponds to the method disclosed in Non-Patent Document 6. It can be seen that the supervised method including the conventional method # 2 is superior to the unsupervised method # 1. In comparison between supervised methods, it can be seen that the proposed method # 3 using similarity as a feature is more accurate than the conventional method # 2 using context words as a feature. Also, in addition to the context-based similarity used in # 3, the accuracy may be improved by using different features such as character redundancy (# 4) and similar character redundancy (# 5) in combination. I understand.

  In the above description, the method of using a word pair not included in the synonym dictionary as a negative example in the process of step 16 in FIG. 4 has been described. This method is a method for avoiding the problem that a word pair that is not included in the synonym dictionary is not necessarily a synonym. As another method, this problem can be avoided by using 1-class SVM as a discriminator. 1-Class SVM is a technology that can learn classifiers only from positive examples. Hideki Aso, Koji Tsuda, Noboru Murata “New concepts and methods of pattern recognition and learning, frontier of statistical science” Iwanami Shoten (2003), the description is omitted. In the case of using 1-class SVM, in the process of step 16 in FIG. 4, only the row assigned “correct” as a label is used as teacher data, and learning is performed using 1-class SVM as a discriminator. . This makes it possible to configure the discriminator only from the positive example, that is, information relating to the word pair included in the synonym dictionary.

  Thus, according to the synonym extraction device of the first exemplary embodiment of the present invention, a synonym dictionary including synonyms that are not included in the existing synonym dictionary is output.

[Second Embodiment]
Hereinafter, a thesaurus extraction apparatus according to a second embodiment of the present invention will be described with reference to the drawings. In the first embodiment, the word semantic relationship extraction problem is solved as a problem for identifying whether a synonym or not. However, there is a more ambiguous situation in actual word semantic relationship extraction. For example, the upper and lower terms are not synonyms in the strict sense, but the meanings are similar. For example, “company” and “manufacturer” correspond. The same applies to siblings, that is, words having a common word as a broader word. For example, “securities company” and “bank” are equivalent.

  In the second embodiment, it is possible to realize a word meaning relationship extraction device that can appropriately handle such a situation. In the second embodiment, the problem is solved by treating the word semantic relationship extraction problem not as a binary identification problem but as a ranking problem. In other words, in the case of synonyms, 1 is given as the rank because it is very similar, and in the case of upper / lower terms and siblings, it is not as synonymous, but it is ranked as if it is somewhat similar. If 2 is assigned and none of them is given, it is considered that the similarity is low and 3 is assigned as a rank. Then, as in the first embodiment, word semantic relationship extraction is performed by learning a function for ranking from teacher data to which a rank is assigned as a correct answer by a manually created dictionary.

  In the second embodiment, Step 16, Step 17, and Step 18 in FIG. 4 of the first embodiment are changed as follows.

  First, the change in step 16 will be described. In the first embodiment, the synonym dictionary is referred to, and a binary label “+1” is set for a positive example and “−1” is set for a negative example. However, unknown word pairs are excluded here. In the second embodiment, a label is set by referring to a thesaurus dictionary including upper / lower relations of words. If a set of words is synonymous with reference to the thesaurus dictionary, “1” is assigned as a label. If the word set is a broader / lower word or sibling, “2” is assigned as a label. The concept of processing in other cases is the same as in the first embodiment. That is, when the word set is not included in the thesaurus dictionary, but each word is included in the thesaurus dictionary, “3” is assigned as the incorrect answer label. If any one of the word pairs is not included in the synonym dictionary, an unknown (-1) label is assigned.

  FIG. 21 shows an example of the similarity matrix in the second embodiment. In the label field, 1 for synonyms such as <computer, computer>, 2 for broader and lower-order words such as <machine, computer>, and none of the above words such as <computer, virtualization technology> Is different from the first embodiment in that a rank of 3 is assigned.

  Step 17 is changed to perform ranking learning rather than binary identification model learning. The classifier that performs ranking learning is disclosed in, for example, T. Joachims, Training Linear SVMs in Linear Time, Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), 2006.

  Step 18 is different in that the value to be set is not a binary value but a value indicating a rank determined according to the learned model. Further, since the set value is not binary, the screen is different even in the dictionary editor. For example, the correction can be performed by performing the display as shown in FIG. In the example of FIG. 22, the ranks given as labels and the word pairs whose determination result ranking is larger than a certain threshold are displayed, and items corresponding to the ranks given as labels as initial values (in the case of FIG. 22) , “Synonyms”, “higher / lower terms”, or “other than that”). When it is determined that the user is wrong, the check is added again, and the dictionary is corrected by reflecting only the portion where the check is changed in the dictionary.

  Thus, according to the thesaurus extraction device of the second exemplary embodiment of the present invention, a thesaurus dictionary including synonyms, broader / lower terms, and siblings not included in the existing thesaurus dictionary is output.

[Third Embodiment]
Hereinafter, a parallel translation extracting apparatus according to a third embodiment of the present invention will be described with reference to the drawings. In the third embodiment, parallel translation relationships between different languages are extracted as word relationships. A bilingual relationship can be viewed as an extension of a synonym relationship between words in different languages. Therefore, it is possible to perform parallel translation relationship extraction based on the same concept as in the first embodiment. In the third embodiment, a system configuration similar to that in the first embodiment is used. However, the configuration differs from that of the first embodiment in that a bilingual dictionary is used instead of the synonym dictionary. An example of the bilingual dictionary 1143 is shown in FIG. The bilingual dictionary has exactly the same format as the synonym dictionary, and translations are stored instead of synonyms.

  FIG. 24 shows an example of a similarity matrix in the case of parallel translation extraction. In the example of FIG. 3, the word pair is made up of a pair of words in the same language, whereas in the example of FIG. 24, a word pair made up of a word in the first language and a word in the second language is stored. .

  The overall processing flow is the same as the flowchart of FIG. However, the details of the processing in step 13 and step 14 are slightly different.

  In step 13, the realization method when acquiring a word pair is different. In the first embodiment, arbitrary different word pairs are extracted from all words in the same language to form word pairs, whereas in this embodiment, the first language word and the second language Get word pairs by word combinations. Specifically, an arbitrary word is obtained from a list of words obtained by morphological analysis of text in the first language and a list of words obtained by morphological analysis of text in the second language, To do.

  In step 14, the similarity calculation method for the word pair is different. Hereinafter, the similarity calculation method in bilingual extraction will be described in detail.

(1) Multilingual context-based similarity In the case of parallel translation extraction, the two words constituting a word pair are different languages. In the following description, it is assumed that one is Japanese and the other is English. Therefore, the context of each word is also a different language. Therefore, the similarity cannot be calculated by matching the context word string. At this time, by using the bilingual dictionary, it is possible to calculate the context-based similarity by associating words in the context with each other in the same manner as in the case of synonym extraction.

  FIG. 25 and FIG. 26 show examples of context matrices in parallel translation extraction. FIG. 25 is an example of a context matrix extracted from Japanese text, and FIG. 26 is an example of a context matrix extracted from English text. The difference from the case of synonym extraction is that in FIG. 25, only verbs are extracted as context without including particles. This is because, in English, there is no particle, and since correspondence is performed in the bilingual dictionary, the character string including the particle is not normally included in the dictionary. However, the point where the particle does not exist can be used in place of the particle by performing a case analysis of the main case, the objective case, etc. by a technique such as syntax analysis.

  By preparing a context matrix for each language and making correspondence between context information using a bilingual dictionary, the similarity based on the context can be calculated as in the first embodiment. For example, the bilingual dictionary shows that “boot” corresponds to “boot”, “stop” corresponds to “shutdown”, etc., so the similarity is calculated from the context information of “computer” and “computer”. be able to.

(2) Multilingual notation base similarity For foreign words in Katakana, a technique for estimating a bilingual relationship based on pronunciation is known. This type of technology is called Transliteration and is disclosed in, for example, K. Knight and J. Graehl: Machine Transliteration, Computational Linguistics, 24 (4), pp. 599-612, 1998. As a simple method, information that “co” can be read as “co”, “m” as “n” or “mu”, and “pu” as “pu” or “pyu” is prepared. , Reading candidates such as “Computer”, “Computer”, and “Computer” are generated from “computer”, and the reading candidates and character strings of Japanese words are generated by the method described in the first embodiment. Similarity can be calculated by comparison.

(3) Multilingual Co-occurrence Based Similarity In the case of parallel translation extraction, it is not possible to obtain from text alone whether Japanese words and English words co-occur as in the case of context-based similarity. Therefore, the co-occurrence base similarity is calculated using the bilingual dictionary. Specifically, the co-occurrence base similarity is calculated from the Japanese text and the English text, respectively, and a co-occurrence similarity table is created. When a bilingual word pair is given, one of the word pairs is converted by the bilingual dictionary and collated with the co-occurrence similarity table. Specifically, the Japanese word of the word pair is converted into English by a bilingual dictionary, and collated with an English co-occurrence similarity table to obtain the similarity. If there are multiple candidates, all are acquired. Similarly, the English word of the word pair is converted into Japanese by the bilingual dictionary and collated with the Japanese co-occurrence similarity table to obtain the similarity. Through the above processing, multilingual co-occurrence-based similarity can be calculated.

  In addition, although a plurality of similarities can be obtained by the above processing, all similarities are calculated, the average of the similarities obtained by English conversion of Japanese words, the similarity obtained by Japanese conversion of English words Variations are conceivable, such as using the average two types. Since which method is suitable depends on the size of the bilingual dictionary and the size of the text, an appropriate method may be adopted depending on the data to be applied.

  Thus, according to the bilingual relationship extraction apparatus of the third exemplary embodiment of the present invention, a bilingual dictionary including words in a bilingual relationship that is not included in the existing bilingual dictionary is output.

100 word meaning relation extraction apparatus 101 CPU
102 Main memory 103 Input / output device 110 Disk device 111 OS
112 Word meaning relation extraction program 1121 Feature vector extraction subprogram 1122 Correct label setting subprogram 1123 Identification model learning subprogram 1124 Identification model application subprogram 113 Text 114 Manual creation dictionary 1141 Synonym dictionary 1142 Thesaurus dictionary 1143 Bilingual dictionary 115 Similarity matrix 116 Context Matrix 117 Part-of-Speech Pattern 118 Co-occurrence Similarity Table 119 Identification Model 120 Character Similarity Table

Claims (9)

Means for generating a feature vector having a plurality of different similarities as elements for a set of words extracted from text;
Means for referring to a known dictionary and assigning a label indicating a word semantic relationship to the feature vector;
Means for learning a word semantic relationship determination rule based on a plurality of feature vectors to which the label is attached;
Means for determining a word semantic relationship for an arbitrary set of words based on the learned word semantic relationship determination rule;
A word meaning relationship extraction device comprising: In the word meaning relationship extraction device according to claim 1,
The means for generating the feature vector includes:
Means for extracting a word in the vicinity of the appearance location in the text of the word of interest as context information of the word of interest;
Means for calculating the similarity between the context information of two words of the word set as the similarity of the word set;
A word meaning relationship extraction device comprising: In the word meaning relationship extraction device according to claim 1,
The means for generating the feature vector includes:
Means for calculating a correspondence between characters included in two words of the set of words based on whether or not they are the same character;
Means for calculating the similarity of the set of words based on the correspondence between the characters;
A word meaning relationship extraction device comprising: In the word meaning relationship extraction device according to claim 1,
The means for generating the feature vector includes:
Means for determining the similarity of characters contained in two words of the set of words;
Means for calculating similarity of the set of words based on similarity of the characters;
A word meaning relationship extraction device comprising: In the word meaning relationship extraction device according to claim 1,
The means for generating the feature vector includes:
Means for extracting two words appearing within a certain distance from the text as a set of co-occurring words;
Means for calculating a statistic indicating the co-occurrence of words using the frequency of the co-occurring word sets as the similarity of the word sets;
A word meaning relationship extraction device comprising: In the word meaning relationship extraction device according to claim 1,
The word semantic relationship is a relationship as to whether or not two words in the set of words are synonyms,
2. The word semantic relationship extracting apparatus according to claim 1, wherein the known dictionary is a synonym dictionary storing entry words and their synonyms. In the word meaning relationship extraction device according to claim 1,
The word semantic relationship is whether the two words of the word set are synonyms, upper / lower relationships, sibling relationships, or neither.
2. The word semantic relationship extraction device according to claim 1, wherein the known dictionary is a thesaurus dictionary storing headwords and their synonyms, upper / lower terms, or siblings. In the word meaning relationship extraction device according to claim 1,
The word semantic relationship is a parallel translation relationship of two words of the word set,
2. The word semantic relationship extracting apparatus according to claim 1, wherein the known dictionary is a bilingual dictionary storing headwords and their translations. In the word meaning relation extraction device according to any one of claims 1 to 8,
Means for determining a label that is highly likely to be incorrect based on information on the assigned label and the determined word semantic relationship;
Means for displaying information about the label likely to be wrong;
Means for accepting user input and correcting the erroneous label;
A word meaning relationship extraction device comprising:

Images