JP3998664B2 - Parallel word classification system, parallel word classification method, and parallel word classification program - Google Patents

Description

  The present invention relates to a bilingual phrase classification system, a bilingual phrase classification method, and a bilingual phrase classification program. For example, the present invention relates to an apparatus, a method, and a program for automatically determining and classifying a field of words in natural language processing.

  In natural language processing, it is necessary to continue to register words in the dictionary in order to cope with new words that are born every day. When classifying words in a dictionary, it is difficult to keep them appropriate and uniform manually.

  For example, it may be used as a specialized dictionary in machine translation. In this case, in order to achieve the purpose of a specialized dictionary, not only words are classified but also hierarchically constructed, and classified into the narrowest possible field, that is, into the lower field as much as possible. Is effective, but in this case, it is difficult to determine in which hierarchy it is most appropriate to classify words. In addition, when classifying words, it is necessary to take into account the translations of the words, so it is still more difficult to classify them.

  In Non-Patent Document 1, unregistered noun phrases collected at the time of translation are hierarchized into two levels called main field and sub-field from information on how many times the pattern included in the field dictionary is hit. A technique for classifying into a dictionary by field is described.

  In Non-Patent Document 1, first, calculation is performed in all sub-fields, and unique and high-frequency items are classified into sub-field dictionaries. After removing them, the same calculation is performed in all main fields and classified into the main field dictionary. It is described that this calculation is performed as a vector in a vector space having a field as an element. Moreover, since the vocabulary of the field-specific dictionary is increased by registering words in the dictionary, the field estimation accuracy can be improved.

Non-Patent Document 1 describes a translation of a word by causing a system that does not know the translation to generate a translation of the word or manually classifying the translation.
Goro Kamiyama and Harumi Ito, “Machine Translation System for Autonomous Vocabulary Expansion”, IPSJ 65th National Convention, 2003, pp. 2-5 to 2-6

  However, in the prior art described in Non-Patent Document 1 described above, when the translated words are classified, for example, even if the translated words are given manually, the meaning of the translated words cannot be considered at the time of field estimation. Since it is not possible to estimate the field by distinguishing the words that you want to change, you may not be able to classify correctly.

  In addition, since the field estimation depends on the contents of the field dictionary, if words that cannot be correctly determined are registered and the vocabulary of the field dictionary increases, there may be cases where the field estimation cannot be correctly determined thereafter. In this case, as the words that could not be judged correctly are registered, the quality of the field-specific dictionary may deteriorate.

  Furthermore, in Non-Patent Document 1 described above, it is necessary to repeat the calculation for each layer. On the other hand, in order to improve the translation quality of machine translation, more hierarchies are desired. Therefore, the number of calculations increases according to the number of hierarchies as the hierarchy becomes deeper, and the calculation method depends on the registered words in the domain dictionary. Becomes enormous.

  Therefore, in order to solve the above-described problems, the translated word / phrase classification system and the translated word / phrase classification for efficiently classifying the words while determining the original word to be determined and the category of the translated word and considering the ambiguous translated word for the original word What is needed is a method and bilingual word classification program.

  In order to solve such a problem, the bilingual phrase classification system according to the first aspect of the present invention determines the original word of the input judgment target word and the category of the translation word, and classifies the judgment target word into categories. In the classification system, (1) one or a plurality of category representative words that characteristically express a certain category are created, and a field relevance degree that indicates the feature level of each category representative word in a document classified by category is created. Category representative word creating means, (2) Category representative word holding means for holding each category representative word created by the category representative word creating means and its field relevance, and (3) 1 co-occurring with the determination target original word Alternatively, a plurality of words are extracted from the original uncategorized document, and among the extracted words having the co-occurrence relationship, the field relevance of one or more words that are category representative words is obtained. (4) One or a plurality of words co-occurring with the judgment target translation word are extracted from the translation uncategorized document, and extracted from the category representative word holding means and creating a field judgment degree list of the judgment target source word. Then, one or more words that are the source language translations of the extracted words in the co-occurrence relationship are determined, and the field relevance of one or more words that are the category representative words among the words that are the source language translations is categorized. (5) each category representative word in the field judgment degree list of the judgment target source word and the field judgment degree list of the judgment target translation word, which is extracted from the representative word holding means and creates a field judgment degree list of the judgment target translation word And (6) a bilingual determination unit that creates an integrated field determination level for each category of a determination target word, and (6) a determination based on the integrated field determination level for each category created by the parallel determination unit. And an outputting means for outputting the category classification information for elephants word.

  The bilingual phrase classification method of the second aspect of the present invention corresponds to the bilingual phrase classification system of the first aspect of the present invention. That is, the bilingual phrase classification method according to the second aspect of the present invention is the bilingual phrase classification method for determining the original word of the input determination target word and each category of the translation word, and classifying the determination target word by category. The category representative word creation means creates one or a plurality of category representative words that characteristically represent a certain category, and creates a field relevance level indicating the characteristic level of each category representative word in a document classified by category. (2) The category representative holding means holds each category representative word created by the category representative word creating means and the field relevance thereof, and (3) the original word judging means is co-occurrence with the judgment target original word 1 or A plurality of words are extracted from the original language uncategorized document, and the category relevance of the category relevance of one or more words that are category representative words is extracted from the extracted co-occurrence words. (4) Translation word determination means extracts one or more words co-occurring with the determination target translation word from the unclassified translation document, and extracts the extracted words One or a plurality of words that are the source language translations of each word in the co-occurrence relationship are determined, and the category relevance of the field relevance of one or more words that are the category representative words among the words that are the source language translations is retained The field judgment degree list of the judgment target translation is extracted from the means, and (5) the field of each category representative word in the field judgment degree list of the judgment target original word and the field judgment degree list of the judgment target translation The determination degrees are integrated, and an integrated field determination degree for each category for the determination target word is created. (6) The output means determines the determination target word based on the integrated field determination degrees for each category created by the parallel translation determination means. Category classification And outputting a broadcast.

  The parallel word / phrase classification program according to the third aspect of the present invention corresponds to the parallel word / phrase classification system according to the first aspect of the present invention. That is, the parallel word / phrase classification program according to the third aspect of the present invention provides a parallel word / phrase classification device that determines the original word of the input determination target word and each category of the translation word, and classifies the determination target word by category (1). Category representative word creating means for creating one or a plurality of category representative words that characteristically express a certain category, and creating a field relevance level indicating a characteristic degree of each category representative word in a document classified by category; (2) Category representative word holding means for holding each category representative word created by the category representative word creating means and its field relevance, and (3) one or more words co-occurring with the determination target original word The category relevance measure is used to extract the field relevance of one or more words that are category representative words out of the extracted co-occurrence words from the classified document. And (4) extracting one or a plurality of words having a co-occurrence relationship with the determination target translation from the uncategorized document and extracting the co-occurrence One or a plurality of words that are the source language translations of each related word are determined, and the field relevance level of the one or more words that are the category representative words out of each word that is the source language translation is extracted from the category representative word holding means. And (5) integrating the field judgment degree of each category representative word in the field judgment degree list of the judgment subject original word and the field judgment degree list of the judgment subject translation word. And (6) the category classification information of the determination target word based on the category-specific integrated field determination degree created by the parallel translation determination unit. output Characterized in that to achieve an output unit that.

  According to the parallel word / phrase classification system, parallel word / phrase classification method, and parallel word / phrase classification program of the present invention, the ambiguity between the original word and the translated word is resolved regardless of the depth of the category, regardless of the degree of the contents of the dictionary. Can categorize bilingual words and phrases.

  The best mode for carrying out the bilingual phrase classification system, bilingual phrase classification method, and bilingual phrase classification program of the present invention will be described below with reference to the drawings.

  In the embodiment described below, a case will be described in which the bilingual phrase classification system, the bilingual phrase classification method, and the bilingual phrase classification program of the present invention are applied to a bilingual phrase classification system that performs field determination of bilingual phrases and classifies them into a dictionary. .

  In the following embodiments, on the source language side, typical words that are characteristic to the field are prepared in each field, the source language is determined using the words, and the source language is determined from those words and the translated word side. By using the word dictionary to the side to determine the field of translations and integrating the field determination results, the ambiguity between the original word and the translated word is resolved regardless of the level of the dictionary contents, regardless of the depth of the hierarchy. So, it was possible to classify parallel translation words and phrases.

(A) First Embodiment A first embodiment of a bilingual phrase classification system, a bilingual phrase classification method, and a bilingual phrase classification program according to the present invention will be described below with reference to the drawings.

  In the following description, a typical word characteristic to a category (field) is defined as a core word. In addition, co-occurrence is that a plurality of elements appear simultaneously within a certain range (for example, words, sentences, paragraphs, sentences, etc.). Words and phrases that appear at the same time are called co-occurrence words or phrases. A word (phrase) that co-occurs with a core word is defined to be in the same category as the core word. In the example, a word in which unnecessary words are excluded from words that appear simultaneously in a sentence up to a line feed is a word having a co-occurrence relationship. Use unnecessary words other than nouns, verbs, adjectives, adjective verbs, and unknown words.

  In order to extract a word having a co-occurrence relationship with a certain word, a sentence including the certain word is searched, and a sentence up to a line feed is extracted and morphological analysis is performed.

  A value indicating the degree of belonging to the category is assigned to the core word. The value is defined as the field relevance. It is assumed that the greater the field relevance, the stronger the degree of belonging to the category.

  A parallel translation consists of an original word and a translated word. The original language desired to be determined as the field is defined as the determination target original word, the translated word desired as the field determination is defined as the determination target translated word, and the parallel translation word desired as the field determination is defined as the determination target word. In the example, the original language is Japanese and the translation is English.

  For the judgment target original word, judgment target translation or judgment target word, the value of the core word indicating the degree belonging to the category is defined as the field judgment degree, and the combination of the category, the core word, and the field judgment degree is a list of the field judgment degrees. It is defined as

  For the determination target word, a value indicating a degree belonging to a category is defined as an integrated field determination degree, and a set of the category and the integrated field determination degree is defined as a list of integrated field determination degrees.

(A-1) Configuration of the First Embodiment FIG. 1 is a block diagram showing a functional configuration of the bilingual phrase classification system of this embodiment. The bilingual phrase classification system 100 according to the first embodiment is realized, for example, by installing a bilingual phrase classification program on an information processing apparatus such as a personal computer provided with input / output means. Can be expressed as

  In FIG. 1, the bilingual phrase classification system 100 of the first embodiment includes an input unit 1, a core word creation unit 2, a source language determination unit 3, a translation word determination unit 4, a translation translation determination unit 5, an output unit 6, and a source language classified document 7. , A core word dictionary 8, an original word unclassified document 9, a translated word unclassified document 10, a word dictionary 11, and a category dictionary 12.

  For general processing such as morphological analysis processing, search processing, extraction processing, database processing, and storage processing described below, known natural language processing techniques can be used.

  For example, the input unit 1 is not only a general input unit such as a keyboard, but also a file reading device such as a recording medium access device, a character recognition device that reads a document as image data and replaces it with text data, and the like. It is a means for inputting a document or a word to be classified, and also indicates an operation mode or the like as appropriate.

  The core word creating means 2 is a means for creating the core word on the original language side and the field relevance. The core word creation means 2 performs morphological analysis of the document of the source language classified document 7 for each category, extracts words excluding unnecessary words from the morpheme analysis result, and creates core words for each category. The core word creation means 2 calculates field relevance and stores the created core word and field relevance in the core word dictionary 8.

  Here, as described above, the field relevance may be a value that can indicate the degree to which the core word belongs to a category (the degree that expresses the characteristics of the category well). Can be sought. In the present embodiment, the field relevance is calculated using a value calculated by the tf * idf method.

  In the tf * idf method, the appearance frequency tf (term frequency within a document: Term Frequency) of a word (here, a core word) in a document d in the document set, and the number of documents df (documents) in which each word t appears one or more times. Frequency: Document Frequency), and the number of documents df is calculated by the following equation (1) to obtain idf (Inverse Document Frequency).

idf (t) = log (N / df (t)) (1)
Here, N is the number of documents.

  The component w (t, d) that determines the field relevance is defined by the following equation using this idf and tf.

w (t, d) = tf (d, t) * idf (t) (2)
The reason for using tf is that a word that repeatedly appears in a document is considered to be an important word in the document. Idf indicates the ability of the word (here, the core word) to identify the document in the document set. This is because idf is small for a general word appearing in many documents in the document set, and conversely, idf is large for a word appearing only in a specific document.

  Since tf * idf can take a larger value as the number of transmitted words is larger, adjustment between categories is necessary. Therefore, for each category, a value obtained by dividing tf * idf by the total number of core words is set as the field relevance.

Field relevance (category, core word)
= Tf * idf / total number of core words per category (3)
The original language determination means 3 is a means for creating a list of field determination degrees of the determination target original language. The original language determination means 3 takes in the input determination target original language, searches the original language document of the original language uncategorized document 9 for the determination target original language, and extracts a word having a co-occurrence relationship with the determination target original language from the original language document It is. Further, the original word determination means 3 searches the extracted word having the co-occurrence relationship from the core word dictionary 8 and calculates the field determination degree by weighting the number of appearances with respect to the field relevance degree of the searched core word. A list of field judgment degrees of the target source language is created.

Judgment level of target language (source language, category, core word)
= Field relevance (category, core word)
× Number of appearances (original words to be judged, core words) (4)
The translated word judging means 4 is a means for creating a list of field judgment degrees of judgment target translated words. The translated word judging means 4 searches the translated word document of the translated word uncategorized document 10 for the inputted judgment target translated word, and extracts the translated word co-occurring with the judged target translated word from the translated word document. Further, the translation word judging means 4 searches the word dictionary 11 for the original word translation of the extracted words having the co-occurrence relationship, extracts the words excluding unnecessary words from the result of the morphological analysis of the source word translation, and extracts the extracted words. A search is made from the core word dictionary 8, the number of appearances is weighted with respect to the field relevance level of the searched core word, the number of source word translations is weighted, a field determination degree is calculated, and a field determination degree of the target translation target word Output a list of

  The reason why the original word translation is subjected to morphological analysis and unnecessary words are removed is to match the core word.

Judgment level of target translation (Judgement target category, core word)
= Field relevance (category, core word)
× Number of appearances (translation target words, core words)
/ Number of source language translations (translations subject to judgment) (5)
In addition, as shown in the above formula (5), as the weighting of the number of source language translations, by dividing the field relevance of matching words by the number of source language translations for the translations, On the other hand, the fact that there are a plurality of source language translations can affect the number of appearances.

  The parallel translation determination means 5 is a means for calculating the integrated field determination degree of the determination target word. The parallel translation determination means 5 takes in the field determination degree list of the determination target original words created by the source word determination means 3 and the field determination degree list of the determination target translation words created by the translation word determination means 4 and integrates them to determine the determination target words The field determination degree is calculated to create a list of field determination degrees of the determination target words. Further, the parallel translation determination means 5 calculates an integrated field determination degree for each category of the determination target word from the created list of field determination degrees of the determination target word, and stores the integrated field determination degree for each category in the category dictionary 12. To do. The bilingual determination means 5 gives the output means 6 the category to which the determination target word determined in the field belongs and outputs it. The category output by this output means is the category with the highest integrated field determination.

  Here, the method of calculating the field determination degree of the determination target word is as follows.

  It is desirable that the field determination degree value of the determination target word be higher as the field determination degree on both the original word side and the translated word side is higher. However, the original word side and the translated word side simply compare the degree of field judgment because the content and quantity of the documents for extracting co-occurrence words are different and the co-occurrence words are ambiguous. Can not do it. Therefore, the field determination degree of the determination target original word and the field determination degree of the determination target translated word are divided by the maximum values of the respective field determination degrees, normalized, and a harmonic average is obtained.

Judgment degree of judgment target word (judgment word, category, core word)
= 2 × (V1 / M1) × (V2 / M2)
/ ((V1 / M1) + (V2 / M2)) (6)
Here, V1 is a field determination degree of the determination target original word (determination target original word, category, core word), M1 is a maximum value of a field determination degree of the determination target original word (determination target original word, category, core word), and V2 is a determination The field judgment degree (judgment target translation, category, core word) of the target translation word, and M2 is the maximum value of the field judgment degree (judgment target translation word, category, core word) of the judgment target translation word.

  In addition, as a method of calculating the integrated field determination degree of the determination target word, if the field determination degree is equal to or greater than a certain threshold, the field determination degrees are totaled for each category. The reason why the threshold value is cut off is to eliminate the adverse effect that the field judgment degree is lower than the number. The threshold value is an arbitrary value that can optimally determine the category. In the present embodiment, description will be made with the optimum value obtained by experiment.

Degree of integrated field judgment (category) of judgment target words
= Field judgment degree of Σ judgment target word (category, core word) (7)
However, the field determination degree of the determination target word ≧ α (α is a threshold value).

  The output means 6 is means for outputting the category of the input word. The output means 6 can correspond to not only general output means such as a display and a printer, but also a storage medium access device for storing in a recording medium. Further, the output means 6 can appropriately output the progress of the operation process in the operation mode, the result, and the like.

  The source language classified document 7 is a database that stores source language documents classified into categories.

  The core word dictionary 8 is a dictionary that stores core word information. The core word dictionary 8 describes pairs of categories, core words, and field relevance.

  The original language unclassified document 9 is a database that stores uncategorized original language documents.

  The translated word uncategorized document 10 is a database that stores unclassified translated word documents.

  The word dictionary 11 is a dictionary that stores information on the original translation of the translated word. The word dictionary 11 describes one or more source language translations for one translated word. FIG. 9 shows an example of information stored in the word dictionary 11. As shown in FIG. 9, the word dictionary 11 stores word information with translated words, parts of speech, and source language translations as item examples.

  The category dictionary 12 is a dictionary that stores information on categories of determination target words. The category dictionary 12 describes the category and the integrated field determination degree for the determination target word.

(A-2) Operation of First Embodiment Next, the operation of the bilingual phrase classification system 100 of this embodiment will be described with reference to the drawings.

  In order for the bilingual phrase classification system 100 of this embodiment to perform field determination, the original language document classified for each category is stored in the original language classified document 7 for each category in advance, and the core code and the field-related are previously stored. It is necessary to create a degree. Then, field determination is performed using various databases and various dictionaries.

  In the following, after describing the overall operation of the bilingual classification system 100 of the present embodiment, specific various processing operations will be described.

(A-2-1) Overall Operation FIG. 2 is a flowchart for explaining the overall operation of the bilingual phrase classification system 100.

  In the bilingual classification system 100, when the input unit 1 is operated by the user, various processes are selected, and the selected processes are operated.

  For example, as shown in FIG. 2, it is accepted that any one of a document storage process, a core word creation process, a bilingual field determination process, or an end (END) process is selected (step 201), and a user operation is input means. 1 takes in and executes various processes according to the operation (steps 202 to 204).

(A-2-2) Document Storage Processing The document storage processing operation in FIG. 2 will be described with reference to FIG. FIG. 3 is a flowchart for explaining the operation of the document storage process.

  In FIG. 2, when the document storage process is selected (Steps 201 and 202), as shown in FIG. 3, the selection of the process by the user is accepted (Step 221).

  The input means 1 captures the user's processing selection, and in accordance with the selection, specifies a category, inputs a source language document classified by category, and stores it in the source language classified document 7 (step 222). Whether a classified original language document is input and stored in the original language unclassified document 9 (step 223), whether an uncategorized translated word document is input and stored in the translated word unclassified document 10 (step 224), or end (END) Do what you want.

(A-2-3) Core Word Creation Processing The operation of the core word creation processing in FIG. 2 will be described with reference to FIG. FIG. 4 is a flowchart for explaining the operation of the core word creation process.

  In FIG. 2, when the core word creation process is selected (steps 201 and 203), the process of FIG. 3 is executed.

  In FIG. 3, the documents stored in the source language classified document 7 are morphologically analyzed for each category by the core word creating means 2 (step 231).

  When the original language document classified by category is subjected to morphological analysis, the core word creation means 2 extracts words excluding unnecessary words from the morphological analysis result, and creates core words for each category (step 232).

  When the core word for each category is created, as described above, the field relevance is calculated according to the above formulas (1) to (3) (step 233), and the created core word and the calculated field relevance are the core words. It is stored in the dictionary 8 (step 234).

(A-2-4) Parallel Translation Field Determination Process The operation of the parallel translation field determination process in FIG. 2 will be described with reference to FIG. FIG. 5 is a flowchart for explaining the operation of the bilingual field determination process.

  In FIG. 2, when the bilingual field determination process is selected (steps 201 and 204), the process of FIG. 5 is executed.

  In this bilingual field determination process, the following is considered. As a weighting that reflects the strength of the co-occurrence relationship, the number of appearances of the words in the extracted co-occurrence relationship is multiplied by the field relevance level that matches the words. In order to eliminate the influence of multiple source language translations on the translation to affect the number of occurrences so that the ambiguity of the translation can be handled, the field relevance level for matching words is used as the weighting of the number of source language translations. Divide by the number of source language translations. A value obtained by weighting the field relevance is set as the field determination degree.

  In FIG. 5, the input means 1 takes in the input determination target word (step 301).

  Then, the source language determination unit 3 calculates the field determination degree of the determination target original word from the document stored in the original word uncategorized document 9 and the core word and the field relevance level of the core word dictionary 8 according to the above-described equation (4). Thus, a list of field determination degrees of the determination target original language is created (step 302). This detailed operation will be described later with reference to FIG.

  Further, the translation determination unit 4 calculates the field determination degree of the determination target translation according to the above-described equation (5) using the document stored in the untranslated document 10 and the word dictionary 11 and the core word dictionary 8. Then, a list of field judgment degrees of judgment target translated words is created (step 303). This detailed operation will be described later with reference to FIG.

  Further, the bilingual determination unit 5 creates a list of field determination degrees of the determination target words from the list of field determination degrees of the determination target original words and the list of field determination degrees of the determination target translation words, according to the above formula (6). The integrated field determination degree for each category of the determination target word is calculated and stored in the category dictionary 12 (step 304). This detailed operation will be described later with reference to FIG.

  The output means 6 outputs the category to which the determination target word determined in the field belongs (step 305).

(A-2-5) Processing for calculating the field determination degree of the determination target original word FIG. 6 is a flowchart illustrating a processing operation for calculating the field determination degree of the determination target original word.

  The input means 1 takes in the input determination target original language (step 321). When the input unit 1 captures the determination target source language, the source language determination unit 3 retrieves the determination target source language from the source language document of the source language unclassified document 9 (step 322).

  When the determination target original word is retrieved by the original language determination unit 3, the source language determination unit 3 extracts a word having a co-occurrence relationship with the retrieved determination target original word from the original language document (step 323).

  When a word having a co-occurrence relationship with the determination target original word is extracted from the original document, the original word determination unit 3 searches the core word dictionary 8 for a word having a co-occurrence relationship with the extracted determination target original word (step 324). ).

  When a word having a co-occurrence relation with the determination target original word is retrieved from the core word dictionary 8, the field relevance level of the retrieved core word is extracted by the original word determination means 3 (step 325), and the extracted field relevance level Is weighted (multiplied) by the number of appearances according to the above equation (4) (step 326), the field determination degree of the determination target original word is calculated, and a list of the field determination degrees of the determination target original word is output (step 327). .

(A-2-6) Processing for calculating the field determination degree of the determination target translation FIG. 7 is a flowchart for explaining the processing operation for calculating the field determination degree of the determination target translation.

  The input target 1 is input by the input means 1 (step 331). When the input unit 1 fetches the determination target translation, the translation determination unit 4 searches for the determination target translation from the translation document of the untranslated document 10 (step 332).

  When the translation target judgment word 4 is searched for the judgment target translation word, the translation word judgment means 4 extracts a translation word co-occurring with the judgment target translation word from the translation word document (step 333).

  When a translation having a co-occurrence relationship with the determination target translation is extracted, the translation determination unit 4 searches the word dictionary 11 for an original translation of the translation having a co-occurrence with the extracted determination target translation (step 334). ).

  When a source language translation of a word having a co-occurrence relationship is searched from the word dictionary 11, the translation word judging means 4 extracts a word from which unnecessary words are removed from the result of the morphological analysis of the searched source translation word (step 335). ).

  When the translated word determining unit 4 extracts a word excluding unnecessary words, the translated word determining unit 4 searches the extracted word from the core word dictionary 8 (step 336).

  When retrieved from the core word dictionary 8, the field relevance of the core word is extracted by the translated word determination means 4 (step 337), and the number of appearances is weighted for the extracted field relevance according to the above equation (5). (Multiplication) is performed (step 338), the number of source language translations is weighted (division) (step 339), the field judgment degree of the judgment target translation is calculated, and a list of field judgment degrees of the judgment target translation is output. (Step 340). The reason why the original word translation is subjected to morphological analysis and unnecessary words are removed is to match the core word.

(A-2-7) Calculation processing of integrated field determination degree for each category of determination target word FIG. 8 is a flowchart for explaining the processing operation for calculating the integrated field determination degree for each category of the determination target word.

  The field judgment degree list of the judgment target original word and the field judgment degree list of the judgment target translation word are input to the parallel translation judgment means 5 (step 341), integrated by the parallel translation judgment means 5, and according to the above equation (6). The field target degree of the determination target word is calculated, and a list of field determination degrees of the determination target word is created (step 342).

  When the list of the field determination degrees of the determination target words is created, the bilingual determination unit 5 calculates the integrated field determination degree for each category of the determination target words from the list of the field determination degrees of the determination target words according to the above formula (7). Is calculated (step 343), and the calculated integrated field determination degree is stored in the category dictionary 12 (step 344).

  When the integrated field determination degree is stored in the category dictionary 12, the category to which the determination target word determined in the field belongs is output (step 345). The category to be output is the category with the highest integrated field determination.

(A-2-8) Example Next, field determination when the determination target word is “judgment” “Judgment” will be described.

  In the following, it is assumed that document storage processing and core word creation processing are performed in advance.

  When the determination target word “judgment” is input to the input unit 1, the source language determination unit 3 creates and outputs a list of field determination degrees of the determination target source word. Since the details of the process of creating the field determination degree list of the determination target original language have been described above, a description thereof will be omitted.

  FIG. 10 is a list example of the field determination degree of the determination target original language. As illustrated in FIG. 10, the list of field determination degrees of the determination target original word includes original words, categories, core words, and field determination degrees as item examples.

  The field determination degree will be described with reference to FIG. 10 using, for example, the word “election” having a co-occurrence relationship with the determination target original word “referee”.

  For example, when the word “election” having a co-occurrence relationship with respect to the determination target original word “referee” appears 195 times and the category “politics” core word “election field relevance is 0.00962, the determination target original word The field determination degree of the category “politics” core word “election” with respect to “the referee” is 0.00962 × 195 = 1.875 according to the above equation (4).

  Further, the translated word judging means 4 extracts the translated word co-occurring with the judgment target translated word “Judgment” from the translated word document. In addition, the word dictionary 11 is searched for a source language translation of the extracted words having a co-occurrence relationship. Further, a word excluding unnecessary words is extracted from the result of morphological analysis of the original word translation, and the extracted word is searched from the core word dictionary 8.

  FIG. 11 is an explanatory diagram for explaining a relationship between core words based on a word excluding a target translation word, a word having a co-occurrence relationship, an original language translation, and an unnecessary word.

  Further, the translated word determining means 4 weights the number of appearances to the field relevance of the searched core word, further weights the number of source word translations, calculates the field determination, and determines the field determination of the translation target word List is created and output.

  FIG. 12 is a list example of the field determination degrees of the determination target translated words. As illustrated in FIG. 12, the list of field determination degrees of the determination target translated words includes translation words, categories, core words, and field determination degrees as item examples.

  The field determination degree will be described using, for example, the word “selection” having a co-occurrence relationship with the determination target translation word “judgement”.

  For example, it is assumed that the word “selection” having a co-occurrence relationship with respect to the determination target translation word “judgement” appears four times, and the original word translation of “selection” is eight such as “election” and “voting”.

  In this case, the field determination degree of the category “politics” core word “election” for the determination target translated word “judgement” is 0.00962 × 4 ÷ 8 = 0.001 according to the above equation (5).

  Next, the bilingual determination means 5 determines the field determination degree of the determination target word as shown in FIG. 13 from the field determination degree list of the determination target translation word in FIG. 10 and the field determination degree list of the determination target translation word in FIG. Create a list of

  FIG. 13 is a list example of field determination degrees of determination target words. In addition, as shown in FIG. 13, the list of field determination degrees of determination target words includes language, translation, category, core word, and field determination degree as item examples.

  The field determination degree will be described with reference to FIG. 13, taking the category “politics” and the “election” for the determination target word “judgment” “judgment” as an example.

  In this case, as shown in FIG. 10, the field determination degree of the category “politics” and the core word “election” is 1.875 (= V1). Further, the maximum value of the field determination degree of the determination target original word “referee” is the case of the category “baseball” and the core word “baseball”, which is 2.581 (= M1).

  Further, as shown in FIG. 12, for the determination target translation word “judgement”, the category determination degree of the category “politics” and the core word “election” is 0.00481 (= V2). In addition, the maximum value of the field determination degree of the determination target translated word “judgement” is in the case of the category “politics” and the core word “National Diet”, which is 0.01076 (= M2).

  Therefore, the field determination degree of the category “politics” core word “election” for the determination target word “judgment” “judgment” is 2 × (1.875 ÷ 2.581) × (0.00481 ÷) according to the above equation (6). 0.01076) ÷ ((1.875 ÷ 2.581) + (0.00481 ÷ 0.01076)) = 0.553.

  Next, the parallel field determination means 5 calculates the integrated field determination degree for each category according to the above equation (7).

  The integrated field determination degree will be described with reference to FIG. 13 by taking the category “politics” for the determination target word “judgment” “judgment” as an example.

  For example, when the threshold value α is set to 0.02, the field determination degree of 0.02 or more is accumulated for each category. For example, the integrated field determination degree of the category “politics” with respect to the determination target word “judgment” is 0.553 + 0.170 + 0.123 + ... + 0.021 = 1.138

  As described above, the integrated field determination degree for each category is calculated and stored in the category dictionary 12. FIG. 14 shows an example of information stored in the category dictionary 12. As shown in FIG. 14, the category dictionary 12 has original words, translated words, categories, and integrated field determination degrees.

  When all the categories are calculated, the bilingual determination unit 5 selects and outputs the category “politics” having the maximum integrated field determination degree.

(A-3) Effects of the First Embodiment As described above, according to the present embodiment, the ambiguity and the original word included in each of the original word and the translated word by creating the core word in advance from the original word document classified in the field. The ambiguity between words and translations can be resolved, and the words and phrases in the translation can be classified. At that time, as long as there is a core word on the source language side, it is possible to cope with multiple languages simply by changing the word dictionary from the translation side to the source language side and the document on the translation side. At this time, the word dictionary need only have a translation of the word, so it does not need to be created specially, which contributes to saving of storage resources. Moreover, classification can be performed without depending on the contents and amount of the field-specific dictionary.

(B) Second Embodiment Next, a second embodiment of the bilingual phrase classification system, bilingual phrase classification method, and bilingual phrase classification program of the present invention will be described with reference to the drawings.

  The second embodiment also uses the definitions of terms described in the first embodiment.

(B-1) Configuration of Second Embodiment FIG. 15 is a block diagram showing a functional configuration of a bilingual phrase classification system according to this embodiment. Similarly to the first embodiment, the bilingual phrase classification system 200 of the second embodiment is realized by installing a bilingual phrase classification program on an information processing apparatus such as a personal computer provided with input / output means, for example. However, it can be functionally represented in FIG.

  In FIG. 15, the bilingual phrase classification system 200 of the first embodiment includes an input unit 1, a core word creation unit 2, a source language determination unit 3, a translation word determination unit 4, a translation translation determination unit 5, an output unit 6, and a source language classified document 7. , A core word dictionary 8, an original word unclassified document 9, a translated word unclassified document 10, a word dictionary 11, a category dictionary 12, and a category relation dictionary 13.

  In FIG. 15, the same and corresponding functional configurations described in FIG. 1 will be described with corresponding reference numerals for convenience of description.

  The bilingual phrase classification system 200 of the second embodiment is different from the functional configuration of the first embodiment in that it includes a category relation dictionary 13. Further, by providing the category relation dictionary 13, part of operations of the document storing process by the core word creating unit 2 and the core word creating process by the original word determining unit 3 are different. Accordingly, the following description focuses on functions different from those of the first embodiment.

  In the second embodiment, attention is paid to classification into hierarchical categories. That is, regarding a document set made up of a plurality of documents, assuming that the categories of the plurality of documents constituting the document set have a hierarchical structure, they are classified into the hierarchical categories.

  Here, the hierarchical category will be described with reference to FIG. FIG. 16 is an image diagram for explaining hierarchical categories assumed in the present embodiment.

  As shown in FIG. 16, “TOP” is the uppermost category, and categories such as “Sports” and “Computer” exist in the lower layer. Furthermore, categories such as “baseball” and “soccer” exist in the hierarchy below “sports”, and categories such as “OS” and “programming” exist in the hierarchy below “computer”. Furthermore, categories such as “OS1” and “OS2” exist in a hierarchy below “OS”.

  Between these categories, there is a parent-child relationship according to the inclusion and inclusion relationships between categories, and categories such as “sports” and “computer” are children of the “TOP” category. Similarly, categories such as “baseball” and “soccer” are children of the “sports” category, and categories such as “OS” and “programming” are children of the “computer” category. Furthermore, the categories such as “OS1” and “OS2” are children of the “OS” category.

  Among these, since the “TOP” category is the uppermost category, it has no parent, and on the contrary, the lowermost categories “baseball”, “soccer”, “programming”, “OS1”, etc. have no children. In FIG. 16, the category indicated by a black circle is the lowest layer category.

  The category relationship dictionary 13 is a dictionary that stores information on hierarchical relationships between categories. For example, when a person defines a hierarchical category configuration as shown in FIG. 16, the category relation dictionary 13 is created. The contents of the category relation dictionary 13 are composed of a category name that uniquely designates the category and information related to the parent category.

  FIG. 17 shows an example of the configuration and contents of the category relation dictionary 13 and is based on the hierarchical category of FIG. In FIG. 17, “−” is a null value (that is, indicates no value).

  The core word creation means 2 creates the core word and the field relevance using the source language classified document 7. The core word creating means 2 first creates a core word for the lowermost category of the source language classified document 7, and then creates a core word for the middle layer category using the core word of the lowermost category.

  As described above, the core word creation means 2 performs the core word creation processing of the lowermost category and the core word creation processing of the middle category by different processes for the following reasons.

  In order to classify into hierarchical categories, it is necessary to assign core words to all categories including intermediate layers other than the lowest layer. In the method of giving a document belonging to a category, there is no problem as a document belonging to a category in the case of a document corresponding to “Other” that does not include the contents of the child category, but in the case of a document corresponding to “General” including the contents of the child category, the parent and child Since the distinction between the two becomes ambiguous, it may not be correctly classified into the narrowest possible scope.

  However, preparing documents appropriately classified into all categories including intermediate levels requires more labor as the levels become deeper, so the core word of the parent category is not created from the documents.

  As the document, only a document classified in the unit of the lowest category is prepared. In the lowest category, a core word is created from the document, and the core word of the parent category is created from the core word of the child category immediately below.

(B-2) Operation | movement of 2nd Embodiment Next, operation | movement of the bilingual phrase classification system 200 of this embodiment is demonstrated.

  In order for the bilingual word / phrase classification system 200 of this embodiment to perform field determination, the source language document classified into the lowest category is stored in the source language classified document 7 for each category in advance, and the source language classified document is stored. 7 must be used to create the core code and field relevance.

  The core word of the lowest layer category is created by the core word creating means 2 from the document of the original language classified document 7, and the core word of the middle layer category is created by the core word creating means 2 using the category relation dictionary 13. Create from.

  Then, field judgment is performed using various databases and various dictionaries.

  Hereinafter, operations of the document storage process and the core word creation process will be described. Note that the overall operation of the bilingual classification system 200 and the operation of the bilingual field determination process correspond to the operation described in the first embodiment, and thus detailed description thereof is omitted.

(B-2-1) Document Storage Processing FIG. 18 is a flowchart for explaining the operation of document storage processing.

  In FIG. 2, when the document storage process is selected (steps 201 and 202), the selection of the process by the user is accepted as shown in FIG. 18 (step 601).

  Whether the input means 1 captures the processing selection by the user, and according to the selection, designates the lowest category and inputs the source language document classified by the lowest category and stores it in the source language classified document 7 by category (Step 602), whether an uncategorized source language document is input and stored in the source language uncategorized document 9 (step 603), or whether an unclassified translated word document is input and stored in the translated word uncategorized document 10 (step 604) Or end (END).

(B-2-2) Core Word Creation Process FIG. 19 is a flowchart for explaining the operation of the core word creation process.

  In FIG. 2, when the core word creation process is selected (steps 201 and 203), the process of FIG. 19 is executed.

  In FIG. 19, the core word creating means 2 creates the core word of the lowest category using the documents of the original language classified documents stored for each lower category in the document storing process (701). Details of the process of creating the core word of the lowest category will be described later with reference to FIG.

  When the core word of the lowest category is created, the core word creation means 2 creates the core word of the middle category using the created core word of the lowest category (702). Details of the process of creating the core word of the middle layer category will be described later with reference to FIG.

(B-2-2-1) Core Word Creation Processing of Lowermost Category Category FIG. 20 is a flowchart for explaining the operation of core word creation processing of the lowest category.

  In FIG. 20, the core word creating means 2 takes out the lowest category among the categories stored in the category relation dictionary 13 (step 711).

  The lowermost category is a category having no child category. For example, in the example of the category dictionary 13 illustrated in FIG. 17, the category name is not listed as a parent category item. For example, in FIG. 17, “baseball” is not listed as a parent category, and is therefore the lowest category.

  When the lowest category is extracted, it is checked whether the core word creation process described below is performed for all the extracted categories, and the core word is processed until the core word creation process is completed for all categories. Creation processing is performed (step 712).

  In step 711, when the lowermost category is extracted from the category relation dictionary 13, the core word creating means 2 extracts documents classified into the extracted category from the source language classified document 7 for each category (step). 713).

  When the document in the lowermost category is extracted from the source language classified document 7, the extracted word is subjected to morphological analysis by the core word creating means 2 (step 714).

  When the core word creation unit 2 performs morphological analysis on the lowermost category document, the core word creation unit 2 extracts a word obtained by removing unnecessary words from the morpheme analysis result as a core word (step 715).

  Then, when the core word is extracted for the lowest category, the core word creation unit 2 calculates the field relevance according to the above equation (3), as described in the first embodiment ( Step 716).

  Thus, when the core word of the lowest category is created by the core word creating means 2 and the field relevance of the created core word is calculated, the core word and the field relevance are stored in the core word dictionary 8. (Step 717).

  As described above, the core word creation process of the lowest category is performed for all the lowest category extracted in S711.

(B-2-2-2) Intermediate Word Category Core Word Creation Processing FIG. 21 is a flowchart for explaining the operation of the intermediate layer category core word creation processing.

  In FIG. 20, the core word creating means 2 takes out the uppermost category from the categories stored in the category dictionary 13 (step 721).

  The uppermost category is, for example, one that does not have a parent category (has a parent category of “−”) in the example of the category relation dictionary 13 shown in FIG.

  Next, when the uppermost category is extracted from the category dictionary 13, the core word creating means 2 searches the category dictionary 13 for a category having the uppermost category as a parent category (step 722).

  It is confirmed whether or not the core word creation processing described below is performed for all categories searched in S722, and the core word creation processing is performed until the core word creation processing is completed for all categories ( Step 723).

  If no core word has been created for the category retrieved from the category dictionary 13 in S723, the category having the category for which the core word has not been created as a parent category is retrieved from the category dictionary 13 by the core word creating means 2 ( Step 724).

  For the categories retrieved from the category dictionary 13 in S724, it is confirmed whether or not a core word has been created for all child categories having the same category (parent category) as a parent (step 725).

  In S725, when the core word creation means 2 confirms that core words have been created for all child categories, the process proceeds to S726, while the core word creation means 2 confirms that core words have not been created for all child categories. If so, the process returns to S724 and is repeated (step 725).

  As a result, the parent category has all the core words created for all the child categories.

  In S726, a core word is created by the core word creating means 2, and the field relevance of the created core word is calculated (step 727).

  Here, the method of creating the core word of the parent category using the core word of the lowest category in the core word creating means 2 is to judge from the degree of deviation of the core words of the child category.

  If it exists in the child category evenly, it will be the core word of the parent category (create it in the parent category), if it protrudes in any child category, it will not be the core word of the parent category (exist only in the child category) This is the idea.

  The procedure is as follows.

  The average value (mean) and standard deviation (sd) of the field relevance level are calculated for each core word in the entire child category immediately under a certain parent category, assuming a normal distribution.

  The field relevance of the category having no core word is calculated as 0. If all the field relevance levels are within the range of variation, the core word is set as the core word of the parent category.

  In the example, the range is “average value + standard deviation × 3” (mean + 3sd). The probability that data falls within the range of “average value + standard deviation × 3” (mean + 3sd) is 99.73%.

  The calculation of the range is correctly the average value ± plus or minus, but since there is no negative value in the field relevance, the negative one can be ignored.

  If tf * idf is used as it is, the field relevance of the lowest category depends on the amount of core word creation documents. Therefore, the value is divided by the total number of core words for each category. Since there is no created document, the field relevance level of the parent category must be set appropriately.

  Here, the upper limit value of the range is “average value + standard deviation × 3” (mean + 3sd). This is because the field relevance of the core word of the parent category is higher than the field relevance of any child category.

  When the core word created in S726 and its field relevance are calculated, the core word creation means 2 stores the core word and its field relevance in the core word dictionary 8 (step 727).

  Then, the process returns to S722, and the operation is repeated until the core words of all categories having the uppermost category as a parent category have been created.

(B-2-3) Example of Core Word Creation Process Next, an example of the core word creation process when the category relation dictionary 13 has the dictionary information shown in FIG. 17 in the hierarchical category of FIG. 16 will be described.

  In FIG. 2, when the core word creation process is selected, first the core word creation process of the lowest category is performed (step 701).

  Here, for example, the core words of the lowest category such as “baseball”, “soccer”, “programming”, “OS1”, “OS2” in FIG. 17 are created, and the field relevance of the created core words is calculated. Is done. Then, the calculated core word and the field relevance are stored in the core word dictionary 8.

  Next, a core word creation process of the category of the intermediate layer is performed (step 702).

The core word creating means 2 extracts the top-level category “T0P” (step 721), and searches the category dictionary 13 for the categories “sports” and “computer” having “T0P” as a parent category (step 721). Step 722).

  When the core word creation means 2 determines that no core word has been created for all of “sports” and “computer”, a category (child category) having “sports” and “computer” as parent categories is searched. (Step 724).

  At this time, for “sports”, “baseball”, “soccer” and the like are searched as child categories, and for “computer”, “OS”, “programming” and the like are searched as child categories.

  Here, since “baseball”, “soccer”, “programming”, and the like are the lowest category, a core word is created in S701. On the other hand, since “OS” or the like is not the lowest layer, no core word is created.

  In step 725, the core word creating means 2 confirms whether or not a core word has been created for each category (child category) searched in step 724, and repeats until all child categories have created a core word.

  That is, for “sports”, since core words are created for all of the child categories “baseball” and “soccer”, the processing of the core word creation means 2 proceeds to step 726 and subsequent steps, and the core word for “sports” is It is created, the field relevance of the core word is given, and stored in the core word dictionary 8.

  On the other hand, for “computer”, since the core word of the child category “OS” has not been created, the process returns to step 724, and the category having “OS” as the parent category is searched from the category relation dictionary 13 (step). 724). That is, “OS1”, “OS2”, etc., which are child categories of “OS”, are searched.

  Since “OS1”, “OS2”, and the like are the lowest-level categories, a core word is created in step 701.

  Accordingly, the processing of the core word creating means 2 for “OS” proceeds to step 726 and subsequent steps, a core word for “OS” is created, the field relevance of the core word is given, and stored in the core word dictionary 8.

  Subsequently, the process returns to S722, and it is determined by the core word creating means 2 that a core word has not been created for “computer” which is a child category of “TOP”. Then, the core word creating means 2 searches for “OS”, “programming”, etc. with “computer” as a parent category (steps 722 to 724).

  In this case, since the core word is created for the “OS” by the above processing, the core word is created in all of the child categories “OS”, “programming”, etc. of the “computer”. The process proceeds to step 726 and subsequent steps, a core word for “computer” is created, the field relevance of the core word is given, and stored in the core word dictionary 8 (steps 725 to 727).

(B-2-4) Middle-tier category field relevance assignment processing FIG. 22 is a flowchart for explaining the operation of the middle-tier category field relevance assignment processing.

  In FIG. 22, the parent word category name is extracted by the core word creation means 2 (step 731). The category name of the child category immediately below the parent category name is extracted (step 732).

  Then, the core word for the immediate child category is extracted from the core word dictionary 8 (step 733), and if the upper limit value has been calculated for all the core words (step 734), the process is ended.

  If the upper limit value has not been calculated for all core words in step 734, the field relevance of the core word is retrieved from the core word dictionary 8 (step 735).

  Then, the maximum value is obtained for the field relevance (step 736), the average value and the standard deviation are calculated (step 737), the upper limit value of the range is calculated (step 738), and the maximum value is calculated from the upper limit value. It is determined whether or not it is smaller (step 739).

  If the maximum value is smaller than the upper limit value, it is within the range, so it is set as the core word of the parent category (step 740).

  For example, when the average value of the field relevance of the core word “player” in all the child categories of the category “sports” is 0.001388, the standard deviation is 0.000516, and the maximum value is 0.00257, the upper limit value of the range is Since the maximum value is within the range, the core word “player” becomes the core word of the category “sports”. The field relevance of the core word “player” in the category “sports” is 0.00294.

  For example, when the average value of the field relevance of the core word “game” in all child categories of the category “sports” is 0.001985, the standard deviation is 0.001821, and the maximum value is 0.00497, the upper limit value of the range Becomes 0.00745, and the maximum value is within the range, so the core word “game” becomes the core word of the category “sports”. The field relevance of the core word “game” of the category “sports” is 0.00745.

  In addition, when the average value of the field relevance of the core word “pitcher” in all child categories of the category “sports” is 0.000783, the standard deviation is 0.002565, and the maximum value is 0.00929, the upper limit value of the range is Since it becomes 0.00848 and the maximum value exceeds the range, the core word “pitcher” does not become the core word of the category “sports”. FIG. 23 shows an example of the core word dictionary 8 created by the above processing.

(B-3) Effects of the Second Embodiment As described above, according to the second embodiment, the same effects as those of the first embodiment can be obtained.

  Further, according to the second embodiment, by generating core words in advance for each category of the hierarchical structure, words can be classified at a time regardless of the depth of the hierarchy. At this time, it is not necessary to prepare documents belonging to all categories, which contributes to saving of storage resources.

(C) Other Embodiments (C-1) The systems according to the first and second embodiments described above register words, compound words, phrases, bilingual phrases, and the like in a dictionary for natural language translation. It can be applied to a dictionary registration system. As a result, it is possible to automatically classify bilingual phrases and register them in the dictionary without the need for human judgment.

(C-2) In the first and second embodiments described above, the determination target word is stored in the determined category and the result is output to the user. However, the result may not be output to the user, and the category The data may be output before being stored in the dictionary so that the user can select a category.

  It is also possible to determine the field of words that are not parallel translations using the procedure up to the creation of the list of field determination degrees. Further, a list of field judgment degrees of words that are not parallel translations may be stored and used when the field of bilingual words is subject to field judgment.

(C-3) Also, the definitions of various terms explained at the beginning of the first embodiment can be changed according to the system operation and the like.

  For example, the creation of a core word or an unnecessary word may be defined by another method such as changing the type of part of speech or cutting out with an n-gram, or may be added or deleted. Furthermore, the co-occurrence relationship may be defined by another method such as a modification relationship, or the range to be extracted may be widened or narrowed. Also, the field relevance and its weight may be calculated by other methods, such as increasing the word included in the heading or reflecting the distance between words, and can be adjusted by specifying the word It may be. The calculation of the field relevance of the middle layer may change the definition of the range and the upper limit value of the range. Each document database may be referred to via a network.

(C-4) The components (1 to 13) described in the first and second embodiments described above may not be mounted on one device, but may be distributed and arranged on a plurality of devices. Good. For example, the source language classified document 7, the source language unclassified document 9, the mistranslation unclassified document 10, and the like may be replaced as a website accessible via a network.

(C-5) Further, the classification target may be not only a word but also a compound word or a phrase. Furthermore, the configurations of various dictionaries and the configurations of the lists described in the first and second embodiments described above are not limited to those illustrated.

(C-6) Most of the functions realized in hardware in the above description can be realized in software, and almost all realized in software can be realized in hardware.

It is a functional block diagram of the bilingual phrase classification system of 1st Embodiment. It is a flowchart which shows the whole operation | movement in 1st Embodiment. It is a flowchart which shows the document storage process in 1st Embodiment. It is a flowchart which shows the core word creation process in 1st Embodiment. It is a flowchart which shows the bilingual field determination process in 1st Embodiment. It is a flowchart which shows the calculation process of the field determination degree of the determination target original word in 1st Embodiment. It is a flowchart which shows the calculation process of the field determination degree of the determination target translated word in 1st Embodiment. It is a flowchart which shows the calculation process of the integrated field determination degree of the determination object word in 1st Embodiment. It is a figure which shows the structural example of the word dictionary in 1st Embodiment. It is a figure which shows the structural example of the field | area determination degree list | wrist of the determination target original language of 1st Embodiment. It is a figure which shows the relationship with the judgment object translation of 1st Embodiment, the word in co-occurrence, an original word translation, and a core word. It is a figure which shows the structural example of the field | area determination degree list | wrist of the determination target translated word of 1st Embodiment. It is a figure which shows the structure of the field | area determination degree list | wrist of the determination target word of 1st Embodiment. It is a figure which shows the structural example of the category dictionary of 1st Embodiment. It is a function block diagram of the bilingual phrase classification system of 2nd Embodiment. It is explanatory drawing explaining the hierarchy category in 2nd Embodiment. It is a figure which shows the structural example of the category relation dictionary of 2nd Embodiment. It is a flowchart which shows the document storage process in 2nd Embodiment. It is a flowchart which shows the core word creation process in 2nd Embodiment. It is a flowchart which shows the core word creation process of the lowest layer in 2nd Embodiment. It is a flowchart which shows the core word creation process of the intermediate | middle layer in 2nd Embodiment. It is a flowchart which shows the field related degree provision process of the category of the intermediate | middle layer in 2nd Embodiment. It is a figure which shows the structural example of the core word dictionary in 2nd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Input means, 2 ... Core word preparation means, 3 ... Original word determination means, 4 ... Translation word determination means, 5 ... Parallel translation determination means, 6 ... Output means, 7 ... Original word classified document, 8 ... Core word dictionary, 9 ... Original word not yet Classification document, 10 ... untranslated word document, 11 ... word dictionary, 12 ... category dictionary, 13 ... category relation dictionary.

Claims (6)

In the bilingual phrase classification system that determines each category of the input original word of the determination target word and its translation, and classifies the determination target word by category,
Category representative word creating means for creating one or a plurality of category representative words that characteristically express a certain category and creating a field relevance level indicating the characteristic degree of each category representative word in a document classified by category; ,
Category representative word holding means for holding each category representative word created by the category representative word creating means and its field relevance;
One or more words co-occurring with the original word to be judged are extracted from the original word uncategorized document, and among the extracted words having the co-occurrence relation, the field relations of one or more words that are the category representative words are extracted. Source word determining means for taking out the degree from the category representative word holding means and creating a field determination degree list of the determination target original words;
One or more words co-occurring with the target translation are extracted from the unclassified translation document, and one or more words that are the source language translations of the extracted co-occurrence words are determined. A word determination means for taking out the field relevance level of one or more words that are the category representative words from the category representative word holding means, and creating a field determination degree list of the determination target translation words,
The field judgment degrees of the category representative words in the field judgment degree list of the judgment subject original word and the field judgment degree list of the judgment subject translation word are integrated, and an integrated field judgment degree for each judgment subject word is created. Bilingual judgment means;
A bilingual phrase classification system comprising: output means for outputting category classification information of the determination target word based on a category-specific integrated field determination degree created by the parallel determination section.   The translated word determining means weights the number of occurrences of the word that becomes the source word translation in the translated word uncategorized document with respect to the field relevance level of each word that becomes the source word translation of each word having the extracted co-occurrence relationship. 2. The bilingual phrase classification system according to claim 1, wherein a field judgment degree list of the judgment target translation words is created by weighting the number of words to be translated.   The bilingual determination means integrates the field determination degrees of the upper category representative words so as to highly evaluate the field determination degrees of both the determination target original word and the determination target translation words. Item 3. The bilingual phrase classification system according to item 1 or 2. The above category representative word creation means is:
For a hierarchical category having a predetermined hierarchical structure, a category relationship dictionary unit that holds hierarchical relationship information between each category,
A lowermost category representative word determining unit that determines one or more category representative words that characteristically express the contents of a document classified into any one of the lowermost categories of the hierarchical category;
Referring to the hierarchical relationship information between each category in the category relationship dictionary section, the category representative word of the category corresponding to the parent of the child category including the lowest category is used as the category representative word of the child category. 4. The bilingual phrase classification system according to claim 1, further comprising: an upper category representative word determining unit that sequentially and repeatedly determines and executes an upper category between hierarchical categories. In the bilingual phrase classification method for determining the category of the input original word of the determination target word and its translation, and classifying the determination target word by category,
The category representative word creation means creates one or a plurality of category representative words that characteristically represent a certain category, and creates a field relevance level indicating the characteristic degree of each category representative word in a document classified by category. And
The category representative holding means holds each category representative word created by the category representative word creating means and its field relevance,
The original language determination means extracts one or more words having a co-occurrence relationship with the determination target original word from the original language uncategorized document, and one or more of the extracted word having the co-occurrence relationship is the category representative word. Taking out the field relevance level of the word from the category representative word holding means, creating a field determination degree list of the determination target original word,
The translation determination unit extracts one or more words co-occurring with the target translation from the untranslated word classification document, and determines one or more words that are the original translations of the extracted co-occurrence words The field relevance of one or more words that are the category representative words out of the words that become the original word translation is extracted from the category representative word holding means, and a field judgment degree list of the judgment target translation words is created.
The bilingual determination unit integrates the field determination degrees of the category representative words in the field determination degree list of the determination target original word and the field determination degree list of the determination target translation word, and integrates the determination target words by category. Create a degree of judgment,
The bilingual phrase classification method, wherein the output means outputs the category classification information of the determination target word based on the category-specific integrated field determination degree created by the parallel translation determination means. In the bilingual phrase classification device that determines each category of the input original word of the determination target word and its translation, and classifies the determination target word by category,
Category representative word creating means for creating one or a plurality of category representative words that characteristically express a certain category and creating a field relevance level indicating the characteristic degree of each category representative word in a document classified by category; ,
Category representative word holding means for holding each category representative word created by the category representative word creating means and its field relevance;
One or more words co-occurring with the original word to be judged are extracted from the original word uncategorized document, and among the extracted words having the co-occurrence relation, the field relations of one or more words that are the category representative words are extracted. Source word determining means for taking out the degree from the category representative word holding means and creating a field determination degree list of the determination target original words;
One or more words co-occurring with the target translation are extracted from the unclassified translation document, and one or more words that are the source language translations of the extracted co-occurrence words are determined. A word determination means for taking out the field relevance level of one or more words that are the category representative words from the category representative word holding means, and creating a field determination degree list of the determination target translation words,
The field judgment degrees of the category representative words in the field judgment degree list of the judgment subject original word and the field judgment degree list of the judgment subject translation word are integrated, and an integrated field judgment degree for each judgment subject word is created. Bilingual judgment means;
A bilingual phrase classification program, characterized in that, based on a category-specific integrated field determination degree created by the bilingual determination means, output means for outputting category classification information of the determination target word is realized.

Images