JP2007018359A - Language processing device, language processing method and language processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a language processing device that can determine the semantic well-formedness of expressions extracted from English text, and output well-formed expressions along with corresponding translations and output nothing about ill-formed expressions. <P>SOLUTION: A noun phrase A is extracted from first language text, the noun phrase A is machine-translated, and nouns MTN<SB>i</SB>(i=1-n) included in the machine translation result are extracted. A noun phrase B is extracted from second language text, and it is determined whether each noun MTN<SB>i</SB>is similar to the noun phrase B. The ratio (relevance rate) of the number of nouns MTN<SB>i</SB>similar to the noun phrase B to the number of nouns MTN<SB>i</SB>is next calculated, and it is determined whether or not the ratio (relevance rate) exceeds a threshold. If the threshold is exceeded, the noun phrase A and noun phrase B are registered as translations. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a language processing apparatus, a language processing method, and a language processing program for extracting an English specific expression having a prepositional phrase and a coordinate structure and a corresponding Japanese expression from a parallel corpus. In particular, a language processing apparatus and language processing method for extracting a parallel translation of a noun phrase including a collocation and a compound word, and a language processing program.

The processing target of the method of automatically acquiring the expression to be registered in the bilingual dictionary used in the machine translation system or the like from the bilingual corpus can be divided into a specific expression and a non-specific expression. When comparing a specific expression and a non-specific expression, there are relatively many specific expressions that are not registered in an existing dictionary. The large number of unregistered expressions in the dictionary is one of the major causes of quality degradation in machine translation systems. In order to solve this problem, the proper expression and its translation should be preferentially acquired.
For this reason, the inventors have been researching a method for acquiring from the bilingual corpus a pair of English specific expressions to be registered in the bilingual dictionary of the English-Japanese machine translation system and the corresponding Japanese expressions.
Conventionally, the research aiming to acquire specific expressions and their translations is not so much compared to research aiming to recognize specific expressions in a single language, but there are non-patent documents 1 to 4 etc. . In these documents, the English specific expressions to be extracted are limited to simple noun phrases with only prefix modifiers, and no discussion of noun phrases with prefix phrases and noun phrases with coordinate structure is not made. . However, there are actually many unique expressions having one or both of post-modification by a prepositional phrase and a coordinate structure.

Therefore, in the present invention, an English proper noun phrase having one or both of a prepositional phrase and an equal structure is selected. In the following, such noun phrases are simply referred to as English noun phrases. When processing English noun phrases, it is necessary to eliminate syntactic ambiguity such as preposition addition as a problem that did not occur when processing simple noun phrases with only prefix modification. For example, the expression “Japanese Embassy in Moscow” extracted from the following English sentence (E1) is semantically qualified and one English noun phrase, but “the United States into World War” extracted from the English sentence (E2). II ”is semantically ineligible and not an English noun phrase.
(E1) The ministry quickly instructed the Japanese Embassy in Moscow to. . . .
(E2) The attack on Pearl Harbor. . . . .drew the United States into World War II.
Therefore, it is necessary to judge the semantic eligibility of expressions extracted from English sentences, and to output appropriate expressions together with their translations, and to output nothing for unqualified expressions.

Patents relating to the present invention include the following. Patent Document 1 translates and generates a second language expression from a first language expression, extracts a parallel translation correspondence between the first language expression and the second language expression, and transmits parallel data. This is a translation dictionary automatic generation method in which this parallel translation data is stored in a parallel translation dictionary.
Patent Document 2 is a dictionary data improvement device that uses bilingual documents describing the same content, extracts compound word candidates by a compound word candidate extraction unit, and registers them in a dictionary database for translation.
Patent Document 3 associates a first language sentence with a second language sentence that is a translation result thereof, and an appearance frequency of a combination (word pair) of a word included in the first language sentence and a word included in the second language sentence. This is a device that automatically obtains a translation by determining that a word pair in which the appearance frequency of the word pair is higher than a predetermined value is a translation of the first language and the second language.
Y. Al-Onaizan and K. Knight.Translating Named Entities Using Monolingual and Bilingual Resources.In Procs.of the 40th Annual Meeting of the ACL, pp.400-408,2002. F. Huang and S. Vogel. Improved Named Entity Translation and Bilingual Named Entity Extraction.In Procs.of the 4th IEEE International Conference on Multimodal Interfaces (ICMI), pp.253-258,2002. F.Huang, S.vogel and A.Waibel.Automatic Extraction of Named Entity Translingual Equivalence Based on Multi-feature Cost Minimization.In Procs.of the ACL Workshop on multilingual and Mixed-language Named Entity Recognition, pp.9-16, 2003. R.Moor.Learning Translations of Named-Entity Phrases from Parallel Corpora.In Procs. Of the 10th conference of the ECAL, pp.259-266,2003. JP-A 62-297972 JP 2001-142882 A Japanese Patent Laid-Open No. 5-298359

According to Patent Document 1, this parallel translation data can be automatically accumulated to generate a translation dictionary. However, the extraction condition is that the words constituting the Japanese noun phrase are mapped to the English noun phrase, all the words of the English noun phrase are generated from the bilingual dictionary, or there are no parallel rules or rules for combinations of multiple words. It is generated by the bilingual dictionary, and the correspondence that can be extracted is limited.
Patent Document 2 is an apparatus for improving dictionary data, but check data to be collated with translated words extracted from translated word candidate data includes a target language morpheme, a source language morpheme, a set of translated words, and the like. Translations that can be checked by these check data are limited.
Patent Document 3 is a device that determines a word in a first language and a translated word in the second language based on the appearance frequency of word pairs, but requires a huge amount of data processing.
Therefore, it is necessary to judge the semantic eligibility of expressions extracted from English sentences, and to output appropriate expressions together with their translations, and to output nothing for unqualified expressions. The present invention presents a solution to this problem.

The language processing apparatus of the present invention extracts a noun phrase A from a first language sentence, machine translates the noun phrase A, and extracts a noun MTN _i (i = 1 to n) included in the machine translation result, , Extract the noun phrase B from the second language sentence, determine whether each noun MTN _i is similar to the noun phrase B, and then the number of nouns MTN _i and the noun MTN similar to the noun phrase B A ratio of the number of _i (participation rate) is calculated, it is determined whether or not this ratio (participation rate) exceeds a threshold, and noun phrases A and B that exceed the threshold are registered as parallel translations.

  More specifically, the language processing apparatus of the present invention includes a first language noun phrase extracting unit that extracts a noun phrase of a first language sentence, a translation unit that translates the first language noun phrase into a second language, and the translation A first language noun extraction means for extracting a noun included in the result, a second language sentence noun phrase extraction means for extracting a noun phrase of a second language sentence, a noun extracted by the first language noun extraction means, Similarities between second language nouns or between noun phrases that determine similarity between nouns extracted by the first language noun extraction means and nouns or noun phrases extracted by the second language noun phrase extraction means A determination means, and an involvement rate determination means for obtaining a ratio (involvement rate) in which the similarity between the noun phrase of the second language sentence and the noun phrase of the second language sentence is equal to or greater than a similarity threshold among the nouns extracted by the first language noun extraction means Provided, and outputs a parallel translation of the first language and the second language.

In the language processing apparatus according to the present invention, the first language noun phrase extraction unit may include a first language noun phrase extraction unit, a first language article list, a first language preposition list, and a first language noun phrase determination pattern. It is desirable to provide.
In the language processing apparatus of the present invention, it is preferable that the second language noun phrase extraction unit includes a second language noun phrase extraction unit and a second language noun phrase determination pattern.
In the language processing apparatus of the present invention, it is preferable that the second language noun extraction unit includes a second language noun extraction unit and a second language noun determination pattern.
Further, in the language processing apparatus of the present invention, the similarity determination means between the second language nouns or noun phrases includes a similarity determination unit between second language nouns or noun phrases, a character similarity determination unit, It is desirable to include a phonological similarity determination unit and a comprehensive determination unit.
In the language processing apparatus of the present invention, it is preferable that the character similarity determination unit includes a comparison unit in units of characters and a comparison unit in consideration of character types.
The language processing apparatus of the present invention preferably further includes a morphological analysis unit, and extracts a word string from the noun phrase extracted by the second language noun phrase extraction unit.
In the language processing apparatus of the present invention, the participation rate determination means calculates a participation rate from a ratio of the number of nouns in the first language and the number of nouns included in the noun phrase in the second language, and the participation rate It is desirable to determine whether or not the threshold value exceeds a threshold value.
Moreover, it is desirable that the language processing apparatus of the present invention outputs the first language noun phrase and the second language noun phrase that exceed the participation rate threshold value in association with each other.

According to another aspect of the present invention, there is provided a dictionary creation device that creates a dictionary to be registered as a parallel translation by associating a first language noun phrase with a corresponding second language noun phrase by the language processing device as described above. is there.
Furthermore, the present invention relates to a language processing method from another viewpoint, the step of extracting a first language noun phrase from a first language sentence, the step of translating the first language noun phrase into a second language, and the translation result Extracting a contained noun; extracting a second language noun phrase from a second language sentence corresponding to the first language sentence; extracting a second language noun phrase including the noun; and the noun Determining a second language noun phrase including a predetermined number or more of second language noun phrases including as a corresponding phrase of the first language noun phrase. Here, “the step of extracting the first language noun phrase from the first language sentence, the step of translating the first language noun phrase into the second language, the step of extracting the noun included in the translation result”, “the above-mentioned Either the step of extracting the second language noun phrase from the second language sentence corresponding to the first language sentence or the step of extracting the second language noun phrase including the noun may be processed first.
The language processing program of the present invention causes a computer to execute the above steps to realize a language processing method.

The language processing apparatus of the present invention extracts a noun phrase A from the first language sentence, machine translates the noun phrase A, and extracts a noun MTNi (i = 1 to n) included in the machine translation result, Extract a noun phrase B from the second language sentence, determine whether each noun MTN _i is similar to the noun phrase B, and then the number of nouns MTNi and the number of nouns MTNi similar to the noun phrase B The ratio (participation rate) is calculated, it is determined whether the ratio (participation rate) exceeds the threshold, and the noun phrase A and noun phrase B exceeding the threshold are registered as parallel translations. It can be output with the bilingual translation, and nothing can be output for unqualified expressions.
The language processing apparatus according to the present invention includes a first language noun phrase extracting unit that extracts a noun phrase of a first language sentence, a translation unit that translates the first language noun phrase into a second language, and a noun included in the translation result. Second language noun extraction means for extracting, second language noun phrase extraction means for extracting a noun phrase of a second language sentence, nouns extracted by the first language noun extraction means, and the second language noun phrases Of the nouns extracted by the first language noun extraction means, the similarity determination means between the second language nouns or between the noun phrases for determining the similarity between the nouns or noun phrases extracted by the extraction means, Since it includes the participation rate determination means for obtaining a ratio (involvement rate) in which the similarity between the bilingual sentence and the noun phrase is equal to or higher than the similarity threshold, a correct translation of the noun phrase can be output.

Moreover, the language processing apparatus of the present invention includes the first language noun phrase extraction unit, the first language article list, the first language preposition list, and the first language noun phrase determination pattern in the first language noun phrase extraction unit. Can extract noun phrases from English with proper expressions.
In the language processing apparatus of the present invention, since the second language noun phrase extraction means includes a second language noun phrase extraction unit and a second language noun phrase determination pattern, the associated noun when the second language is Japanese, It is also possible to extract an adjunct representing a combination modification.
In the language processing apparatus according to the present invention, the similarity determination means between the second language nouns or noun phrases includes a similarity determination unit between the second language nouns or noun phrases, a character similarity determination unit, Since the similarity determination unit and the comprehensive determination unit are provided, when the first language is English and the second language is Japanese, it is possible to collate the English unique expression with the Japanese expression.

In the language processing apparatus according to the present invention, the character similarity determination unit includes a character-by-character comparison unit and a comparison unit that considers the character type. Can be distinguished and handled. The processing considering the character type can prevent characters other than kanji from being matched in units that are too short.
In addition, the language processing apparatus of the present invention further includes a morphological analysis unit and extracts a word string from the noun phrase extracted by the second language noun phrase extraction unit, so that a noun can be easily extracted.
In the language processing apparatus of the present invention, the participation rate determination unit calculates a ratio (participation rate) between the number of nouns in the first language and the number of nouns included in the noun phrase in the second language, and the ratio ( Since it is determined whether or not (participation rate) exceeds a threshold value, the relevance rate and the F value can be improved. Thereby, an association error can be suppressed.
In addition, since the dictionary creation device of the present invention outputs the first language noun phrase and the second language noun phrase that exceed the threshold of the participation rate in correspondence with each other, it is possible to obtain a dictionary in which eligible expressions are registered as parallel translations. it can.
Also, the dictionary creation device of the present invention is registered by the device of the present invention because the language processing device as described above associates the first language noun phrase with the corresponding second language noun phrase and registers it as a parallel translation. A more accurate translation is possible by using the dictionary.

  In general, English noun phrases do not always correspond to Japanese noun phrases, but it is considered that English noun phrases that are the subject of the present invention often correspond to Japanese noun phrases. For this reason, it is assumed that noun phrases correspond to each other. The present invention comprises three stages: (1) extraction of English noun phrases, (2) extraction of Japanese noun phrases, and (3) collation of English noun phrases and Japanese noun phrases. Although it is considered that the syntactic ambiguity can be resolved in the first stage or the third stage, in the present invention, it is performed by a collation process.

The present invention is an apparatus, method, and program for extracting an English specific expression including a prepositional phrase and a coordinate structure and a corresponding Japanese expression from a bilingual corpus. The present invention (1) evaluates a pair of an English specific expression and a Japanese expression from two viewpoints of semantic similarity and phonological similarity and integrates two kinds of similarities to obtain an overall similarity. And (2) processing to resolve syntactic ambiguity such as adding prepositions in English proper expressions.
The apparatus and method of the present invention were tested using the bilingual corpus of Yomiuri Shimbun and The Daily Yomiuri, and an experiment was performed comparing the performance of the present invention with the performance of the baseline without the above processing. From the experimental results, the F value of 0.678 obtained by the present invention exceeded the F value of 0.583 at the baseline, indicating the effectiveness of the present invention.
Hereinafter, the present invention will be described in detail with examples.

(Example 1)
FIG. 1 shows a block diagram of the language processing apparatus of the present invention. Reference numeral 1 denotes a bilingual sentence input means for inputting a bilingual sentence using a keyboard, a mouse, and a scanner. The input unit 1 may be a communication interface that acquires a document via a network. Reference numeral 2 denotes a result output means for displaying the processing result of the present invention by a display device such as a liquid crystal display device, a plasma display device or a CRT. The output means 2 may be configured to output information to a storage device or a network in addition to displaying with characters. Reference numeral 3 denotes an operation main memory such as a RAM, and 4 denotes a mass storage means such as a hard disk or an optical disk. The storage unit 4 stores a bilingual corpus in which a first language sentence is associated with a second language sentence corresponding to the first language sentence. The program, language processing program and calculation program of the present invention are stored.
The following 5 to 10 are parts constituting the language processing apparatus of the present invention, 5 is a noun phrase extracting means for extracting and extracting a noun phrase from the first language sentence, and 6 is a noun phrase from the second language sentence. Noun phrase extraction means for extracting and extracting, 7 for machine translation means for translating from the first language to the second language, and 8 for extracting and extracting nouns as words from the noun phrases extracted by the second language noun phrase extraction means 6 Second language noun extraction means, 9 is a similarity determination means between nouns or noun phrases in the second language, and 10 is an involvement rate determination for determining an involvement rate of the translation result of the first language noun phrase with respect to the second language noun phrase. Means.

A more detailed block diagram of FIG. 1 is shown in FIG.
The first language noun phrase extraction means 5 of FIG. 1 includes a first language noun phrase extraction unit 5a, a first language article list 5b, a first language preposition list 5c, and a first language noun phrase determination pattern 5d. The second language noun phrase extraction means 6 includes a second language noun phrase extraction unit 6a and a second language noun phrase determination pattern 6b. The second language noun extraction means 8 includes a second language noun extraction unit 8a and a second language noun determination pattern 8b. A second language morphological analyzer 11 is provided that extracts a second language noun from the noun phrase extracted by the second language noun phrase extractor 6a and stores it in the second language noun extractor 8a. Similarity determination means 9 between nouns or noun phrases in the second language includes a similarity determination unit 9a, a character similarity determination unit 9b, and a phonological similarity determination unit 9c. And a comprehensive determination unit 9d.

Here, the first language and the second language are not limited to foreign languages and may be dialects as long as they are interpreted in a broad sense and are for transmitting information. Alternatively, the language is not limited to the language in which humans speak or record, and may be a language for computer processing. The first language sentence and the second language sentence are not limited to a single sentence but include a plurality of sentences, words, and phrases. A phrase includes not only words but also collocations and compound words.
The language processing device of the present invention may be configured as a language processing device, a dictionary creation device, or a dictionary creation method alone, or provided as software installed in a general-purpose processing device such as a personal computer. By executing the processing, it may function as a language processing device, a dictionary creation device, or a dictionary creation method.
The program of the present invention may be stored in a rewritable storage means such as a flash memory or a hard disk, or a non-rewritable storage means such as a mask ROM, CD, or DVD.

The apparatus of the present invention executes processing as shown in the flowchart of FIG. In this flowchart, the first language sentence is assumed to be English, and the second language sentence is assumed to be Japanese.
First, in step s01, the bilingual sentence input unit 1 acquires the first language sentence and the second language sentence corresponding thereto from the storage unit 4 that stores the bilingual corpus. The first language sentence and the second language sentence are as follows. The acquisition of the first language sentence and the second language sentence is performed by using the first language sentence as a keyword from the database storing a large number of the first language sentence and the second language sentence in addition to the acquisition from the bilingual corpus. You can also search for and retrieve sentences.
First language sentence: They strongly request Iraq to remove the obstacles which prevent t he Diplomatic and Consular Missions in Kuwait from executing their f unctions.
Second language sentence: Both sides refrain from taking any action against Iraq that removes the obstacles to the activities of the diplomatic mission in Kuwait and infringe the status and non-aggression of the mission members. Urgently ask.

Step s02 is a process of listing noun phrase candidates in the first language by the first language noun phrase extracting unit 5a.
In the English noun phrase extraction process, the longest word string represented by the following expression (1) is extracted (in this case, parsing is not performed).
Base NP ((P / C) BaseNP) ⁺ ... (1)
Here, Base NP is assumed to be a simple noun phrase with only a pre-modification phrase, and is a word string consisting of a capital letter or a number. However, the beginning may be a constant / indefinite article beginning with a lowercase letter. P is a preposition and C is an or comma. ⁺ Represents one or more times.
In order to prevent the extraction of syntactically unqualified English noun phrases, several conditions are imposed on the word string extracted by the expression (1). The main conditions are as follows.
Condition 1: If a comma is included in the word string, and must exist after the comma.
Condition 2: The word at the end of the word string is not a generic noun.
Condition 3: When a word string appears at the beginning of a sentence, the part of speech of the first word is neither a preposition, a conjunction, nor an adverb.

Condition 1 can prevent “Moscow in April, The Yomiuri Shimbun” from being extracted as an English noun phrase from the following English sentence (E3), for example.
(E3) The Russian government has decided to sign. . . . in Moscow in April, The Yomiuri Shimbun learned on Wednesday.
Condition 2 can prevent, for example, the extraction of “IBM and the USs” in “IBM and the USs” big three auto manufacturers. However, such an error cannot be prevented if the last word of an English noun phrase is not a generic noun. For example, from “the Self-Defense Forces and US forces” to “the Self-Defense Forces and US” It will be extracted by mistake.
Condition 3 is set to extract “Japan and the EC” instead of “Hopefully Japan and the EC”, for example. This determination is made by comparing the word with a list of prepositions, conjunctions, and adverbs. Of the words registered in the adverb list, 15 words such as East, West, North, and South are English words that function as nouns or adjectives, such as “West Germany, Britain and Italy”. Since noun phrases were relatively common in training data, they were deleted from the adverb list.

The features of this step s02 are as follows.
(A) Do not use English dictionary lookup or syntax analysis.
(B) Do not use an English noun dictionary.
(C) An article / preposition list (first language article list 5b, first language preposition list 5c shown in FIG. 2) is used.
(D) A word beginning with a capital letter in the middle of a sentence is regarded as a noun (with a proper expression).
(E) Extract noun phrase candidates as patterns of combinations of prepositions, articles, nouns of proper expressions, and commas mentioned above. The pattern is as shown in the above formula (1). Here, the first language noun phrase determination pattern 5d in FIG. 2 is determined for a noun phrase that always includes either a preposition phrase or a parallel by “a nd”.
(F) Among the noun phrases that match the above pattern, if multiple noun phrases with overlapping ranges are listed as candidates, only the longest of them is selected as a candidate. In this embodiment, the longest word is extracted in this way, but as another embodiment, all words may be listed as in Japanese.

As a result, the English noun phrase extracted from the first language sentence is only “the Diplomatic and Consular Missions in Kuwait”. The longest noun phrase is extracted here, but if this English noun phrase is too long and the corresponding Japanese noun phrase is not extracted, the extracted longest noun phrase is decomposed and shorter. All noun phrases may be extracted.
It should be noted that an arrangement such as “the Diplomatic and Consular Missions in Kuwait from executing their functions” is a part of a noun phrase because the part of “executing their functions” does not start with a capital letter in the process of step s01 of this embodiment. Not considered.

  The extracted English noun phrase is translated by the experimental system of the present invention, and the translation result is analyzed by morpheme using “tea bowl” which is a morphological analysis method. "Chaya" is a Japanese morphological analysis software for natural language processing published by Nara Institute of Science and Technology. Hereinafter, the machine translation result of an English noun phrase is referred to as an MT translation of the English noun phrase.

The next step s03 is a process of listing up the second language noun phrase candidates by the second language noun phrase extraction unit 6a.
In the Japanese noun phrase extraction process, a noun phrase that does not include the adjunct modification is extracted as a parallel translation candidate corresponding to the English noun phrase. The Japanese text of the bilingual corpus is morphologically analyzed by the morphological analysis method software “tea bowl”, and the word string expressed by the following equation (2) is extracted based on the result.
(JNF) ^* JN (2)
Here, it is assumed that JN is a compound noun and F is an adjunct. ^* Represents 0 or more times.
Unlike the extraction of English noun phrases described in step s01 above, not only the longest word string that satisfies the condition but also its partial word string is extracted. However, compound nouns are not divided in principle. In addition to the longest Japanese noun phrase "Amendments to the Employees' Pension Insurance Law and the National Pension Law" from a certain Japanese sentence, the "Employees' Pension Insurance Law and the National Pension Law", "Amendments to the National Pension Law", "Employees'Pension" “Insurance Law”, “National Pension Law”, and “Revision Plan” are extracted.

The features of this step s03 are as follows.
(A) Use a Japanese morphological analyzer with a part-of-speech dictionary. (Second language morphological analyzer 11 in FIG. 3)
(B) Extract noun phrase candidates as patterns of part-of-speech combinations. The part of speech that is the target is a noun and an adjunct that represents the combination modification, and the pattern is as shown in the above formula (2). (Second Language Noun Phrase Determination Pattern 6b in FIG. 2) As an appendix representing the combination modification, for example, a noun phrase such as “... It is.
(C) Among the noun phrases that match the above pattern, when a plurality of noun phrases with overlapping ranges are listed as candidates, all of them are candidates. However, parts where only nouns are continuous are treated as a group, and only some of the nouns are not extracted.

As a result, Japanese noun phrase candidates extracted from the second language sentence are:
"both"
"Iraq"
"Kuwaiti diplomatic mission in Kuwait"
"Activity"
“Failure on”
"Up"
"Failure"
"The status and non-aggression of the mission members"
"The status of the mission members"
"Member of the mission"
"Position and non-aggression"
"Position"
"Non-aggression"
"Action"
It becomes.

Next, in step s04, the English noun phrase candidate extracted in step s02 is translated into Japanese by the machine translation means 7 from the first language to the second language. In the case of the first language sentence, the only English noun phrase candidate is “the Diplomatic and Consular Missions in Kuwait”, and its machine translation (MT translation) is
"Diplomatic and consular missions in Kuwait"
Suppose that The accuracy of machine translation at this stage is not important in the present invention, and any machine translation method and machine translation apparatus can be used. Alternatively, instead of a so-called machine translation device, a translation device called an example-driven translation device or an example sentence-driven translation device may be used.
If a correct translation can be obtained by machine translation in this step s04, the processing of the present invention described below becomes unnecessary. Thus, the processing time of the present invention can be shortened by eliminating the processing. Alternatively, the final processing of the present invention is processed so that it is not output as a parallel noun phrase. Alternatively, by performing the processing of the present invention, it is possible to create a pair of English proper expressions and Japanese expressions that are more desirable than the results obtained by machine translation, and it is possible to collect only those that are not registered in the dictionary. , The dictionary level can be higher. The present invention is characterized in that a second language that matches the first language is extracted from the translated result.

The machine translation result (MT translation) of the English noun phrase candidate thus obtained is subjected to Japanese morphological analysis by the Japanese morphological analysis unit 11 shown in FIG. 2 in step s04-1. The morphological analysis result of the MT translation of the above English noun phrase candidate is, for example, as follows.
" Kuwait / in / in / diplomatic / above / of / and / and / consular / of / mission "
Here, the slash mark represents a morpheme boundary, and an underlined morpheme indicates that the part of speech is a noun.
The order of steps s01 to s04-1 described above is not limited to this order, and may be changed as appropriate, or may be processed in parallel.

Next, in loop 1 from step s05 to step s14, all of the MT translation of the English noun phrase obtained in step s02 and the English noun phrase obtained in step s04 and the Japanese noun phrase candidate obtained in step s03 Try the combinations in order.
In the matching process in loop 1, (1) MT translation of Japanese noun phrases and Japanese noun phrases are evaluated from the two viewpoints of semantic similarity and phonological similarity, and the two types of similarity are integrated. The overall similarity is obtained. Furthermore, (2) the semantic eligibility of English noun phrases is judged (syntactic ambiguity is resolved), and correspondence with Japanese noun phrases is performed only when they are qualified.
Steps s05 to s06 included in the loop 1 will be described.
In step s05, one MT translation of the English noun phrase candidate obtained as described above and one Japanese noun phrase candidate obtained as described above are converted into the second language noun phrase in FIG. Alternatively, the similarity is determined by the similarity determination means 9 between noun phrases.

The similarity is determined by obtaining each score by evaluation of semantic similarity by the character similarity determination unit 9b in FIG. 2 and evaluation of phonological similarity by the phonological similarity determination unit 9c, and comprehensively determining the result. To do. Semantic similarity can be evaluated in character units or character types.
(1) Semantic similarity evaluation: Comparison in character units The method of collating two Japanese character strings and obtaining the similarity between them is divided into a method of performing collation in units of words and a method of performing in units of characters. It is done. The MT translation of English noun phrases and Japanese noun phrases are collated on a character basis. The target of collation is the MT translation of English noun phrases and the compound noun parts of Japanese noun phrases, and the attached word part is excluded.
The Jackard coefficient is used as a scale representing the MT translation of English noun phrases and the semantic similarity between Japanese noun phrases. (The Jackard coefficient is, for example, “Example Cluster Analysis” by Romesburg, translated by Hideo Nishida, translated by Shinji Sato, translated by Utsuda Otsukuru, 1992), that is, a set of characters that appear in the compound noun part of the MT translation of the English noun phrase ( Let X be the character's appearance order without allowing for the appearance of characters), Y be the set of characters that appear in the compound noun part of the Japanese noun phrase, and U be the set U of characters that appear in both X and Y Then, the semantic similarity Ssem for a pair of an MT translation of an English noun phrase and a Japanese noun phrase is obtained by the following equation (3).

  The translation of “The Headquarters of the Struggle against Consumption Tax Raise” extracted from an English sentence using a conventional experimental system yields the MT translation of “Headquarters of struggle against consumption tax rise”. In this MT translation, the compound noun parts are “consumption tax rise”, “struggle”, and “headquarters” (number of characters is 9 characters). On the other hand, the “promotion headquarters against consumption tax rate increase” extracted from the Japanese sentence corresponding to this English sentence is composed of compound nouns only (16 characters). At this time, the characters appearing in both the set of characters constituting the MT translation compound noun of the English noun phrase and the set of characters constituting the compound noun of the Japanese noun phrase are “erasure”, “expense”, “tax”. , “Up”, “book”, and “part”. Therefore, from the expression (3), Ssem = 6 / () as the semantic similarity between the MT translation of the English noun phrase “Headquarters for struggling against consumption tax rise” and the Japanese noun phrase “Headquarters for promoting movement against consumption tax rate” The value 9 + 16−6) = 0.316 is given. This value is interpreted as the semantic similarity between “The Headquarters of the Struggle against Consumption Tax Raise” and the “Promotion Headquarters against Consumption Tax Rate Increase”.

(2) Semantic similarity evaluation: Comparison in consideration of character types All character types (kanji, hiragana, katakana, etc.) are used in the above-mentioned (1) [Evaluation of semantic similarity: comparison in character units]. English letters, numbers, symbols, etc.) are treated equally. In such processing that does not take into account differences in character types, characters other than Kanji characters may be inappropriately matched. For example, the MT translation of “Kasumigaseki Station of the Teito Rapid Transit Authority” extracted from a certain English sentence will be “Kasumigaseki Station of the Imperial High-Speed Transit Authority”. The Japanese text corresponding to this English sentence is extracted from “Keidan Subway / Kasumigaseki Station” and “Attache Case”. In this case, the semantic similarity between “Kasumigaseki Station of the Imperial City High-Speed Transportation Company” and “Kasumigaseki Station of the Imperial City High-Speed Transportation Company” is more similar between “Kasumigaseki Station of the Imperial City High-speed Transportation Company” and “Attache Case”. Desirably higher than the degree.

  However, the latter semantic similarity is actually higher as follows. The correct common correspondence between the compound nouns of “Kasumigaseki Station of Teikoku High Speed Transportation Corps” and “Keidan Subway / Kasumigaseki Station” is “Ying” and “Dang”. The semantic similarity Ssem between them is 2 / (27 + 9−2) = 0.059. On the other hand, the common characters between the compound nouns of “Kasumigaseki Station of the Imperial Capital High-Speed Transportation Company” and “Attache Case” are “shi”, “-”, “su”, so the semantic similarity Ssem is 3 / (27 + 8−3) = 0.094. Since this similarity value is relatively low, it is possible to prevent this association from being output due to a restriction by a threshold value. However, it is not desirable that an incorrect association is given higher similarity than a correct association. .

  Observing these cases, it is clear that even a single character is meaningful, but other characters are not, so it is not appropriate to treat kanji and non-kanji equally. . As a countermeasure against such a problem, a method of weighting by character type has been shown so far. On the other hand, in the present invention, the problem is solved by changing the collation unit according to the character type during collation. Specifically, collation of kanji is performed in units of characters, and collation of non-kanji is performed in units of the longest character string of the same character type (however, word boundaries are not exceeded). In the case of non-kanji characters, a condition is set that the longest character string of the same character type must be matched. However, in calculating the similarity, the method for obtaining the Jackard coefficient is the same for both kanji characters and non-kanji characters. That is, X is the set of characters that appear in the compound noun part of the MT translation of the English noun phrase, Y is the set of characters that appear in the compound noun part of the Japanese noun phrase, and the longest string unit of the same character type from X and Y When the character sets from which non-kanji characters that do not match each other are deleted are X ′ and Y ′, and the character set U ′ appears in both X ′ and Y ′, the MT translation of the English noun phrase and the Japanese noun phrase The semantic similarity S′sem in consideration of the character type for the pair is obtained by the following equation (4).

  In the case of “Kasumigaseki Station of the Emperor Metropolitan High Speed Transportation Company” and “Attache Case”, the characters that make up “Kasumigaseki Station” are deleted from X. ”,“ Speed ”,“ Degree ”,“ Communication ”,“ Through ”,“ Ying ”,“ Group ”, and the characters that make up“ Attache Case ”are deleted from Y, so Y ′ is an empty set . For this reason, there is no character common to X ′ and Y ′, and the semantic similarity S′sem according to Expression (4) is zero. By performing processing in consideration of the character type in this way, it is possible to prevent characters other than kanji from being matched in units that are too short.

(3) Evaluation of phonological similarity Semantic similarity is effective when the constituent elements of an English noun phrase are registered in the dictionary of a machine translation system and the MT translation is obtained. If not, it will not work effectively. In particular, it is expected that the English noun phrases (components) targeted in the present invention are not registered in the dictionary. In such cases, transliteration is effective.
When processing a Yomiuri Shimbun used in the experiment of the present invention, an article about Japan related matters such as The Daily Yomiuri, and a corpus containing many parallel translations, the English noun phrase related to Japan is Japanese. There is a high possibility that many words written in Roman letters are included. Focusing on this, Romaji reading is collated as an evaluation of phonological similarity. Readings of unregistered words that appear in the MT translation of English noun phrases are obtained by collating with the Japanese syllabary table. Readings other than unregistered words in the MT translation of English noun phrases and readings of Japanese noun phrases can be obtained by the above morphological analysis software “tea bowl”.

The MT translation of the English noun phrase and the phonological similarity of the Japanese noun phrase are also measured by the Jackard coefficient as in the case of the semantic similarity. However, in the case of phonological similarity, collation is performed in units of words, not in units of characters. That is, the set of readings of the compound noun part of the MT translation of the English noun phrase is X, the set of readings of the compound noun part of the Japanese noun phrase is Y, and further, the set of readings appearing in both X and Y is U. The phonological similarity Spho is obtained by the same expression as Expression (3).
As mentioned in (2) “Evaluation of semantic similarity: Comparison in consideration of character types”, the Japanese noun phrase to be associated with “Kasumigaseki Station of the Teito Rapid Transit Authority” is “Keidan Subway / Kasumigaseki Station”. It is. However, since “Kasumigaseki” is not registered in the dictionary of the experimental system, this Japanese noun phrase and the “Kasumigaseki Station of the Teito Rapid Transit Authority” MT translation “Kasumigaseki Station of the Imperial Capital High Speed Transit Authority” are meaningful. The similarity is as low as 0.059. “Tate”, “Kosokudo”, “Kotsuku”, “Aidan”, “Kasumigaseki”, and “Station” are read as MT translations of English noun phrases by the “Chaya”, a collation with the Japanese syllabary and morpheme analysis software. It is done. In addition, “Aidan”, “Chikatsutsu”, “Kasumigaseki”, and “Eki” are obtained as readings of Japanese noun phrases by the morphological analysis software “tea bowl”. Therefore, the readings that appear in both MT translations of English noun phrases and Japanese noun phrases are “Aidan” and “Kasumigaseki”, “Kasumigaseki Station of the Imperial Capital High Speed Transit Authority” and “Keidan Subway / Kasumigaseki Station” The phonological similarity Spho between is 2 / (6 + 4-2) = 0.250.

(4) Integration of Semantic Similarity Evaluation and Phonological Similarity Evaluation In the present invention, when an unregistered word is included in the MT translation of an English noun phrase by the comprehensive determination unit 9d, the following formula ( The total similarity S between the MT translation of the English noun phrase and the Japanese noun phrase is obtained by combining the semantic similarity S'sem and the phonological similarity Spho based on the weighted sum calculation formula as in 5). If the MT translation of the English noun phrase contains no unregistered word, the semantic similarity S′sem is set as the total similarity S.

βは現在のところ0.5としている。
そして、総合類似度Ｓが類似度閾値ｔｈ以上である英語名詞句と日本語名詞句の対を出力する。現在のところ類似度閾値ｔｈは0.1としている。

β is currently set to 0.5.
Then, a pair of an English noun phrase and a Japanese noun phrase whose total similarity S is equal to or greater than the similarity threshold th is output. At present, the similarity threshold th is 0.1.

Here, the following set is considered as one of the combination of the MT translation of the English noun phrase candidate obtained as described above and the Japanese noun phrase candidate.
MT translation of English noun phrase candidates: diplomatic and consular missions in Kuwait Candidates for Japanese noun phrases: diplomatic consular delegations in Kuwait
MT translation of the English noun phrase candidate: Kuwait / to / definitive / diplomacy / above / Bruno /, / and / consul / mission Japanese noun phrase candidate: the meaning of the presence / Kuwait / diplomatic / consular / mission slash and underline As described in step s04, the slash mark represents a morpheme boundary, and the underlined morpheme is a noun. Further, the morphological structure of the Japanese noun phrase candidate is obtained at the stage of morphological analysis of the Japanese sentence before obtaining the Japanese noun phrase candidate in step S03.

The semantic similarity is obtained by looking at the commonality of characters in the present invention. More specifically, the noun part in the noun phrase, that is, the noun commonality shown in the underlined portion of the above example is used. That is, as described in (1) “Evaluation of semantic similarity: comparison in character units”.
In addition, since it is considered that the weight of one character is different between kanji and non-kanji (for example, katakana), in the present invention, even if one character is common, it is regarded as “common”. As described in (2) “Evaluation of semantic similarity: Comparison considering character types”, non-kanji characters are not considered “common” if they do not match in morphological analysis. Yes.

Therefore, when the semantic similarity is obtained by applying the above equation (4), the following is obtained.
X ... Kuwaiti Consular Mission (15 characters)
Y: Diplomatic mission in Kuwait (13 characters)
U '... Kuwaiti diplomatic consul (9 characters)
Semantic similarity = 9 / (15 + 13-9) = 0.4737
With respect to phonological similarity, the phonological similarity referred to in the present invention is obtained only when the MT translation of an English noun phrase includes a morpheme consisting only of English letters (ie, a word that was not translated by the MT translation). So we don't ask for this group. When phonological similarity is not obtained, the effect of phonological similarity is not seen in the comprehensive determination as described in the above formula (5).
An example in which it is effective to obtain phonological similarity has been described in (3) Evaluation of phonological similarity.

As a result, the overall similarity of this group is the same as the semantic similarity, which is 0.4737.
If the similarity threshold is set to 0.1, the similarity of this group exceeds the threshold, and the process advances from step s06 to step s07. If the similarity is less than or equal to the threshold value 0.1, the process proceeds to step s13. Step s13 returns to step s05 in order to try the next combination in step s14 when there is still a combination of noun phrase candidates. The loop in FIG. 3 returns to step s05, but the process may return to step s04 to translate the noun phrase candidate of the first language into the second language. If there is no combination of noun phrase candidates in step s13, the process proceeds to step s15, the bilingual noun phrase determined to be output is output, and the process ends.
In addition to the “Kuwaiti Diplomatic Consular Mission”, the Japanese noun phrase candidate obtained in step s03 is similar to the MT translation of the English noun phrase candidate “diplomatic and consular mission in Kuwait”. In view of the commonality of characters, it is clear that the value will be zero or extremely low (obviously less than 0.1 when the similarity threshold is 0.1), so detailed description will be omitted. In this case, the process proceeds from step s06 to step s13.

Next, in step s07, preparation is made for determining the participation rate of the MT translation of the English noun phrase with respect to the Japanese noun phrase candidate. That is, the noun is extracted from the MT translation of the first language noun phrase candidate.
MT translation of the above English noun phrase candidates " Kuwait / Ni / Koku / Diplomatic / Up ///// Consul / Mission"
To extract nouns. Here, the noun refers to a portion where only “nouns” extracted from the morphological analysis result by the second language noun extraction unit 8a by the second language noun determination pattern 8b of FIG. 2 are continuous. In other words, “Kuwait”
"Diplomatic"
"Consul"
"Mission"
Are extracted as nouns.

Next, in step s08, for each noun extracted from the MT translation of the first language noun phrase candidate, the degree of similarity with the second language noun phrase candidate is obtained.
Noun each extracted in step S07, the Japanese noun phrase candidate "presence / Kuwait / foreign / consul / mission"
The degree of similarity is obtained respectively. The calculation method is as described in step s05. The result is as follows when it is obtained from equation (4).
「Kuwait」 ・ ・ ・ ・ ・ 0.3846
“Diplomatic” ・ ・ ・ ・ ・ 0.1429
`` Consul '' ・ ・ ・ ・ ・ 0.1538
"Mission" ... 0
Next, in step s09, it is determined whether or not each similarity is equal to or greater than a similarity threshold.
Here, when the threshold is 0.1, “Kuwait”, “Diplomatic”, and “Consular” are equal to or higher than the similarity threshold, and “Mission” is lower than the similarity threshold.
Next, in step s10, the ratio (participation rate) of the nouns extracted from the MT translation of the noun phrase in the first language whose similarity with the noun phrase candidate in the second language is equal to or greater than the similarity threshold is obtained. In other words, among the four nouns extracted from the MT translation of the English noun phrase candidate "Diplomatic and consular mission in Kuwait", it is involved in the Japanese noun phrase candidate "Kuwait diplomatic mission in Kuwait" Since there are three, the participation rate is 0.75.
In the next step s11, the participation rate determination unit 10 determines whether the participation rate exceeds a threshold value. If the threshold of participation rate is 0.5, the above-mentioned participation rate 0.75 exceeds the threshold value.

Therefore, in step s12, it is determined that a set of noun phrase candidates that exceed the participation rate threshold is output as a parallel noun phrase. The following English noun phrase candidate, the MT translation of the English noun phrase candidate, and the Japanese noun phrase candidate described in the first embodiment are determined to be correct as a parallel translation of the noun phrase. Or it may be judged correct.
English noun phrase candidates: the Diplomatic and Consular Missions in Kuwait
MT translation of English noun phrase candidates: Diplomatic and consular mission in Kuwait Japanese noun phrase candidates: Kuwaiti diplomatic mission in Kuwait If the participation rate does not exceed the threshold in step s11, the noun phrase candidate The set is not output as a bilingual noun phrase, and the process proceeds to step s13. Further, after it is determined in step s12 that the candidate phrase set is output as a parallel noun phrase, the process proceeds to step s13.

(Example 2)
The second embodiment is an example for explaining that a misjudgment can be avoided if a participation rate judgment is adopted where there is a possibility that an incorrect noun phrase may be associated with a judgment based only on the degree of similarity.
In the second embodiment, a process of extracting a bilingual noun phrase from the following English sentence and Japanese sentence pair will be described.
English: The attack on Pearl Harbor was the trigger that drew the United States into World War II.
Japanese sentence: Pearl Harbor attack was also a turning point that triggered the US intervention in World War II.

In step s01, the English sentence is input, and in step s02, English noun phrase candidates are extracted. The English noun phrase candidate is “the United States into World War II”. “Attack on Pearl Harbor” is not extracted as an English noun phrase candidate because it does not begin with a capital letter.
In step s03, a Japanese noun phrase candidate is extracted from the Japanese sentence. In this example, the Japanese noun phrase candidates extracted from the Japanese sentence are as follows.
"Attack on Pearl Harbor"
"USA"
"Second World War"
"Trigger"
"Turning point"

Next, step s04 is a step of translating the English noun phrase candidate extracted in step s02 into Japanese by machine translation means. The machine translation result (MT translation) of the above English noun phrase candidate “the United States into World War II” is as follows.
"United States to World War II"
Then, it is assumed that the morphological analysis result of the MT translation of this English noun phrase candidate is as follows, for example.
" World War II / To / No / USA "
A slash symbol represents a morpheme boundary, and an underlined morpheme indicates that the part of speech is a noun.

The next steps s05 to s14 are loops that try all combinations of the English noun phrase candidate obtained in step s02 and the MT translation obtained in s04 and the Japanese noun phrase candidate obtained in step s03.
First, steps s05 and s06 will be described.
In step s05, the degree of similarity is determined between one MT translation of the English noun phrase candidate obtained up to now and one of the Japanese noun phrase candidates obtained up to the above. As one of the combinations of the MT translation of English noun phrase candidates obtained above and the Japanese noun phrase candidates obtained above,
MT translation of English noun phrase candidates: United States to World War II Japanese noun phrase candidates: Consider a group of World War II.

The result of this set of morphological analysis is
World War II / to / of / USA
World War II . By applying equation (4) to obtain the semantic similarity,
X ... World War II USA (9 characters)
Y ... World War II (7 characters)
U '... World War II (7 characters)
Semantic similarity = 7 / (9 + 7-7) = 0.7778
Phonological similarity is out of scope. Therefore, the overall similarity in this group is 0.7778. If the similarity threshold is 0.1, the similarity is higher than the threshold in this set, and the process advances from step s06 to step s07. If the similarity is less than or equal to the threshold value 0.1, the process proceeds to step s13.

Step s07 is a step of extracting a noun from the MT translation of the English noun phrase candidate. In this step s07, the nouns “second world war” and “US” are extracted from “ second world war / to / no / US ”.

Next, the participation rate is obtained in steps s08 to s11.
Step s08 is a step for obtaining the similarity between each noun extracted from the MT translation of the English noun phrase candidate and the Japanese noun phrase candidate. That is, for each noun extracted in step s07, the degree of similarity with the Japanese noun phrase candidate “second world war” is obtained. Result is,
World War II ... 1.0
United States 0.0
It becomes. Next, in step s09, it is determined whether or not each similarity is equal to or greater than a similarity threshold.
Here, if the similarity threshold is 0.1, “second world war” is equal to or greater than the similarity threshold, and “US” is less than the similarity threshold. Next, in step s10, the ratio (participation rate) of the nouns extracted from the MT translation of the noun phrase in the first language whose similarity with the noun phrase candidate in the second language is equal to or greater than the similarity threshold is obtained. From the above results, among the two nouns extracted from the MT translation of the English noun phrase candidate "US to World War II", it is involved in the Japanese noun phrase candidate "World War II" Since there is one, the participation rate is 0.5. If the participation rate threshold is 0.5, this set does not exceed the threshold.

In the next step s11, the participation rate determination unit 10 determines whether the participation rate exceeds a threshold value. If the participation rate threshold is 0.5, the above-mentioned participation rate is 0.5, so the threshold is not exceeded.
Therefore, since the threshold value of the participation rate is not exceeded in step s11,
English noun phrase candidates: the United States into World War II
MT translation of English noun phrase candidates: US to World War II Candidates for Japanese noun phrase: World War II candidates are judged to be incorrect translations of noun phrases. Or, there is a high possibility of being judged as an error. For this reason, this set is not output as a bilingual noun phrase, but proceeds from step s13 to step s14 and step s05, and another combination is tried.

As another combination of the MT translation of the English noun phrase candidate obtained above and the Japanese noun phrase candidate, the following combination is considered.
MT translation of English noun phrase candidates: USA to World War II Japanese noun phrase candidates: USA The morphological analysis of this group is as follows.
World War II / to / of / USA
For this morphological analysis, applying equation (4),
X ... World War II USA (9 characters)
Y ... United States (2 characters)
U '... United States (2 characters)
Semantic similarity = 2 / (9 + 2-1) = 0.2222
Phonological similarity is out of scope. Therefore, in this group, the overall similarity is 0.2222. If the similarity threshold is 0.1, the similarity is higher than the threshold in this set, and the process advances from step s06 to step s07. If the similarity is less than or equal to the threshold value 0.1, the process proceeds to step s13.

Step s07 is a step of extracting nouns from the MT translation of English noun phrase candidates. In step s07, nouns are extracted from “ second world war / he /// US ”. “Great War” and “USA” are extracted.
Next, the participation rate is obtained in steps s08 to s11. In step s08, for each noun extracted in step s07, the degree of similarity with the Japanese noun phrase candidate “US” is obtained. Result is,
World War II ... 0.0
United States ・ ・ ・ ・ ・ 1.0
It is. Next, in step s09, it is determined whether or not each similarity is equal to or higher than the similarity threshold. If the similarity threshold is 0.1, “US” is equal to or higher than the similarity threshold, and “WWII USA” Is less than the similarity threshold. Next, in step s10, a step of obtaining a ratio (participation rate) of the nouns extracted from the MT translation of the first language noun phrase whose similarity with the second language noun phrase candidate is equal to or greater than the similarity threshold. Because. Based on the above results, one of the two nouns extracted from the MT translation of the English noun phrase candidate “US to World War II” is involved in the Japanese noun phrase candidate “US”. Therefore, the participation rate is 0.5. If the participation rate threshold is 0.5, this set does not exceed the threshold.

The next step s11 is a step in which the participation rate determination unit 10 determines whether or not the participation rate exceeds a threshold value. If the participation rate threshold is 0.5, the above-mentioned participation rate is 0.5, so the threshold is not exceeded.
Therefore, since the threshold value of the participation rate is not exceeded in step s11,
English noun phrase candidates: the United States into World War II
MT translation of English noun phrase candidates: United States to World War II Japanese noun phrase candidates: The US group is judged to be an error as a translation of noun phrases. Or, there is a high possibility of being judged as an error. For this reason, this set is not output as a bilingual noun phrase, but proceeds from step s13 to step s14 and step s05, and another combination is tried.
In addition to “World War II” and “United States”, the Japanese noun phrase candidate obtained in step s03 is similar to the MT translation of the English noun phrase candidate “input to World War II”. From the viewpoint of character commonality, it is clear that the value is zero or extremely low (obviously less than 0.1 when the similarity threshold is 0.1).

Next, resolution of syntactic ambiguity will be described.
As described in step s02 above, in the second language noun phrase candidate extraction step, in the case of an English sentence, the English noun phrase extracted by the process of extracting the English noun phrase is originally composed of the constituent elements of the English noun phrase. It may contain prepositional phrases that do not qualify. An English noun phrase that includes such a prepositional phrase is semantically ineligible and should not be associated with any Japanese noun phrase. At the beginning of this specification, as mentioned in English (E2), “the United States into World War II” extracted from the following English (E4) is semantically ineligible.
(E4) The attack on Pearl Harbor was the trigger that drew the United States into World War II.
(J4) The attack on Pearl Harbor was also a turning point that triggered the US intervention in World War II.

  However, in the processing described in (1) “Comparison in character units” and (2) “Comparison in consideration of character types” in step s02, a certain English noun phrase includes a preposition phrase that is semantically ineligible. It is difficult to identify whether or not it is present, and in such an example, inappropriate association is performed. That is, the MT translation of “the United States into World War II” is “US to World War II” (the number of characters in the compound noun part is 9 characters), which is extracted from the above Japanese sentence (J4) Since the seven letters of “World War II” are shared with the noun phrase “World War II”, the semantic similarity Ssem is a relatively high value of 7 / (9 + 7−7) = 0.778. Thus, an inappropriate association is obtained. Therefore, a new process is introduced to deal with this point.

If an English noun phrase is semantically ineligible and does not originally form a noun phrase, the Japanese expression corresponding to this English noun phrase is not a single Japanese noun phrase, but another expression form. It is thought that it is easy to become. According to this hypothesis, whether or not an English noun phrase is semantically qualified can be determined by whether or not there is one Japanese noun phrase corresponding to the MT translation of the English noun phrase. This determination is made based on the following rule 1, and only when the MT translation of an English noun phrase and the Japanese noun phrase comply with rule 1, it is considered that there is one Japanese noun phrase corresponding to the MT translation of the English noun phrase. I will decide.
Rule 1: If the total similarity Si between a compound name MTNi (i = 1, 2,..., M) that constitutes an MT translation of an English noun phrase and a Japanese noun phrase is greater than or equal to the threshold th, MTNi is English It is said to be involved in the correspondence between noun phrases and Japanese noun phrases. When the number of compound nouns involved in the correspondence between English noun phrases and Japanese noun phrases is n, out of all n compound nouns that compose the MT translation of English noun phrases, the participation rate m / n is the threshold th _part Must be exceeded.
In the determination by rule 1, the processing for obtaining the overall similarity between the entire MT translation of the English noun phrase and the Japanese noun phrase, that is, (1) “Comparison in character units” to (4) “Semantic similarity” The process of “integration of evaluation and evaluation of phonological similarity” is applied between a component (compound noun) of an MT translation of an English noun phrase and a Japanese noun phrase. The threshold th _part is currently set to 0.5.

Rule 1 for the MT translation of “the United States into World War II”, a semantically ineligible English noun phrase “the United States into World War II”, and the Japanese noun phrase “World War II” Of the compound nouns that make up “United States to World War II”, “US” is the correspondence between “United States to World War II” and “World War II”. Since it is not involved, the participation rate is 0.5, and the threshold th _part is not exceeded, so the association between “the United States into World War II” and “World War II” is rejected.
This process not only eliminates the ambiguity of adding prepositions, but also eliminates the ambiguity of the parallel structure and suppresses association errors when there is no Japanese noun phrase in the Japanese sentence that correctly corresponds to an English noun phrase. Also effective.

The evaluation of the present invention was confirmed by the following experiment.
(experimental method)
For the evaluation experiment, we used the Yomiuri Shimbun and the top 30,000 sentences in the sentence correspondence score among the articles from mid-July 1996 to 2001 in the bilingual corpus of The Daily Yomiuri. A matching process is performed for this 30,000 sentence pair, and for each English noun phrase, the total similarity S is the threshold th = 0.1 or more, and the Japanese noun phrase having the highest total similarity among those that follow rule 1 Is output, otherwise nothing is output.
200 sentence pairs were sampled from the obtained data, and correct answer data was manually created from the 200 sentence pairs, and compared with the matching results according to the present invention. When the numbers of correct answers, omissions of correspondence, and association errors are C, M, and N, respectively, the recall, precision, and F value calculated by the following equations were used as indices for evaluating the performance of the present invention.
Reproducibility = C / (C + M)
Compliance rate = C / (C + N)
F value = (2 x recall x precision) / (fit precision + recall)

(Experimental results and discussion)
In the present invention, matching is performed by three processes: collation in character units that consider character types, collation of Roman readings, and collation that considers the involvement of each compound noun that constitutes the MT translation of an English noun phrase. On the other hand, the table of FIG. 4 shows the evaluation results in the case where the association is performed only by collation by character unit without considering the character type, and the association is performed by each of the present invention and the baseline. According to FIG. 4, the F value of the present invention is 0.678, which is an improvement from the baseline F value of 0.583 to 0.095. Comparing the recall rate and precision rate of the present invention with the baseline, the recall rate is 0.101 higher for the baseline, but the precision rate is 0.271 higher for the proposed method.In the present invention, the precision rate is improved while suppressing the fall of the recall rate. Has been realized. In the present invention, emphasis is placed on suppressing an association error rather than suppressing an omission of correspondence, so that it is considered that the intended goal is achieved by this result.

  Next, it was examined how much each process according to the present invention contributes to the improvement of performance. The results are shown in the table of FIG. In FIG. 5, the symbol “+” in the processing column means that the processing is introduced and associated, and the symbol “−” means that the processing is performed without the processing. (A) in which no processing is introduced is the performance of the baseline, and (h) in which all the processing is introduced is the performance of the present invention. Comparing (a) and (b), the introduction of processing that takes into account the involvement of compound nouns increased the number of correspondence omissions by 16, but the correspondence errors were greatly reduced to 81 cases. I understand. From this, (1) English noun phrases extracted from English sentences are syntactically or semantically ineligible, or (2) English noun phrases extracted from English sentences are syntactically and semantically qualified. However, when the correct answer does not exist in the Japanese sentence, the processing that considers the involvement of compound nouns works effectively to suppress the error that some Japanese noun phrase is associated with this English noun phrase. I can say that.

The degree to which the processing considering the similarity of reading contributes to the performance improvement is only one case in which the correspondence was not correct in the baseline, as can be seen by comparing (a) and (c). Because it is, it looks very low. (4) As described in “Integration of Semantic Similarity Evaluation and Phonological Similarity Evaluation”, this processing works when an MT translation of an English noun phrase includes an unregistered word. It is. Of the English noun phrases in the correct answer data, there were 20 cases where the MT translation contained unregistered words. Of these 20 cases, 8 did not get the correct answer at baseline. Of these, one case was improved by processing that considered the similarity of reading.
Comparing (a) and (e), the number of correspondence errors is reduced by introducing a process that considers the character type. The purpose of this process is to prevent characters other than kanji from matching in units that are too short, but the expected effect is obtained with the four correspondence errors that have been eliminated. On the other hand, the number of correct answers decreased by two and the number of correspondence omissions increased by two. This was a case where the two cases that were correct in the baseline resulted in an omission of correspondence in the process considering the character type. In this experiment, English letters and numbers are regarded as the same character type, and alphanumeric characters must be matched in units of the longest character string. For this reason, the correspondence between the MT translation of the English noun phrase “the Group of Seven” “Group of Seven” and the Japanese noun phrase “G7” cannot be obtained.

  Next, we will see how the F value changed by introducing multiple processes simultaneously. The F value 0.678 of (d), which introduced the processing considering the similarity of reading and the processing considering the involvement of compound nouns at the same time, is higher than the F value 0.586 of (c) where only the former processing was introduced, and It is higher than F value 0.667 of (b) which introduced only the latter process. The F value 0.669 of (f), which introduced the processing considering the character type and the processing considering the involvement of compound nouns at the same time, also applied the F values 0.581, 0.667 of (e) and (b), which introduced these processing individually. Exceed. The F0.585 value of (g), which introduced the processing considering the character type and the processing considering the similarity of reading at the same time, is higher than the F value 0.581 of (e) when only the processing considering the character type was introduced. When the processing considering the similarity of reading is introduced at the same time, it is slightly lower than the F value of 0.586 in (c). The F bond of the present invention (h) that introduced all three types of treatment is the highest. From the above, there are some side effects when the processing considering the character type and the processing considering the similarity of reading are introduced at the same time. In general, it can be said that these processing do not suppress the effects of each other.

  As described above, the present invention has shown a method for extracting an English proper expression having a prepositional phrase or a coordinate structure and a corresponding Japanese expression from a bilingual corpus. Although it is considered that the introduction of the participation rate shown in the present invention may not always be effective, the introduction of the participation rate shown in the present invention is a general tendency from the results of experiments using a bilingual corpus. , It was found that the matching accuracy is improved. In the present invention. The English proper expression and the Japanese expression were collated by the process considering the character type, the process considering the Roman reading, and the process considering the involvement of the constituent elements of the noun phrase. In the experiment using the bilingual corpus of Yomiuri Shimbun and The Daily Yomiuri, these three types of processing resulted in a precision of 0.764, recall of 0.610, and F value of 0.678. This result exceeds the result obtained when these processes are not performed and the result obtained when a syntax analysis system is used.

1 shows a block diagram of the present invention. Figure 2 shows a more detailed block diagram of the present invention. The flowchart figure of this invention is shown. The table | surface explaining evaluation of this invention is shown. The table | surface explaining the performance comparison for every process of this invention is shown.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input means 2 Output means 3 Operation | movement main memory 4 Memory | storage means 5 1st language sentence noun phrase extraction means 6 2nd language sentence noun phrase extraction means 7 Machine translation means 8 which translates from a 1st language to a 2nd language 2nd language Noun extraction means 9 Similarity determination means 10 between nouns or noun phrases in the second language Participation rate determination means 11 Second language morpheme analysis section 5a First language noun phrase extraction section 5b Article list of the first language,
5c First language preposition list 5d First language noun phrase determination pattern 6a Second language noun phrase extraction unit 6b Second language noun phrase determination pattern 8a Second language noun extraction unit 8b Second language noun determination pattern 9a Second language noun phrase determination pattern 9a Similarity determination unit 9b between nouns or noun phrases Characteristic similarity determination unit 9c Phonological similarity determination unit 9d General determination unit

Claims (12)

First language noun phrase extracting means for extracting a noun phrase of the first language sentence;
Translation means for translating the first language noun phrase into a second language;
First language noun extraction means for extracting nouns included in the translation result;
Second language sentence noun phrase extracting means for extracting a noun phrase of the second language sentence;
Similarities between second language nouns or between noun phrases that determine similarity between nouns extracted by the first language noun extraction means and nouns or noun phrases extracted by the second language noun phrase extraction means A determination means;
A participation rate determination unit for obtaining a ratio (participation rate) in which the similarity with the noun phrase of the second language sentence is equal to or higher than a similarity threshold among the nouns extracted by the first language noun extraction unit; A language processing device.   2. The first language noun phrase extraction unit includes a first language noun phrase extraction unit, a first language article list, a first language preposition list, and a first language noun phrase determination pattern. The language processing device described in 1.   The language processing apparatus according to claim 1, wherein the second language noun phrase extraction unit includes a second language noun phrase extraction unit and a second language noun phrase determination pattern.   4. The language processing apparatus according to claim 1, wherein the second language noun extraction unit includes a second language noun extraction unit and a second language noun determination pattern. 5.   The similarity determination means between the second language nouns or noun phrases includes a similarity determination unit between second language nouns or between noun phrases, a character similarity determination unit, a phonological similarity determination unit, The language processing apparatus according to claim 1, further comprising a determination unit.   The language processing apparatus according to claim 5, wherein the character similarity determination unit includes a comparison unit in units of characters and a comparison unit in consideration of a character type.   The language processing apparatus according to claim 1, further comprising a morphological analysis unit, wherein a word string is extracted from the noun phrase extracted by the second language noun phrase extraction unit.   The participation rate determination means calculates a participation rate from a ratio between the number of nouns in the first language and the number of nouns included in the noun phrase in the second language, and determines whether the participation rate exceeds a threshold value. The language processing apparatus according to claim 1, wherein:   The language processing apparatus according to claim 1, wherein a first language noun phrase and a second language noun phrase that exceed the participation rate threshold value are output in correspondence with each other.   A language processing apparatus according to any one of claims 1 to 9, wherein a dictionary for registering as a parallel translation is created by associating a first language noun phrase with a second language noun phrase corresponding thereto. Dictionary creation device. Extracting a first language noun phrase from the first language sentence;
Translating the first language noun phrase into a second language;
Extracting a noun included in the translation result;
Extracting a second language noun phrase from a second language sentence corresponding to the first language sentence;
Extracting a second language noun phrase containing the noun;
Determining a second language noun phrase including at least a predetermined number of second language noun phrases including the noun as a corresponding phrase of the first language noun phrase.   A language processing program for causing a computer to execute the steps according to claim 11.

Images