JP3305953B2 - Translation pattern creation method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a translation pattern creating apparatus and a method for creating a translation pattern by performing statistical processing on a bilingual document and used in a machine translation apparatus or the like.

[0002]

2. Description of the Related Art Generally, it is difficult to determine a translation word of a bilingual dictionary used for a machine translation device. This is because the translation of a word is not uniquely determined and depends on the word co-occurring with the word or on the context before and after. In addition, idioms and fixed phrases such as “you are welcome” and “welcome to you” are not always translated in units of words, and one translation is given when multiple words co-occur. Such translated words relating to the correspondence of plural words are also required.

In order to more easily create such a bilingual dictionary for a machine translation apparatus, for example, Japanese Patent Laid-Open No. 5-1512 is disclosed.
As disclosed in Japanese Unexamined Patent Publication No. 60, there is a system that reduces the work of creating a bilingual dictionary by using an existing bilingual sentence. Specifically, the existing bilingual sentence is correlated between the two languages using the parsing result and the bilingual dictionary or using human judgment, and based on the result, “you are welcom
e ”and a translation pattern like“ I'm glad you are ”.

In the above-mentioned system, a user simply prepares a bilingual sentence by using a mechanism for automatically recognizing the correspondence between words and phrases between bilingual sentences in the bilingual sentence. It is possible to easily create a translation pattern relating to a simple wording. It is not necessary for the user to identify a translation unit (a source language unit necessary for assigning a certain translated word) or associate words between an original sentence and a target sentence.

[0005]

However, the above-mentioned conventional system has the following problems. First, since the correspondence between words in a bilingual sentence is automatically identified using only the linguistic information in one bilingual sentence for which a bilingual dictionary is to be created, the conventional system requires a parsing process and a bilingual dictionary. I do. However, actual documents have many sentences that are difficult to parse. Furthermore, special documents such as technical documents contain many technical terms that are not registered in a general bilingual dictionary, but the conventional system could not identify a bilingual relationship in such a document.

Second, since the syntax analysis result and the bilingual dictionary are used as information for identifying the correspondence, if the syntax structures in the two languages are significantly different, or if a word pair not registered in the bilingual dictionary is used, If it was included in the bilingual sentence, the correspondence could not be identified.

Third, when the prepared translation is a document composed of a plurality of sentences such as a dissertation and a manual, linguistic information obtained from the entire document is useful. For example,
If a plurality of words always appear simultaneously in a document in one language, they can be regarded as one translation unit. To explain by giving a specific example, when "symbolic" and "link" appear simultaneously in a bilingual Japanese document, "symbolic" and "link" always appear simultaneously in the corresponding English document Then
"Symbolic link" can be identified as corresponding to "symbolic link". However, in the above-mentioned conventional system, linguistic information obtained from the entire document is not considered at all.

Fourth, in the above-mentioned conventional system, a bilingual dictionary is used as information for identifying a correspondence. However, a bilingual dictionary is only used as a material for identification, and a bilingual dictionary is extracted using a translation pattern extracted by itself. Did not have the means to improve. For this reason, it has not been easy to create a bilingual dictionary for each field (document), such as using a bilingual dictionary as information for identifying a correspondence, while adapting the bilingual dictionary itself to a given document.

In order to solve the above-mentioned problems of the prior art, linguistic information obtained from the entire bilingual document can be used, a syntactic analysis process, a translation pattern that does not require a bilingual dictionary can be performed, and an existing translation pattern can be created. It has been desired to realize a translation pattern creation method and apparatus having means for adapting a bilingual dictionary to a given bilingual document.

[0010]

The present invention employs the following structure to solve the above problems. <Structure of Claim 1> A translation pattern creating method according to the present invention comprises a bilingual sentence table.
And a target language, which is a translation of the source language document and the document
A bilingual document formed by associating word documents with each other in sentence units
The translation pattern creation device
The number of occurrences of the translated word pair and the original words in the word pair
The number of occurrences of the word in the source language document, and
Of the target language words in the target language document
The degree of correspondence between word pairs is calculated based on
Is extracted as a bilingual word, and the extracted unit
Remove word pairs from the bilingual sentence table and form a bilingual document again
The degree of correspondence is again obtained for the remaining word pairs,
The word pair with the highest degree of correspondence is extracted as a bilingual word.
And repeat these steps to sequentially extract bilingual words
It is characterized by.

<Explanation of Claim 1> The invention of claim 1 is characterized in that a bilingual sentence is extracted only by statistical processing without performing processing such as syntactic analysis on the bilingual sentence. For this purpose, first, the number of occurrences of an arbitrary word in the source language and the number of occurrences of an arbitrary word in the target language in the bilingual sentence are obtained. Further, the number of occurrences as an arbitrary word pair in the source language word and the target language word is determined. Find the number of times. Then, as the number of appearances becomes closer to each other, the higher the word correspondence, the higher the word correspondence is determined, and the word pair having the highest value of the word correspondence is identified as a bilingual word. That is, in the case of a bilingual word corresponding to a certain word, such a method is employed by paying attention to the point that the number of appearances of these words is a close value.

The source language and the target language are, for example, languages such as Japanese and English, but these languages may be any languages. A bilingual document that is correlated in units of sentences is only required to correspond to a source language sentence and a corresponding target language translation in unit of sentence. This does not have to be one sentence to one sentence. A plurality of sentences or a plurality of sentences versus a plurality of sentences may be used.

The degree of word correspondence is obtained by, for example, (2 × the number of appearances of a Japanese-English word pair) / (the number of appearances of a Japanese word + the number of appearances of an English word), but is not limited to this equation. If the number of occurrences of the word in the source language, the number of occurrences of the word in the target language, and the number of occurrences of these words in the same translation are closer, the higher the word correspondence, the higher the word correspondence. Various calculation methods are also available.

[0014] Then, the invention of claim 1, further for the previous translation word extraction processing for extracting a word pairs that could not be extracted, removed from the object of counting the word pairs extracted by the first time, again Find the degree of correspondence. As a result, the degree of word correspondence of a certain word pair becomes higher than the first time, and this word pair can be extracted in the second and subsequent processing.
By repeating the processing in this manner, all word pairs existing in the bilingual sentence can be extracted.

[0015] In the translation pattern generating method according to claim 1 <Configuration of claims 2>, the translation pattern forming apparatus, when the extracted word pairs and the original word pair, among the original word pair, Looking at the other word from the one word, there is another word that has the same sentence as the other word in the original word pair and the same number of occurrences in the pair as the one word, and A translation pattern creation method for extracting a word from the other word of the original word pair and extracting the word as a multiple word-to-multiword translation pattern co-occurring with the original word pair when there is a similar word. It is.

[0016] invention <claims described in Section 2> claim 2, in which so as to extract a plurality of word pairs plurality of word translation pattern. Here, the one-sided word and the other-sided word are, for example, English and Japanese,
These languages can correspond to any language. Further, the plurality of words need not be a continuous word string in the document, and can be applied even when any other word is included between the words.

[0017]

[0018] The invention of claim 3 <Configuration of Claim 3> is stored in the translated sentence table was formed in association with sentences and target language documents each other is the translation of the source language document and the document For the bilingual document, the translation pattern device uses the number of appearances of the word sequence pair for the bilingual and word sequence pair consisting of continuous words in the bilingual document;
The degree of correspondence between word string pairs is determined based on the number of appearances of the source language word string in the word string pair in the source language document and the number of appearances of the target language word string in the word string pair in the target language document. The word string pair having the highest degree of correspondence is extracted as a translation pattern, the extracted word string pair is removed from the bilingual sentence table, and a bilingual document is formed again. And a word string pair having the highest value of the degree of correspondence is extracted as a translation pattern, and these are repeated to sequentially extract a translation pattern.

[0019] <Description of claim 3> The invention of claim 3, obtains the word string corresponding degree, by the word string corresponding degree, is obtained so as to extract a word string pairs. The degree of correspondence between the word string pairs is obtained by, for example, (2 × the number of appearances of a Japanese-English word string pair) / (the number of appearances of a Japanese word string + the number of appearances of an English word string), but is limited to this equation. Rather than the number of occurrences of the source language word sequence,
Any calculation method can be used as long as the number of appearances of the word string in the target language and the number of times these word strings appear in the same translation are closer to each other, as long as the word string correspondence degree becomes higher. .

Further , as a word string composed of continuous words,
A plurality of consecutive times such as two consecutive words and three consecutive words can be arbitrarily set, and also includes one word. According to such a translation pattern creating method, it is possible to obtain a translation pattern in which not only a translated word in a translated document but also a fixed phrase appearing in the translated document is used as a translation unit.

[0021] Then, it is performed by repeating the further generation and a corresponding degree of calculation of the word sequence pairs remaining in the new translation document the generated extracted word sequence pairs translated document obtained by removing all the word sequence pairs Are sequentially extracted from the bilingual document.

[0022] The invention <claimed structure of claim 4> claim 4, an input unit for inputting a bilingual document documents in the target language are respectively associated with the sentence is a document and translation of this document in the original language, For a bilingual document, the number of occurrences of a specific word in the source language document, the number of occurrences of a specific word in the target language document, and the specific words in the same translation of the source language and target language documents The higher the number of occurrences of the word is, the higher the word correspondence that is higher, the word pair having the highest value of the word correspondence is extracted as a bilingual word, and any word pair is taken as a bilingual word. If extracted, a new bilingual sentence excluding the word pair is generated from the bilingual sentence targeted at this time, and the degree of word correspondence is calculated for the new bilingual sentence. Word pairs as translated words A translation pattern estimating unit that sequentially extracts bilingual words by repeating the process, and an output unit that outputs the bilingual words extracted by the bilingual word estimating unit. is there.

[0023] <claims described in the section 4> input section is for user inputs a bilingual document. When a bilingual document is input from the input unit, the bilingual word estimating unit performs the same processing as the method of claim 2 and sequentially outputs the bilingual word. The output unit outputs the translated word extracted by the translated word estimating unit on a screen or prints. Therefore, even if the user does not follow the grammar,
By simply inputting the bilingual document, a bilingual word can be obtained, and there is an effect that a bilingual dictionary is not required.

[0024] In the translation pattern forming apparatus according to claim 4 <Configuration of Claim 5>, and bilingual dictionary with translation of advance source language and target language, the word pairs that appear in the bilingual document, registered in the bilingual dictionary If so, a bilingual sentence excluding the word pair is generated, a word correspondence is calculated for the translated sentence, and a word pair having the highest value of the word correspondence is extracted as a bilingual word. And a translation pattern creating apparatus comprising:

The invention <claims described in Section 5> claim 5 is the invention of claim 4, in which to use a previously prepared translation dictionary. Thus, similarly to the invention of claim 4, a bilingual word can be extracted even from a sentence that does not follow the grammar, and information of a bilingual dictionary can be used when extracting a bilingual word. A highly accurate bilingual word can be extracted by effectively utilizing both the information of the bilingual word and the information of the bilingual word from the bilingual dictionary.

The invention of claim 6 <Configuration of claims 6>, in the translation pattern forming apparatus according to claim 5, a pattern table for storing word pairs extracted as translation, any of the source language in the translation dictionary It is determined whether the word and the translation of the target language for the word match the word pair in the pattern table. If not, the translation for the arbitrary word is deleted in the bilingual dictionary, and the word pair in the pattern table is deleted. A translation pattern creation device is provided with a bilingual word estimation unit for registering as a translation word.

The invention <claims described in Section 6> claim 6, in addition to the invention of claim 5, in which to perform the tuning bilingual dictionary. This tuning is a process of registering a word pair registered in the pattern table in the bilingual dictionary when the word pair is not in the bilingual dictionary. Therefore, by simultaneously providing a bilingual dictionary and a bilingual document having the same field or characteristics as the document to be translated by the user to the present apparatus, the bilingual dictionary is tuned to a bilingual dictionary capable of obtaining a translation result equivalent to the bilingual document. be able to. For example, by performing this work for each document in a different field, a bilingual dictionary specialized for the field can be created. Therefore, this can be used as a technical term dictionary for each field in a machine translation device or the like.

The invention of claim 7 <Configuration of claims 7>, in the translation pattern forming apparatus according to claim 4, the word pairs extracted by the translation word estimator and the original word pair, the words of the source In the pair, looking at the other word from the one word, the other word having the same sentence as the other word in the original word pair and having the same number of occurrences in the pair as the one word is found. If there is and there is one similar word when viewed from the other word of the original word pair, these words are extracted as a multi-word-multi-word translation pattern co-occurring with the original word pair A translation pattern creation device comprising a co-occurrence word estimation unit.

The co-occurrence word estimator <Description of claims 7>, when translated document from the input unit is inputted, the same processing as method of claim 5, outputs the multi-word pairs of multi-word translation pattern Then, the output unit performs output such as screen display and printing. Therefore, even if a user does not follow the grammar, the user can input the bilingual document, and not only the bilingual word, but also the translation pattern using the fixed phrase appearing in the bilingual document as a translation unit. Can be acquired.

The word invention of claim 8 <Configuration of claims 8> is the translation pattern forming apparatus according to claim 7, appearing with bilingual dictionary with translation of advance source language and target language, in the bilingual document If the pair is registered in the bilingual dictionary, a bilingual sentence excluding the word pair is generated, the degree of word correspondence is calculated for the bilingual sentence, and the word pair having the highest value of the word corresponding degree is translated. A translation pattern creating device comprising a bilingual word estimating unit for extracting as a word.

The invention <claims described in Section 8> 8. is the invention of claim 7 is obtained by adding the ability to use bilingual dictionary. That is, claim 5
And a function of a co-occurrence word estimating unit according to the seventh aspect of the present invention. With this configuration, when extracting a bilingual word, the information in the bilingual dictionary can be used, so that both the information of the input bilingual document and the information of the bilingual word from the bilingual dictionary can be effectively used. It is possible to extract highly accurate bilingual words, and by simply inputting a bilingual document, obtain not only the bilingual word but also a translation pattern that uses fixed phrases that appear in the bilingual document as translation units. be able to.

The invention of claim 9 <Configuration of claims 9>, in the translation pattern forming apparatus according to claim 8, a pattern table for storing word pairs extracted as translation, any of the source language in the translation dictionary It is determined whether the word and the translation of the target language for the word match the word pair in the pattern table. If not, the translation for the arbitrary word is deleted in the bilingual dictionary, and the word pair in the pattern table is deleted. A translation pattern creation device is provided with a bilingual word estimation unit for registering as a translation word.

The invention <claims described in Section 9> claim 9, in addition to the invention of claim 8, in which to perform the tuning bilingual dictionary. Therefore,
In addition to the effect of the invention of claim 8, by providing a bilingual dictionary and a bilingual document having the same field or characteristics as the document to be translated by the user to the apparatus at the same time, the bilingual dictionary is translated equivalent to the bilingual document. There is an effect that tuning can be performed on a bilingual dictionary that can obtain a result. For example, by performing this work for each document in a different field, a bilingual dictionary specialized for the field can be created. Therefore,
This can be used for a machine translation device or the like as a technical term dictionary for each field.

[0034]

The invention of claim 10 <Configuration of claims 10> is an input unit for inputting a bilingual document formed in association with each other sentences and target language document is translated sentence of the source language document and the document And the number of appearances of the word string pair, the number of appearances of the source language word string in the word string pair in the source language document, and the word string pair Based on the number of appearances of the target language word string in the target language document, the degree of correspondence between the word string pairs is determined, and the word string pair with the highest degree of correspondence is extracted as a translation pattern, and the extracted word string is extracted. Except for the pair, a bilingual document is formed again, the degree of correspondence is again obtained for the word string pair, the word string pair having the highest value of the degree of correspondence is extracted as a translation pattern, and these are repeated to sequentially execute the translation pattern. Extract The translation pattern forming apparatus characterized by comprising a translation pattern estimation unit.

The translation pattern estimation unit <Description of claims 10>, when translated document from the input unit is inputted, the same processing as method of claim 3, and outputs a word string translation pattern, the output unit Performs output such as screen display and printing. Therefore, even if a user does not follow the grammar, the user can input the bilingual document, and not only the bilingual word, but also the translation pattern using the fixed phrase appearing in the bilingual document as a translation unit. Can be acquired. In addition, one word can be extracted as the translation pattern.

[0037] Then, the translation pattern estimation unit, after extracting the highest degree of correspondence word string pairs from bilingual documents stored in the translation table, a new translation document to erase the word sequence pairs issued extract The method according to claim 3, wherein a word string pair is extracted from the generated bilingual document based on the degree of correspondence again, and these are repeatedly performed to extract all word string pairs in the bilingual document. The method can be realized.

[0038]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below using specific examples. << Specific Example 1 >><Configuration> FIG. 1 is a configuration diagram of a specific example 1 in the translation pattern creating apparatus of the present invention. This specific example 1 is characterized in that a bilingual word (translation pattern in word units) can be automatically extracted without using a bilingual dictionary.

The apparatus shown in the figure has an input / output unit 11 for a user to input a bilingual document and to display a created bilingual word.
0, a bilingual word extraction unit 120 for extracting a bilingual word, and a resource management unit 130 for storing resources used when creating a translation pattern.

The input / output unit 110 includes an input unit 111 for a user to input a bilingual document serving as a source of a bilingual word, and an output unit 11 for displaying and correcting a bilingual word.
2. Parameter setting unit 11 for setting output format and the like
3

The bilingual word extraction unit 120 performs a morphological analysis on the bilingual document for each document in each language, a bilingual word estimating unit 122 that identifies the correspondence between words in both languages, and estimates the bilingual word. , A bilingual word generation unit 123 that generates a bilingual word in a format set by the user.

The resource management unit 130 includes a morphological dictionary 131 for storing a dictionary for morphologically analyzing bilingual documents in a bilingual document, a bilingual document storage unit 132 for storing morphological analysis results of an input bilingual document, and a bilingual word. At the time of extraction, a table storage unit 133 for storing various tables created temporarily
Consists of

<Operation> FIG. 2 is a flowchart of the translation pattern creation processing of the first embodiment. Hereinafter, a description will be given assuming that the source language of the translation pattern creation processing is Japanese and the target language is English. In the illustrated flowchart, the translation pattern creation device extracts bilingual words from bilingual documents in English and Japanese,
This is an explanation of the flow of processing up to the display. Here, the operation will be described by taking as an example the case where the user inputs the following bilingual document.

FIG. 27 shows an example of a bilingual document for which a bilingual dictionary is created by the apparatus. The bilingual document is composed of a Japanese document and an English document, and it is assumed that the two documents are associated with each other as indicated by the line number at the left end of the line. By using the sentence association algorithm shown in Japanese Patent Application No. 7-192134 filed earlier, it is possible to input a sentence that is not associated. Further, the sentence need not be a sentence-to-sentence translation, but may be a sentence-to-plurality sentence (for example, an English translation for one sentence in Japanese is composed of a plurality of sentences), or a plural-sentence-to-multiple sentence. It does not matter.

First, the user inputs a bilingual document for which a bilingual dictionary is to be created by the input unit 111 (step S).
101), the morphological analysis unit 121 divides the bilingual document into a Japanese document and an English document, and performs morphological analysis using the morphological dictionary 131 of each language (step S102).

FIGS. 28, 29 and 30 are explanatory diagrams showing the results of morphological analysis in the bilingual document of FIG. 27. As shown in the figure, the results are shown in each of the Japanese and English documents as the appearance results (1) and (2) shown in FIG.
There are two types of results, standard results (1) and (2) shown in FIG. The appearance form result is obtained by storing the result of morphological analysis of the input sentence as an appearance form heading (as it appears), using blanks as delimiters. On the other hand, the standard result is
The result of the morphological analysis of the input sentence is stored in a form in which a blank is used as a delimiter and a standard form heading is followed by “symbol shown + candidate part of speech”. When there are a plurality of candidate parts of speech, all candidate parts of speech are listed using "/" as a delimiter. Then, the respective morphological analysis results are stored in the bilingual document storage unit 132.

Next, the translated word estimation section 122 performs a translated word estimation process (step S103). Here, bilingual Japanese and English word pairs are extracted. The bilingual word estimation process (step S103 → step S104) is repeated until there is no more word pair to be extracted. If no word pair is found in step S104, the bilingual word generation process of the bilingual word generation unit 123 (step S105)
Move on to In the bilingual word generation processing, the bilingual word is shaped into a format set by the user in the parameter setting unit 113. Then, the result is displayed on the output unit 112 (step S10).
6), end the process.

Next, details of the algorithm of the translated word estimation unit 122 and an example of specific processing will be described. FIG.
FIG. 4 is a flowchart showing the algorithm of the translated word estimation process (steps S103 and S104 in FIG. 2) of the translated word estimation unit 122.

First, 1 is set to a pattern creation counter n (step S201). The pattern creation counter indicates the number of parallel translation word estimation processes that are performed repeatedly,
It is used to manage a bilingual sentence table (SentJn, SentEn) described later.

The standard form results in the morphological analysis results (FIG. 2)
9. Japanese words that include independent words (name (noun), verb (verb), form (adjective or adjective verb), and adverb (adverb) in candidate words, from (1) and (2) in FIG. 30) , N (noun), v (verb), adj (adjective), adv
The word headings of (English words including (adverb)) are translated into the bilingual sentence table of each language (SentJn [line number] = {Japanese bilingual word list}, SentEn [line number] = {English bilingual word list}
(N: pattern creation counter (n = 1))) (step S202).

FIG. 31 shows the results of the morphological analysis described above (FIG. 2).
30 is an example of a bilingual sentence table from the first to third lines of the bilingual document of FIGS. 8 to 30). The bilingual sentence table for each language is composed of a line number serving as an index and a list of independent word headings in the sentence serving as its value. Hereinafter, each table is represented by “variable name [index] = value”.

Next, the number of appearances of each word appearing in the bilingual sentence table is counted, and a word table (FreqJ [Japanese word] = frequency of appearance of Japanese word, FreqE [English word] ] = Number of appearances of English words) (step S203). FIG. 32 shows an example of a word table for each language.

Furthermore, the number of appearances of a combination (word pair) of a Japanese word and an English word included in a bilingual sentence is counted, and the number of appearances and the line number of the appearing sentence are used as a word pair table using each word pair as an index. (FreqJE [Japanese word English word] = the number of times a Japanese word and an English word appear as a bilingual sentence, SentJE [Japanese word English word] = {List of appearance line numbers}) (step S204). FIG. 33, FIG.
4 is an explanatory diagram of the word pair table, and FIG. 33A shows a state after counting the number of appearances.

Next, the degree of word correspondence in the word pair stored in the word pair table is calculated (step S20).
5). The degree of word correspondence is calculated by the equation (SimJE [JW EW] = 2 * FreqJE [JW EW] / (FreqJ) shown in FIG.
[JW] + FreqE [EW]) (JW: Japanese word, EW: English word) is calculated (hereinafter this value SimJE [JWEW] is called Japanese-English correspondence), and the value is stored in the word pair table. to add. The degree of word correspondence takes a value from 0 to 1, and a word pair having a bilingual relationship takes a value closer to 1. The result of calculating such word correspondence is the value shown in (2) of FIG.

Next, a word pair table ((2) in FIG. 34)
The Japanese words and English words stored in the index of the Japanese language) have the highest Japanese-English correspondence (SimJE [Japanese words English words]) among the English words paired with the Japanese words Extract English words. If the Japanese word having the highest Japanese-English correspondence among the Japanese words paired with the English word is the original Japanese word, the word pair is identified as a translated word (step S206). A pattern table (Pattern [Japanese word English word]
= {List of Appeared Line Numbers}) (Step S2)
07). FIG. 35 is an example of a pattern table.

If there are a plurality of word pairs having the highest degree of Japanese-English correspondence, all the word pairs are stored.
For example, in the case of FIG. 34 (2), the English words forming the index together with the Japanese word [information] are “information”,
"Display" and "resource", the English word with the highest Japanese-English correspondence (SimJE) is "information"
On the other hand, the Japanese word with the highest Japanese-English correspondence in “information” is “information”. As a result, both words are identified as parallel words, and stored as Pattern [information information] = {314...} In the pattern table.

If a bilingual word is extracted (step S
208), 1 is added to the pattern creation counter n (step S209), and the bilingual sentence table (SentJn [line number], Se
ntEn [line number] n: pattern creation counter (n = 2)
It returns to the creation process (step S202). Where n is 2
In the above case, when the morphological analysis result is read into the bilingual sentence table, if a word pair registered in the pattern table (Pattern [Japanese word English word]) exists in the bilingual sentence, the word is stored in the bilingual sentence table. Do not store.
As an example, if the pattern creation counter n is 2, and FIG.
If there are 5 pattern tables, “initial initial” and “information informa” stored in the pattern table
“tion” is not stored in the bilingual sentence table.
FIG. 4 is an explanatory diagram of a bilingual sentence table in the case of No. 2;

As shown in the figure, "initial initial" and "information" stored in the first bilingual sentence table (see FIG. 31).
“information” is not stored in the new bilingual sentence table.
6), the above processing is repeated (steps S202 to S202).
208) When there are no more word pairs to be extracted, the process proceeds to a bilingual word generation process (step S210). The reason why the process is repeated here is that by performing the repetition process, a word pair that could not be extracted in the previous bilingual word estimation process can be extracted. For example, in the word pair table (FIG. 34 (2)), "Information informati
on ”exists with a word correspondence of 0.43,
Since there is “information information” having a higher degree of correspondence, it cannot be extracted at the first time. But the second time, "Information
By removing "information" from the list,
The word correspondence of “information information” has a high value, and the “information i”
nformation ”can be extracted.

FIG. 37 and FIG. 38 are explanatory diagrams when the elements are simplified in the above-described bilingual word estimation processing. Now, as shown in the bilingual sentence table for n = 1 in FIG. 37, three “information” and two “information”
And one "display" appear in the English document
"Appear six times. This value is stored in the word table, and a word pair table is generated. For example, SimJE of “information information” is 2 × 3 / (3
+6) ≒ 0.67, which is the highest price compared to other word pairs, and the highest value Japanese word in “information” is “information”, so this word pair is used as a bilingual word. Extract. Then, this word pair is stored in the pattern table.

In the case of the second translation word estimation process (n = 2), since "information" is removed from the first translation text, the contents of the translation text table are as shown in FIG. As a result, this time, "Information
SimJE of "information" has the highest value (0.8), and the Japanese word having the highest value of "information" is "information", so this word pair is extracted as a bilingual word. Then, this word pair is added to the pattern table. In this way, bilingual words can be sequentially extracted.

Returning to FIG. 2, the translated word generation process (step S105 in FIG. 2) is performed by a pattern table (Pattern [Japanese word English] created by the user in accordance with the request generated by the parameter setting unit 113 in the translated word estimation process. Word]). For example, when the user gives a request to “output a bilingual word in a word-to-word format”, the bilingual word generation unit 123 receives the instruction from the parameter setting unit 113, shapes and outputs the result of the pattern table. . When a command of “output with part of speech information” is given, the bilingual word generation unit 123 extracts one line number of the line number list stored in the pattern table, and generates a bilingual sentence table indicated by the line number. refer. Then, the number of the corresponding word is extracted as an independent word, the words located in that order are extracted from the result of the morphological analysis, and the part of speech stored therein is output together with the part of speech. In addition, the bilingual sentence of the line number stored in the pattern table (the morphological analysis result appearance form result (FIG. 29, 3
0 (1), (2)) can also be displayed.

FIG. 39 is a diagram showing a display example of the first embodiment. This example is an example of an output result of a translated word from the output unit 112 when the user gives an instruction of “outputting a translated word in a word-to-word format”. Note that the user can check whether the result is correct or not by using the input unit 111 and correct the result.

<Effects> As described above, the first embodiment has the following effects. In the past, parsing was required, so that sentences that did not follow the grammar could not be automatically extracted.However, the present apparatus and method do not require parsing, so even if the sentence does not follow the grammar,
Bilingual words can be extracted.

Conventionally, a bilingual dictionary was required. However, since the degree of word correspondence is determined from the number of appearances of words and word pairs in the entire bilingual document and is used for extraction, even if the bilingual dictionary does not exist, bilingual translations in the bilingual sentence can be obtained. You can get words.

<< Specific Example 2 >><Structure> FIG. 5 shows a specific example 2 of the translation pattern creating apparatus.
FIG. In this specific example, when extracting a bilingual word, an existing bilingual dictionary is used. The feature of this example is that an existing bilingual dictionary and a bilingual document having the same field or characteristics as the document to be translated by the user are given to the system at the same time, so that the existing bilingual dictionary is translated equivalent to the bilingual document. The point is that you can tune to a bilingual dictionary that gives you the results.

The configuration of the specific example 2 includes an input / output unit 210, a bilingual word extraction unit 220, and a resource management unit 230. These configurations are similar to the specific example 1, but the bilingual dictionary 234 exists in the resource management unit 230. Is different. In addition, in the input / output unit 210, the input unit 211 receives not only a bilingual document but also a bilingual dictionary, and the output unit 212 has a function of displaying an existing bilingual dictionary tuned by the input bilingual document. Is different.

In the figure, the parameter setting unit 213 in the input / output unit 210, the morphological analysis unit 221 in the bilingual word extraction unit 220, the bilingual word estimating unit 222, the bilingual word generation unit 223, and the resource management unit 230 have other configurations. The morpheme dictionary 231, the bilingual document storage unit 232, and the table storage unit 233 in
13, the morphological analysis unit 121, the bilingual word estimation unit 122, the bilingual word generation unit 123, the morphological dictionary 131, the bilingual document storage unit 132, and the table storage unit 133 have the same functions.

<Operation> FIG. 6 is a flowchart showing an algorithm for creating a bilingual dictionary in the second embodiment. The operation of this specific example 2 is basically the same as that of the specific example 1, except that (1) at the time of input, a bilingual dictionary to be tuned is read simultaneously with the bilingual document (step S301), (2) ) Bilingual word estimation processing (step S3)
03), using a bilingual dictionary, and (3) performing bilingual dictionary tuning in the bilingual word generation processing (step S305). The following description focuses on the differences.

The bilingual document used in Example 1 (see FIG. 27)
The following describes a case in which the following bilingual dictionary (Dic [Japanese word English word] (value omitted)) is tuned. FIG. 40 shows an example of the bilingual dictionary.

First, the point of inputting a bilingual document in step S301 and the Japanese / English morphological analysis processing in step S302 are the same as those in the first embodiment, and a description thereof will be omitted. On the other hand, in this specific example 2, step S3
At 01, the user inputs a bilingual dictionary to be tuned at the same time as the bilingual document through the input unit 11. This bilingual dictionary is stored in the bilingual dictionary 234 of the resource management unit 230.

After each morphological analysis result is stored in the bilingual document storage unit 232 by the morphological analysis process (step S302), the bilingual word estimation unit 222 performs a bilingual word estimation process (step S303). Here, the bilingual dictionary 234 is used to extract Japanese and English word pairs that are bilingual candidates. In the determination of the next step S304,
If there is a word pair that is a translation candidate, the process returns to the translation word estimation process in step S303. Such a bilingual word estimation process is repeated until there is no more word pair to be extracted. If there is no more word pair in step S304, the processing is transferred to the bilingual word generation unit 223, and the format specified by the user in the parameter setting unit 213. Tunes the bilingual dictionary (step S305), displays the result on the output unit 212 (step S306), and ends the processing.

The details of the algorithm after the parallel word estimation process and an example of a specific process will be described below. FIG.
FIG. 8 and FIG. 9 are flowcharts showing the algorithm of the translated word estimation processing (the above-described steps S303 and 304) of the translated word estimation unit 222.

First, 1 is set to the pattern creation counter n (step S401). The pattern creation counter indicates the number of parallel translation word estimation processes that are performed repeatedly,
It is used to manage a bilingual sentence table (SentJn, SentEn) described later. Next, in the standard form results (FIGS. 29 and 30 (1) and (2)) in the morphological analysis results, the bilingual words stored in the bilingual dictionary (Dic [Japanese word English word]) are included in the bilingual sentence. Is included in the pattern table (Pattern [Japanese word English word] =
(List of occurrence line numbers)) (step S40).
2).

FIG. 41 shows the pattern table after the translation of the bilingual dictionary. The pattern table is composed of the bilingual dictionary shown in FIG.
9 shows a pattern table created from the bilingual sentence table shown in FIG.

The translated words stored in the bilingual dictionary 234 and included in the bilingual sentence are regarded as more reliable correspondences. By excluding the bilingual word from the target of counting, the value of the word correspondence degree of a word pair having a bilingual relationship other than the bilingual word stored in the bilingual dictionary 234 becomes higher, thereby increasing the value as compared with the case where the bilingual dictionary 234 does not exist. Can also extract many word pairs. Further, this processing is to detect a translation word that appears in the bilingual document and is stored in the bilingual dictionary 234, so that the detected bilingual word is prioritized to adapt to the bilingual document. It is helpful. That is, when there are a plurality of different translations in the bilingual dictionary 234, it is possible to specify the translation in the bilingual document.

Next, based on the standard form results in the morphological analysis results, Japanese words containing independent words (name (noun), verb (verb), form (adjective or adjective verb), adverb (adverb) in candidate parts of speech The word heading of the word and the candidate part of speech including n (noun), v (verb), and adv (adverb) in the bilingual sentence table of each language (SentJn [line number] = {Japanese bilingual sentence word list) ｝, SentEn [line number] = {English translation word list}
(N: pattern creation counter (n = 1))) (step S403). However, when reading the morphological analysis result into the bilingual sentence table, the pattern table (Pattern
[Japanese word English word]), and if a bilingual word stored in the pattern table exists in the bilingual sentence, each word of the bilingual word is not stored in the bilingual sentence table. As an example, when the pattern table of FIG. 41 exists, “start s” stored in FIG.
tart ”,“ instruction direction ”,“ field field ”
Is not stored in the bilingual sentence table. Subsequent processing (steps S404 to S411) in the bilingual word estimation processing conforms to the first specific example, and a description thereof will be omitted.

The bilingual word generation process in the next step S411 is a process of integrating the bilingual dictionary 234 with the bilingual words stored in the pattern table.

FIG. 10 is a flowchart showing the algorithm in the bilingual word generation processing, and shows the bilingual dictionary tuning processing. First, the bilingual dictionary 234 is read (step S501), and one element (Dic [Japanese word English word]) is taken out (step S502). In the next step S503, it is checked whether any of the extracted Japanese words or English words is registered in the pattern table. If not, nothing is performed. On the other hand, if any one is registered, it is checked whether a bilingual translation between a Japanese word and an English word is stored (step S504). If it is stored, other translations are stored in the translation dictionary, and the stored translation is deleted from the pattern table (step S505). Step S
If it is not stored in 504, the bilingual dictionary 23
4. Storage of Japanese and English words (Dic [Japanese word
English words]) and delete the other translations into the bilingual dictionary 234.
And deletes the translation from the pattern table (step S506). Such a process is performed on the bilingual words in all the bilingual dictionaries 234, and finally, all the bilingual words remaining in the pattern table are stored in the bilingual dictionary 234 (step S507).

As a result, the bilingual words in the bilingual dictionary 234 are
If a bilingual word extracted from the bilingual document exists, the bilingual word is stored. If not, the original bilingual dictionary 23 is stored.
4 bilingual words are stored. Therefore, the translation word stored in the bilingual dictionary 234 is a translation word that is more adapted to the bilingual document. Finally, the bilingual dictionary 234 tuned here
Is displayed on the output unit 212 by the user. After confirming the correctness, the user can correct the result.

<Effects> As described above, the specific example 2 has the following effects. In the past, parsing was required, so that sentences that did not follow the grammar could not be automatically extracted.However, with the present apparatus and method, bilingual words can be extracted even for sentences that do not follow the grammar, because parsing is not required. it can.

Since the bilingual dictionary can be used to determine the degree of word correspondence from the number of appearances of words and word pairs in the entire bilingual document, a high-precision bilingual translation effectively utilizing the information of both information (the bilingual document and the bilingual word) Word extraction can be performed.

By simultaneously providing an existing bilingual dictionary and a bilingual document having the same field or characteristics as a document to be translated by a user to the present system, the existing bilingual dictionary can be translated to obtain a translation result equivalent to that of the bilingual document. Can tune to a dictionary. By performing this work for each document in a different field, a bilingual dictionary specialized for the field can be created. This can be used for a machine translation device or the like as a technical term dictionary for each field.

<< Embodiment 3 >><Structure> FIG. 11 is a block diagram of Embodiment 3 in the translation pattern creating apparatus of the present invention. This specific example is characterized in that a translation pattern can be created using a fixed pattern that appears multiple times in a document as a translation unit without using a bilingual dictionary. For example, if "symbolic" and "link" appear more than once at the same time, if "symbolic" and "link" always appear at the same time in the corresponding English sentence,
“Symbol link” can create a translation pattern using “symbolic link” as a translation unit.

The configuration of the third embodiment includes an input / output unit 310, a resource management unit 330, and a translation pattern extraction unit 340. The input / output unit 310 and the resource management unit 330 conform to the configuration of the first embodiment. Further, the translation pattern extraction unit 340
It has the same function as the bilingual word extraction units 120 and 220 in the specific examples 1 and 2, but the bilingual word estimation units 122 and 222
Are different in the translation pattern estimation unit 342 corresponding to
It comprises a bilingual word estimator 342a for estimating the correspondence between words in both languages and a co-occurrence word estimator 342b for estimating the co-occurrence between words in the language. Further, the translation pattern generation unit 343 has a function corresponding to the translated word generation units 123 and 223 in the first and second specific examples.

In FIG. 11, the other components of the input / output unit 310, the morphological analysis unit 341 of the translation pattern extraction unit 340, and the resource management unit 330 are the same as those of the first embodiment. It has the same function as each component of the unit 110, the morphological analysis unit 121, and each component of the resource management unit 130.

<Operation> FIG. 12 is a flowchart showing a process in which the present apparatus creates and displays a translation pattern from a bilingual document in English and Japanese. Also in this specific example, the user
The operation will be described by taking as an example a case where a bilingual document is input.

The input of the bilingual document (step S 601) and the Japanese / English morphological analysis processing (step S 602) are the same as those in the first embodiment, so that the description will be omitted. Then, after each morphological analysis result is stored in the bilingual document storage unit 332 by the morphological analysis process of step S602, the bilingual word estimation unit 342a of the translation pattern estimation unit 342 performs the bilingual word estimation process (step S603). Here, as in the first specific example, Japanese and English word pairs serving as translation candidates are extracted.

In the next step S 604, when there is a word pair to be a translation candidate,
The co-occurrence word estimating unit 342b performs the co-occurrence word estimation processing (step S605) and the translation pattern word order determination processing (step S606), and returns to the bilingual word estimation processing of step S603 again. The bilingual word estimation process and the co-occurrence word estimation process are repeated until there are no more word pairs to be extracted. If there is no word pair in step S604, the process proceeds to the translation pattern generation process by the translation pattern generation unit 343 (step S607). Thereafter, the translation pattern is shaped into the output format designated by the user with the parameter setting unit 313 (step S607), the result is displayed on the output unit 312 (step S608), and the process ends.

The details and specific examples of the algorithm after the bilingual word estimation processing will be described below. 13, 14,
FIG. 15 is a flowchart illustrating an algorithm of the translated word estimation process (steps S603 and S604) performed by the translated word estimation unit 342a.

First, 1 is set to the pattern creation counter n (step S701). The pattern creation counter indicates the number of parallel translation word estimation processes that are performed repeatedly,
It is used to manage a bilingual sentence table (SentJn, SentEn) described later.

The standard form results in the morphological analysis results (FIG. 2)
9. From FIG. 30 (1) and (2)), a Japanese word including a self-sufficient word (name (noun), verb (verb), form (adjective or adjective verb), adverb (adverb) in candidate part of speech) N in part of speech
(English words including (noun), v (verb), and adj (adverb))
Of the word heading of each language in the bilingual sentence table (SentJn [line number] = {Japanese bilingual sentence word list}, SentEn [line number]
= {English bilingual sentence word list} (n: pattern creation counter (n = 1))) (step S702). It is assumed that the result is the bilingual sentence table shown in FIG.

Next, the number of appearances of words appearing in the bilingual sentence table is counted, and the word table (FreqJ [Japanese word] = the number of appearances of Japanese words, FreqE [English word]) is indexed by the word heading of each word. = The number of occurrences of English words)
(Step S703). It is assumed that the created word table is the word table shown in FIG.

Further, the number of appearances of a combination (word pair) of a Japanese word and an English word included in the bilingual sentence is counted, and the number of appearances and the line number of the appearing sentence are used as a word pair with each word pair as an index. Table (FreqJE [Japanese word
(English word) = the number of times a Japanese word and an English word appear as a bilingual sentence, and are stored in SentJE [Japanese word English word] = {appearance line number list} (step S704).

Next, the degree of word correspondence in the word pair stored in the word pair table is calculated (step S70).
5). The degree of word correspondence is calculated by three expressions (SimJE [JW EW] = 2 * FreqJE [JW EW] /) shown in step S705 in FIG.
(FreqJ [JW] + FreqE [EW]) (called Japanese-English correspondence), SimJ [JWEW] = FreqJE [JWEW] / (FreqJ
[JW]) (called Japanese language support), SimE [JW EW] = Fr
It is calculated from eqJE [JW EW] / (FreqE [EW]) (called the degree of English correspondence) and (JW: Japanese word EW: English word)), and the value is added to the word pair table. However, on condition that the value of the degree of Japanese-English correspondence SimJE [JW EW] is equal to or greater than the threshold value x (step S706), if the threshold value is not satisfied, 0 is added to the values of SimEJ, SimJ, and SimE. Each word correspondence takes a value from 0 to 1, and a word pair having a bilingual relationship has a value closer to 1. The value of the threshold value x can be changed by the user according to his / her preference in the parameter setting unit 313. In this case, when the value is increased, the error rate of the word candidate decreases, but the number of translation patterns that can be obtained decreases. In this specific example, x = 0.1. FIG. 42 shows an example of the word pair table in the specific example 3.

Next, of the Japanese words and English words stored in the index of the word pair table, the highest degree of Japanese-English correspondence (SimJ
E Extract English words that have [Japanese word English word]).
If the Japanese word having the highest degree of Japanese-English correspondence among the English words paired with the English word is the original Japanese word, the word pair is estimated as a bilingual candidate (step S707). , And the word pair to the word pair candidate list (Cand
JE [Japanese word English word]) (step S7)
08). If there are a plurality of word pairs having the highest degree of correspondence between Japanese and English, all the word pairs are stored.

For example, in the case of FIG. 42, the English words forming the index together with the Japanese word “link” are “link”,
Among them, “symbolic”, “owner” and “change”, the English word having the highest Japanese-English word correspondence (SimJE) is “link” (SimJE = 0.75). On the other hand, similarly, the Japanese word having the highest degree of Japanese-English correspondence in the English word “link” is “link”. As a result, both are estimated as bilingual candidates, and a word pair candidate list (CandJE [link link
]) (Step S708). If there is a bilingual candidate extracted here (step S709), co-occurrence word estimation processing by co-occurrence word estimation unit 342b (FIG. 12)
, The translation pattern word order determination process (step S606 in FIG. 12) is performed. Less than,
The co-occurrence word estimation processing (step S605) and the translation pattern word order determination processing (step S606) will be described.

FIG. 16 is a flowchart showing the algorithm of the co-occurrence word estimation processing and the translation pattern word order determination processing. An outline of these processes is as follows. For each word pair extracted in the above-described bilingual word estimation process in FIGS. 13 to 15, a co-occurrence word estimation process and a pattern word order determination process are performed. The pattern creation counter is advanced by one (step S808), and the process returns to the parallel word estimation process (step S702 in FIG. 13). Hereinafter, the details of the processing will be described.

First, a word pair candidate list (CandJE [Japanese word English word]) created by the parallel word estimation process is read (step S801), and a word pair of a Japanese word and an English word is extracted from the word pair candidate list. One is taken out (step S802).

FIG. 43 shows a word pair candidate list. Here, the processing of steps S803 to S807 will be described using the element “link” of the word pair candidate list as an example. First, paying attention to the English word “link”, a Japanese word (word table (SimJE [JW “link”] registered as a word table (SimJE [JW “link”]) as a translation candidate of “link” stored in the word pair table JW (However, SimJE [JW “lin
k "]! = 0)) are all extracted (step S80).
3). In the case of “link”, the Japanese word JW extracted from FIG. 42 is {symbol, temporary, link}.

Further, at step S803, the English word correspondence (SimE [JW English word]) and the appearance line number (SentJE [JW English word]) of the extracted word pair relating to the Japanese word are used as the original word pair ("link"). link ”) equals JW (SimE [JW“ li
nk ”] = SimE [“ link ”“ link ”] and SentJE [JW
“Link”] = SentJE [All JWs that satisfy the condition of “link” “link”] are extracted. Here, {symbol, link} is extracted from FIG. That is, "link l
Ink ”has a SimE of 0.63 and a SentJE of 4 153 154...
link ”and the original word pair“ link link. ”Therefore, {symbol, link} is extracted as the Japanese word.

On the other hand, this time, attention is paid to the Japanese word “link”, and an English word (word table (SimJE [Japanese word EW]) as a translation candidate of “link” stored in the word pair table is used. (Step S804) In the case of "link", the English words extracted from FIG. 42 are @link, owner, change, symboli
c｝.

Further, in step S804, the Japanese word correspondence (SimJ [“link” EW]) and the appearance line number (SentJE [“link” EW]) of the word pair relating to the extracted English word match the original word pair. Equal EW (SimJ [“link” EW] = SimJ
[“Link” “link”] and SentJE [“link” EW] =
All Es that satisfy the conditions of SentJE [“link” “link”]
W) extract. Here, from FIG. 42, ｛link, symboli
c｝ is selected. That is, SimJ of "link" is
0.92, SentJE is 4 153 154..., And the word pair having a value equal thereto is “link symbolic” and the original word pair “link link”. Therefore, {link, symbolic} is extracted as the English word.

Next, translation pattern word order determination processing (steps S805 and S806) is performed. From the word pair table, the value (one of the elements of the list) of the line number (SentJE [Japanese word English word]) of the bilingual sentence in which the Japanese word and English word of the word pair candidate appear (step S805). In step S806, the word order of the Japanese words to be the translation candidates extracted in the co-occurrence word estimation processing is determined with reference to the Japanese translation table (SentJ1 [line number]).
Similarly, the word order of English words that are translation candidates is determined with reference to the English translation table (SentE1 [line number]). If another word string interrupts a word that is a translation candidate, the interrupt word string is represented as “*” as a variable.
In the case of “link link”, the appearance line number 4 is obtained from the word pair table (see FIG. 42), and the “symbolic link” is obtained from the Japanese bilingual sentence table (SentJ1 [4]).
On the other hand, “symbolic link” is obtained from the English bilingual sentence table (SentE1 [4]).

The obtained result is stored in a pattern table (Patt
ern [Japanese pattern English pattern] = {List of occurrence line numbers}). In this example, Pattern
[Symbolic link] = $ 4,153,
154,...｝ (“_” Indicates a space between words in a pattern including a plurality of words). FIG. 44 shows an example of the obtained pattern table.

After performing the above processing for all elements of the word pair candidate list, 1 is added to the pattern creation counter n (step S807), and the process returns to the bilingual word estimation processing of FIG. However, when n is 2 or more, when reading the morphological analysis result into the bilingual sentence table, the pattern table (Pa
ttern [Japanese pattern English pattern])
If the translation pattern stored in the pattern table exists in the bilingual sentence, words constituting the translation pattern are not stored in the bilingual sentence table.

To give an example, if the pattern creation counter is 2 and the pattern table of FIG.
4 "information" and "instruction_give answer" are not stored in the bilingual sentence table. Then, the bilingual word estimation processing and the co-occurrence word estimation processing are repeatedly performed on the newly created bilingual sentence table. When there are no more word pair candidates extracted in the bilingual word estimation processing, the translation pattern generation processing (step in FIG. 15) S711)
Move on to Here, the reason why the repetitive processing is performed is as follows:
In the same way as above, it is to extract word pairs that could not be extracted in the previous bilingual word estimation processing by repeating the processing.

In the translation pattern generation processing in step S711, the pattern table (Pattern [Japanese pattern English pattern]) created in the bilingual word estimation processing is shaped according to the request set by the user through the parameter setting unit 312. For example, when the user gives a request to “output a translation pattern in the form of a Japanese pattern versus an English pattern”, the translation pattern generation unit 343 receives the instruction from the parameter setting unit 312, and receives a pattern table (Pattern [Japan Output the formatted result of the word pattern English pattern]). When a command of “output with part of speech information” is given, the translation pattern generation unit 343
Receives the instruction from the parameter setting unit 312,
Take out one line number of the line number list stored in the pattern table, refer to the bilingual sentence table indicated by the line number, extract the number of the independent word, and in that order The located word is extracted from the morphological analysis result, and the part of speech stored therein is output together.
The user can confirm whether the result is correct or not and correct the result.

<Effects> As described above, Embodiment 3 has the following effects in addition to the effects of Embodiment 1. That is, by extracting a co-occurrence relationship in a single language, it is possible to obtain not only a translated word in a bilingual document but also a translation pattern using a fixed phrase appearing in the bilingual document as a translation unit. .

<< Embodiment 4 >><Structure> FIG. 17 is a block diagram of Embodiment 4 of the translation pattern creating apparatus of the present invention. This specific example is characterized in that a translation pattern can be created using a fixed pattern that appears multiple times in a document as a translation unit, as in the above-described specific example 3. Is limited to consecutive words.

The configuration of the fourth embodiment is similar to that of the third embodiment except that the function of the translation pattern estimating section 344 is different. The translation pattern estimating unit 344 calculates the translation pattern
It has a function of obtaining a word string correspondence degree of an arbitrary word string composed of a plurality of consecutive words, and extracting a word pair serving as a translation pattern based on the word string correspondence degree.
Since the other components are the same as those of the third embodiment, the corresponding portions are denoted by the same reference numerals and description thereof will be omitted.

<Operation> FIG. 18 is a flowchart showing a process in which the apparatus creates and displays a translation pattern from bilingual documents in English and Japanese. Input of bilingual document (step S9)
01), Japanese / English morphological analysis processing (step S90)
2) is the same as in the first embodiment, and a description thereof will not be repeated. Then, after each morphological analysis result is stored in the bilingual document storage unit 332 by the morphological analysis processing in step S902,
The translation pattern estimation unit 344 performs a translation pattern estimation process (step S903). Here, pairs of Japanese and English word strings that are translation pattern candidates are extracted. If it is determined in the next step S904 that there is a pair of word strings serving as translation pattern candidates, the process returns to the translation pattern estimation processing in step S903. Such a translation pattern estimation process is repeated until there are no more pairs of word strings to be extracted. If there is no longer any pair of word strings in step S904, the process proceeds to a translation pattern generation process by the translation pattern generation unit 343 (step S905).
After that, the translation pattern is shaped into the output format specified by the user in the parameter setting unit 313 (step S905),
The result is displayed on the output unit 312 (step S906), and the process ends.

The details of the algorithm after the translation pattern estimation processing will be described below. FIGS. 19 and 20 are flowcharts showing the algorithm of the translation pattern estimation processing (steps S903 and S904) by the translation pattern estimation unit 344.

First, the threshold of the number of appearances of Japanese and English (f
_min ) is set (step S1001). This is set by the user in the parameter setting unit 312 or the present apparatus sets an appropriate value (according to the type of input text, an average of f _min = 10 is appropriate according to experiments).

Next, 1 is set to the pattern creation counter n (step S1002). The pattern creation counter indicates the number of times of the parallel translation word string estimation process to be performed repeatedly, and is used to manage a translation text table (SentJn, SentEn) described later.

Next, a bilingual sentence table storing word strings that appear a plurality of times is created from the standard form results (FIG. 29 (1), FIG. 30 (2)) in the morphological analysis results (step S1003). The creation method is as follows.

The number of appearances of the Japanese independent word and the English independent word appearing in the standard form result in the morphological analysis result is counted, and only the word that appears twice or more is extracted. Next, two consecutive words starting with the word are created, the number of appearances is counted, and headings of the two consecutive words that appear two or more times are extracted.
Further, the same processing is performed on three consecutive words starting with two consecutive words that appear two or more times, and thereafter, the headings of the consecutive words that appear two or more times are extracted while expanding the number of consecutive words. Is stored in the bilingual sentence table (hereinafter, a heading of a continuous word composed of n words extracted here is simply referred to as a word string).

Next, the number of appearances of the word strings appearing in the bilingual sentence table is counted, and the word string table (FreqJ [Japanese word string] =
The number of appearances of the Japanese word string, FreqE [English word string] = the number of appearances of the English word string) are stored (step S100).
4).

Further, the number of appearances of a combination of a Japanese word string and an English word string (word string pair) included in the bilingual sentence is counted, and the number of appearances and the line number of the sentence are identified by using each word string pair as an index. Word string pair table (FreqJE
[Japanese word string English word string] = the number of times a Japanese word string and an English word string appear as a bilingual sentence, and stored in SentJE [Japanese word string English word string] = appearance line number list) (step S1005).

Next, the degree of correspondence (SimJE [Japanese word string English word string]) of the word strings in the word string pairs stored in the word string pair table is calculated (step S100).
6). Here, as an expression for calculating the degree of correspondence between word strings,
The equation (1006a) in FIG. 20 is used. However, the degree of correspondence is
The threshold value f of the number of appearances set in advance in step S1002
_mi provided that satisfies the formula (1006b) by _n.

Next, in the Japanese word string and the English word string stored in the index of the word string pair table,
The English word string having the highest correspondence (SimJE [Japanese word string English word string]) among the English word strings paired with the Japanese word string is extracted. Then, if the Japanese word string having the highest correspondence among the Japanese word strings paired with the English word string is the original Japanese word string, the word string pair is determined as a bilingual pattern ( In step S1007, the word string pair is stored in the bilingual pattern list (step S1008).

As a specific example, first, if the English word string having the highest correspondence to the Japanese word string "Internet address" is "internet address", the "Internet address" and "internet addres
On the other hand, if the English word string having the highest correspondence in the Japanese word string “Internet” is “internet address”, the “Internet” and “Internet” An internet address "pair is extracted. Next, the highest Japanese word string is extracted for the English word string “internet address”. If it is "Internet address", it has already been confirmed that the English word string having the highest correspondence of "Internet address" is "internet address". Is determined as the bilingual pattern and stored in the bilingual pattern list. In other words, at this point, "Internet" and "in
The "ternet address" pair is rejected.

If there is a word string pair stored in the bilingual pattern list (step S1009), 1 is added to the pattern creation counter n (step S1010).
It returns to step S1003 of FIG. However, step S
In the process of 1003, when n is 2 or more, when the word string is stored in the bilingual sentence table, the word string pair stored in the pattern table is referred to by referring to the pattern table ([Japanese pattern English pattern]). If it exists in the bilingual sentence, the word string including the word string is not stored in the bilingual sentence table. Thereafter, the processing is continued as in the case of n = 1.

[0123] On the other hand, if there is no word sequence pairs stored in the bilingual pattern list (Step S1009), it resets the threshold value f _min number of occurrences to a value decremented by one (step S1011).

This resetting means that the number of Japanese and English word strings that are candidates for the translation pattern is gradually increased.
Since the number of appearances is reflected in the accuracy (that is, it can be said that the correspondence is more reliable as the number of occurrences as a pair increases), extracting bilingual patterns in descending order of appearances is a high-quality bilingual pattern. This leads to the extraction of the bilingual pattern in order. Therefore, there is an advantage that the quality of the bilingual pattern and the number of generated translation patterns can be adjusted by changing the condition of z by the user using the parameter setting unit 312.

Next, the threshold value f _min of the number of appearances is checked (step S1012). If the threshold of the number of appearances is f _min
If a value becomes equal to or less than a certain value z (z = 2 in FIG. 20),
The processing shifts to the translation pattern generation processing (that is, z is the minimum condition of the number of appearances. Therefore, as z increases, the accuracy of the extracted bilingual pattern increases, but the number of extractions decreases.).

On the other hand, if the threshold f _{min of the} number of appearances is equal to or greater than a certain value z, the flow returns to step S1002 in FIG.
After the pattern creation counter is initialized to 1, the subsequent processing is repeated.

The subsequent translation pattern generation processing is the same as in the third embodiment. That is, the user sets the parameter setting unit 3
According to the request set in step 13, the translation pattern obtained by the translation pattern estimation processing is shaped and output. Then, the user can confirm whether the result is correct or not and correct the result.

FIG. 45 is an explanatory diagram of the extracted translation pattern. As shown in the figure, as shown in an example of four translation patterns including “Internet”, a plurality of translation patterns related to a certain word can be extracted. Also,
Even within a single bilingual sentence, the translated word differs depending on the word connected after it. Translation patterns such as "trade secret" and "business hours: business hour" can also be accurately extracted.

<Effects> As described above, in this specific example, as in the specific example 3, not only the translation word in the bilingual document but also the translation using the fixed phrase appearing in the bilingual document as the translation unit You can get the pattern.

<< Embodiment 5 >><Configuration> FIG. 21 is a block diagram of Embodiment 5 in the translation pattern creating apparatus. Here, an existing bilingual dictionary is used when creating a translation template in a bilingual sentence format in which some words are replaced with replaceable variables. The feature of this example is that an existing bilingual dictionary and a bilingual document having the same field or characteristics as the document to be translated by the user are given to the system at the same time, so that the existing bilingual dictionary is translated equivalent to the bilingual document. The point is that you can tune to a bilingual dictionary that gives you the results. Further, since a translation pattern including a plurality of words can be acquired, tuning can be performed not only in a word-to-word bilingual dictionary but also in an existing idiom dictionary. That is, in this specific example, the bilingual word estimation processing using the bilingual dictionary in the specific example 2 and the specific example 3
Are performed together with the co-occurrence word estimation processing in.

The configuration of the fifth embodiment includes an input / output unit 410, a resource management unit 430, and a translation pattern extraction unit 440. These configurations are similar to those of the third embodiment.
A bilingual dictionary 434 exists, and the translation pattern extraction unit 44
0 in the translation pattern estimating section 442.
34 is used. Also, the input / output unit 41
At 0, the input unit 411 receives not only a bilingual document but also a bilingual dictionary, and the output unit 412 displays an existing bilingual dictionary tuned by the input bilingual document. The bilingual dictionary may be not only a correspondence in word units but also a translation pattern composed of a plurality of words as described in the third embodiment. The other input / output units 410,
The configurations of the resource management unit 430 and the translation pattern extraction unit 440 are the same as those of the input / output unit 310 and the resource management unit 330 of the third embodiment.
The configuration is the same as that of the translation pattern extraction unit 340.

<Operation> FIG. 22 is a flowchart showing an algorithm for creating a translation pattern according to this example. The operation of this specific example is basically the same as that of the specific example 3. The differences are (1) at the time of input, a bilingual dictionary to be tuned is read at the same time as the bilingual document, (2) a bilingual dictionary is used for the bilingual word estimation processing, and (3) a bilingual dictionary tuning processing in the translation pattern generation processing. Do, three points. The following description focuses on the differences.

In this specific example, the bilingual dictionary (see FIG. 46) (Dic [Japanese pattern English pattern] (value is omitted)) is tuned by the bilingual document (see FIG. 27) used in specific example 1. The case will be described. FIG.
6 is an explanatory diagram of the bilingual dictionary.

In FIG. 22, the input of the bilingual document (step S1101) and the Japanese / English morphological analysis (step S1102) are the same as those in the above-described specific examples, and therefore description thereof will be omitted. In step S1101, the user inputs a bilingual dictionary to be tuned by the input unit 411 simultaneously with the bilingual document. This bilingual dictionary is stored in the bilingual dictionary 434 of the resource management unit 430.

After each morphological analysis result is stored in the bilingual document storage section 432 by the morphological analysis processing (step S1102), the bilingual word estimating section 442a in the translation pattern estimating section 442 performs the bilingual word estimating processing (step S1102).
1103). Here, a bilingual dictionary 434 is used to extract Japanese and English word pairs that are bilingual candidates. If there is a word pair that is a translation candidate (step S1104), the process proceeds to a co-occurrence word estimation process (step S1105) by the co-occurrence word estimation unit 442b, and further performs a translation pattern word order determination process (step S1106).
The process returns to the translated word estimation process of step S1103 again. Such a bilingual word estimation process and a co-occurrence word estimation process are repeated until there are no more word pairs to be extracted, and when there are no more word pairs, the process is transferred to the translation pattern generation unit 443, and the user specifies the word in the parameter setting unit 413. The bilingual dictionary is tuned in the specified format (step S1107), and the output unit 4
The result is displayed in step 12 (step S1108), and the process is terminated.

The details of the algorithm after the parallel word estimation process and an example thereof will be described below. FIG. 23, FIG.
4 and FIG. 25 show the translation word estimation processing (steps S1103 and S11 in FIG. 22 described above) performed by the translation word estimation unit 442a.
It is a flowchart which shows the algorithm of 04).

First, 1 is set to the pattern creation counter n (step S1201). The pattern creation counter indicates the number of parallel translation word estimation processes that are performed repeatedly,
It is used to manage a bilingual sentence table (SentJn, SentEn) described later.

Next, in the standard form results (FIGS. 29, 30 (1), (2)) in the morphological analysis results, the translation patterns stored in the bilingual dictionary (Dic [Japanese pattern English pattern]) Is included in the bilingual sentence, the translation pattern is stored in the pattern table (Pattern [Japanese pattern English pattern] = {list of occurrence line numbers}) (step S1002). From the bilingual dictionary of FIG. 46 and the bilingual sentence table shown in FIG. 31, the following pattern table is created. FIG.
Is an explanatory diagram of the pattern table.

The translation pattern stored in the bilingual dictionary 434 and included in the bilingual sentence is regarded as a more reliable correspondence.
By excluding the translation pattern from the objects to be counted, the value of the word correspondence degree of a word pair having a translation relation other than the translation pattern stored in the bilingual dictionary 434 becomes higher, thereby increasing the value as compared with the case where the bilingual dictionary 434 does not exist. Can also extract many word pairs. Furthermore, this processing means that the translation pattern that appears in the bilingual document and is stored in the bilingual dictionary 434 is detected, and the detected translation pattern is prioritized to be adapted to the bilingual document. I have.

Next, from the standard form results in the morphological analysis results, Japanese words including independent words (name (noun), verb (verb), form (adjective or adjective verb), adverb (adverb) in the candidate part of speech) Word headings of n (noun), v (verb), adj (adjective) and adv (adverb) in words and candidate parts of speech are translated into bilingual sentence tables for each language (SentJn [line number] = {Japan Bilingual sentence word list {, SentEn [line number] = {English bilingual sentence word list} (n: pattern creation counter (n =
1))) is read (step S1203). However, when the morphological analysis result is read into the bilingual sentence table, the pattern table (Pattern [Japanese pattern English pattern]) is referred to, and if the translation pattern stored in the pattern table exists in the bilingual sentence, the translation is performed. The words constituting the pattern are not stored in the bilingual sentence table.

As a specific example, when the pattern table shown in FIG. 47 exists, “start start”, “instruction direction”, and “field field” stored in this pattern table are not stored in the bilingual sentence table. Subsequent processing in the bilingual word estimation processing (step S1204)
To S1210) and co-occurrence word estimation processing (step S121).
1) is similar to the specific example 3, and the description is omitted.

The translation pattern generation processing (step S
1212) is a process of integrating the bilingual dictionary 434 with the translation patterns stored in the pattern table.

FIG. 26 is a flowchart showing an algorithm in the translation pattern generation processing. First, the bilingual dictionary 434 is read (step S1301), and one element (Dic [Japanese pattern English pattern]) is extracted (step S1302). It is checked whether the extracted Japanese pattern or English pattern is registered in the pattern table (step S1303),
Do nothing if not registered. On the other hand, if any one is registered and a pair of a Japanese pattern and an English pattern is registered in the next step S1304,
The other translations are stored in the translation dictionary 434, and the stored translations are deleted from the pattern table (step S13).
05).

If the pair of the Japanese pattern and the English pattern is not registered in step S1304, the storage of the Japanese pattern and the English pattern in the bilingual dictionary 434 (Dic [Japanese pattern English pattern]) is deleted. Then, the other translations are stored in the translation dictionary 434, and the translations are deleted from the pattern table (step S1).
306).

The above processing is performed on all the translation patterns in the bilingual dictionary 434, and finally, all the translation patterns remaining in the pattern table are stored in the bilingual dictionary 434 (step S1307). As a result, the bilingual dictionary 43
As for the translation pattern No. 4, if a translation pattern extracted from the bilingual document exists, the translation pattern is stored. If there is no translation pattern, the translation pattern of the original bilingual dictionary 434 is stored. As a result, the translation word stored in the bilingual dictionary 434 becomes a translation word more suitable for the bilingual document. Finally, the user displays the bilingual dictionary 434 tuned here by the output unit 412. After confirming the correctness, the user can correct the result.

In the specific example 5, the same processing as that of the specific example 3 is used as the translation pattern extraction processing. However, this processing may be performed by the processing of the specific example 4.

<Effects> As described above, Embodiment 5 has the following effects in addition to the effects of Embodiment 2. By extracting the co-occurrence relation in a single language, it is possible to obtain not only a bilingual word in the bilingual document but also a translation pattern using a fixed phrase appearing in the bilingual document as a translation unit. At this time, since an existing bilingual dictionary can be used, a highly accurate translation pattern dictionary that effectively utilizes information of both information (a bilingual document and a bilingual dictionary) can be generated.

In addition to word-to-word bilingual dictionaries, bilingual data relating to idioms and translation fixed patterns can be used for translation pattern extraction. In addition, the data can be tuned to apply to the translation of existing bilingual documents.

In each of the above specific examples, the equation (SimJE [JW EW] = 2 * FreqJE [JW EW]
] / (FreqJ [JW] + FreqE [EW]) (JW: Japanese word, EW: English word) is used, but it is not limited to such a calculation method, and calculation of the degree of correspondence between other words is performed. If the number of occurrences of Japanese words, the number of occurrences of English words, and the number of times these words appear in the same translation are closer to each other, the higher the word correspondence, the higher the word correspondence. Such a calculation method can also be used.

In each of the above specific examples, the source language is Japanese and the target language is English. However, the present invention can be applied not only to these languages but also to other languages. Further, the dictionary obtained from the present apparatus and method has no translation directionality. Therefore, it can be used by both an English-Japanese machine translator and a Japanese-English machine translator.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a specific example 1 of a translation pattern creation device of the present invention.

FIG. 2 is a flowchart of a translation pattern creation process in a specific example 1 of the translation pattern creation device of the present invention.

FIG. 3 is a flowchart (part 1) of a bilingual word estimation process in a specific example 1 of the translation pattern creation device of the present invention.

FIG. 4 is a flowchart (part 2) of a bilingual word estimation process in the specific example 1 of the translation pattern creation device of the present invention.

FIG. 5 is a configuration diagram of a specific example 2 of the translation pattern creation device of the present invention.

FIG. 6 is a flowchart of a translation pattern creation process in a specific example 2 of the translation pattern creation device of the present invention.

FIG. 7 is a flowchart (part 1) of a bilingual word estimation process in a specific example 2 of the translation pattern creation device of the present invention.

FIG. 8 is a flowchart (part 2) of a bilingual word estimation process in the specific example 2 of the translation pattern creation device of the present invention.

FIG. 9 is a flowchart (part 3) of a bilingual word estimation process in the specific example 2 of the translation pattern creation device of the present invention.

FIG. 10 is a flowchart of a bilingual dictionary tuning process in the specific example 2 of the translation pattern creation device of the present invention.

FIG. 11 is a configuration diagram of a specific example 3 of the translation pattern creation device of the present invention.

FIG. 12 is a flowchart of a translation pattern creation process in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 13 is a flowchart (part 1) of a bilingual word estimation process in specific example 3 of the translation pattern creation device of the present invention.

FIG. 14 is a flowchart (part 2) of a bilingual word estimation process in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 15 is a flowchart (part 3) of a bilingual word estimation process in the specific example 3 of the translation pattern creation device of the present invention.

FIG. 16 is a flowchart of co-occurrence word estimation processing in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 17 is a configuration diagram of a specific example 4 of the translation pattern creation device of the present invention.

FIG. 18 is a flowchart of a translation pattern creation process in a specific example 4 of the translation pattern creation device of the present invention.

FIG. 19 is a flowchart (part 1) of a translation pattern estimation process in a specific example 4 of the translation pattern creation device of the present invention.

FIG. 20 is a flowchart (part 2) of a translation pattern estimation process in a specific example 4 of the translation pattern creation device of the present invention.

FIG. 21 is a configuration diagram of a specific example 5 of the translation pattern creation device of the present invention.

FIG. 22 is a flowchart of a translation pattern creation process in a specific example 5 of the translation pattern creation device of the present invention.

FIG. 23 is a flowchart (part 1) of a bilingual word estimation process in a specific example 5 of the translation pattern creation device of the present invention.

FIG. 24 is a flowchart (part 2) of a bilingual word estimation process in a specific example 5 of the translation pattern creating device of the present invention.

FIG. 25 is a flowchart (part 3) of a bilingual word estimation process in the specific example 5 of the translation pattern creation device of the present invention.

FIG. 26 is a flowchart of a bilingual dictionary tuning process in the specific example 5 of the translation pattern creation device of the present invention.

FIG. 27 is a diagram showing an example of a bilingual document for explaining in each specific example of the translation pattern creation device of the present invention.

FIG. 28 is an explanatory diagram of an appearance result of morphological analysis in the bilingual document of FIG. 27;

FIG. 29 is an explanatory diagram (No. 1) of a standard form result of morphological analysis in the bilingual document of FIG. 27;

30 is an explanatory diagram (No. 2) of a standard form result of morphological analysis in the bilingual document of FIG. 27;

FIG. 31 is a diagram showing an example of a bilingual sentence table in the translation pattern creation device of the present invention.

FIG. 32 is a diagram showing an example of a word table in the translation pattern creation device of the present invention.

FIG. 33 is a diagram (part 1) illustrating an example of a word pair table in the translation pattern creation device of the present invention.

FIG. 34 is a diagram (part 2) illustrating one example of a word pair table in the translation pattern creation device of the present invention.

FIG. 35 is a diagram showing an example of a pattern table in the translation pattern creation device of the present invention.

FIG. 36 is a diagram illustrating a translation pattern creation apparatus according to the present invention.
6 is a diagram illustrating an example of a bilingual sentence table in the case of No. 2. FIG.

FIG. 37 is an explanatory diagram (part 1) of a case where elements are simplified in the bilingual word estimation processing of the translation pattern creation device of the present invention.

FIG. 38 is an explanatory diagram (part 2) of a case where elements are simplified in the translation word estimation processing of the translation pattern creation device of the present invention.

FIG. 39 is an explanatory diagram of a display example of a specific example 1 of the translation pattern creation device of the present invention.

FIG. 40 is an explanatory diagram of a bilingual dictionary in a specific example 2 of the translation pattern creation device of the present invention.

FIG. 41 is an explanatory diagram of a pattern table after a bilingual dictionary is consulted in a specific example 2 of the translation pattern creating apparatus of the present invention.

FIG. 42 is a diagram showing an example of a word pair table in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 43 is an explanatory diagram of a word pair candidate list in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 44 is an explanatory diagram showing an example of a pattern table in a specific example 3 of the translation pattern creation device of the present invention.

FIG. 45 is an explanatory diagram showing a translation pattern extraction result in the specific example 4 of the translation pattern creation device of the present invention.

FIG. 46 is an explanatory diagram showing a bilingual dictionary in a specific example 5 of the translation pattern creation device of the present invention.

FIG. 47 is an explanatory diagram of a pattern table after bilingual dictionary lookup in a specific example 5 of the translation pattern creation device of the present invention.

[Description of Signs] 111, 211, 311, 411 Input unit 112, 212, 312, 412 Output unit 122, 222, 342a, 442a Bilingual word estimation unit 234, 434 Bilingual dictionary 342b, 442b Co-occurrence word estimation unit

Continuation of front page (56) References JP-A-5-282361 (JP, A) Smadja, K .; R. McKeown, and V.W. Hatzivas siroglow, Translatating Collations for Bilingual Lexicons: A Statistical Aproach, Computational Linguistics, USA, April 24, 1996, Vol. 22, No. 1, p. 1-p. 38 Mihoko Kitamura (Oki Electric Industry Co., Ltd.), Yuji Matsumoto (Nara Institute of Science and Technology), "Automatic acquisition of translation rules using a bilingual corpus", Transactions of Information Processing Society of Japan, Japan, June 15, 1996 Date, Vol. 37, No. 6, p. 1030-p. 1040 Mihoko Kitamura (Oki Electric Industry Co., Ltd.), Yuji Matsumoto (Nara Institute of Science and Technology), "Automatic acquisition of translation knowledge from bilingual corpus", IEICE Technical Report, Japan, 1994 May 13, Vol. 94, no. 32, p. 9-p. 16 Mihoko Kitamura (Oki Electric Industry Co., Ltd.), Yuji Matsumoto (Nara Institute of Science and Technology), "Automatic Extraction of Dialogue Expression Using Bilingual Corpus", Transactions of Information Processing Society of Japan, Japan, April 15, 1997 Date, Vol. 38, No. 4, p. 727-p. 736 Mihoko Kitamura (Oki Electric Industry Co., Ltd.), Yuji Matsumoto (Nara Institute of Science and Technology), "Automatic Extraction of Dialogue Expression Based on Co-occurrence Frequency in Bilingual Corpus", IEICE Technical Report, Japan, July 18, 1996, Vol. 96, No. 157, p. 69-p. 76 Mihoko Kitamura (Oki Electric Industry Co., Ltd.), Yuji Matsumoto (Nara Institute of Science and Technology), “Automatic Extraction of Dialogue Expression Based on Co-occurrence Frequency in Bilingual Corpus,” Information Processing Society of Japan, Japan, July 19, 1996, Vol. 96, No. 65, p. 69-p. 76 (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 17/21-17/28

Claims (10)

(57) [Claims] 1. A translation pattern creation device stores a bilingual document stored in a bilingual sentence table in which a source language document and a target language document, which is a bilingual sentence of the document, are associated with each other on a sentence basis. The number of appearances of the translated word pair in the bilingual document, the number of appearances of the source language word in the word pair in the source language document, and the number of occurrences of the target language word in the word pair in the target language document. The word pair having the highest value of the correspondence is extracted as a translation word based on the word pair, and the extracted word pair is removed from the bilingual sentence table and the translated document is re-translated. Forming and re-determining the degree of correspondence for the remaining word pairs, extracting the word pair having the highest value of the degree of correspondence as a bilingual word, repeating these, and sequentially extracting bilingual words. Translation pattern generation method for the butterflies. 2. The translation pattern creation method according to claim 1, wherein the extracted word pair is an original word pair by the translation pattern creation device, and one of the original word pairs is included in the original word pair. Looking at the other word from the other word of the original word pair, the same sentence appears, and there is another word whose number of appearances in the pair with the one word is the same, and A translation pattern creation method for extracting a word from the other word of the original word pair and extracting the word as a multiple word-to-multiword translation pattern co-occurring with the original word pair when there is a similar word. . 3. A translation pattern device stores a bilingual document stored in a bilingual sentence table and forming a source language document and a target language document which is a bilingual sentence of the document in correspondence with each other on a sentence basis. For a word string pair consisting of a translated word and a continuous word in a bilingual document, the number of appearances of the word string pair, the number of appearances of the source language word string in the word string pair in the source language document, and the word string Based on the number of appearances of the target language word string in the pair in the target language document, the degree of correspondence of the word string pair is obtained, and the word string pair having the highest value of the degree of correspondence is extracted as a translation pattern. A bilingual document is formed again by removing the extracted word string pair from the bilingual sentence table, and the degree of correspondence is again determined for the remaining word string pair, and the word string pair having the highest value of the degree of correspondence is obtained. Translate patter Extracted as down, by repeating these, sequentially, a translation pattern forming method and extracting the translation pattern. 4. An input section for inputting a source language document and a source document in which a target language document which is a translation of the source document is associated with each sentence, and a source language document for the source document. The number of occurrences of the specific word in the document, the number of occurrences of the specific word in the target language document, and the number of times that the specific word appears in the same translation of the source language and target language documents are close to each other Is obtained, the word correspondence with a higher value is obtained, and the word pair having the highest value of the word correspondence is extracted as a bilingual word, and an arbitrary word pair is extracted as a bilingual word.
At this time, a new bilingual sentence excluding the word pair is generated from the target bilingual sentence, a word correspondence is calculated for the new bilingual sentence, and the word pair having the highest value of the word correspondence is determined. A translation characterized by comprising: a bilingual word estimating unit that sequentially extracts bilingual words by extracting the bilingual words and repeating the process, and an output unit that outputs the bilingual words extracted by the bilingual word estimating unit. Pattern creation device. 5. The translation pattern creating apparatus according to claim 4, wherein a bilingual dictionary having a bilingual translation between the source language and the target language and a word pair appearing in the bilingual document are registered in the bilingual dictionary in advance. Is provided with a bilingual word estimating unit that generates a bilingual sentence excluding the word pair, obtains a word correspondence degree with respect to the bilingual sentence, and extracts a word pair having the highest value of the word correspondence degree as a bilingual word. A translation pattern creating apparatus, characterized in that: 6. A translation pattern creating apparatus according to claim 5, wherein the pattern table storing the word pairs extracted as the bilingual translation, an arbitrary word of the source language in the bilingual dictionary, and a translation of the target language for the word. It is determined whether or not the word pair matches the word pair in the pattern table. If the word pair does not match, a translation word for the arbitrary word is deleted in the bilingual dictionary, and a translation word estimation unit that registers the word pair in the pattern table as a translation word is provided. A translation pattern creation device, comprising: 7. The translation pattern forming apparatus according to claim 4, the word pairs extracted by the translation word estimator and the original word pair, among the original word pair, whereas the other side word from side word See, the other word of the original word pair, the same sentence appears, and there is another word that has the same number of appearances in the pair with the one word, and the original word pair A co-occurrence word estimator is provided that extracts a word as a translation pattern of a plurality of words that co-occurs with the original word pair when there is a similar word when viewed from the other word. A translation pattern creating apparatus characterized by the following. 8. The translation pattern creating apparatus according to claim 7, wherein a bilingual dictionary having a bilingual translation between the source language and the target language and a word pair appearing in the bilingual document are registered in the bilingual dictionary in advance. Is provided with a bilingual word estimating unit that generates a bilingual sentence excluding the word pair, obtains a word correspondence degree with respect to the bilingual sentence, and extracts a word pair having the highest value of the word correspondence degree as a bilingual word. A translation pattern creating apparatus, characterized in that: 9. A translation pattern creating apparatus according to claim 8, wherein a pattern table for storing the word pairs extracted as the parallel translation, an arbitrary word in the source language in the bilingual dictionary, and a translation in the target language for the word. It is determined whether or not the word pair matches the word pair in the pattern table. If the word pair does not match, a translation word for the arbitrary word is deleted in the bilingual dictionary, and a translation word estimation unit that registers the word pair in the pattern table as a translation word is provided. A translation pattern creation device, comprising: 10. An input unit for inputting a bilingual document formed by associating a source language document and a target language document, which is a bilingual sentence of the document, with each other on a sentence basis, and inputting a bilingual document in the bilingual document. For a word string pair consisting of continuous words, the number of appearances of the word string pair, the number of appearances of the source language word string in the word string pair in the source language document, and the target language word string in the word string pair The degree of correspondence of the word string pair is determined based on the number of appearances in the target language document, and the word string pair having the highest degree of correspondence is extracted as a translation pattern, and the extracted word string pair By again forming a bilingual document, obtaining the degree of correspondence again for the word string pair, extracting the word string pair having the highest value of the degree of correspondence as a translation pattern, and repeatedly performing these, Sequentially, translation putter A translation pattern estimating unit for extracting a translation pattern.