JP2000250914A - Machine translation method and device and recording medium recording machine translation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation method/device which can suppress the scale of a co-occurrence information data base and also to obtain a most likelihood translation equivalent string against an original language word string without using an analyzed corpus. SOLUTION: The word pairs occurring concurrently in a decided range of a sentence of an object language corpus and the co-occurrence frequency information on these word pairs are stored in an object language co-occurrence information data base 300. A translation equivalent candidate generation part 220 inputs an original language word string to be translated and retrieves the translation equivalent candidates of an object language of every original language word included in the original language word string by means of a translation dictionary 410. A co-occurrence intensity detection part 230 retrieves the relative words and near-synonyms of every object language translation equivalent by means of the dictionary 410 or a semantic category dictionary 420 and calculates the co-occurrence intensity of a translation equivalent candidate pair against the original language word by means of the data base 300. A translation equivalent decision part 240 selects an object language translation equivalent against the original language word according to the co-occurrence intensity.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a machine translation method and apparatus for selecting a translation word using word co-occurrence information of a target language.

[0002]

2. Description of the Related Art In a rule-based machine translation system using a dictionary or a rule, a process of generating a translated sentence after analyzing an input sentence is performed by a bilingual dictionary or a conversion rule. As a method of selecting an appropriate translation word at the stage of generating this translation, for example, a method of determining the semantic category of the source language word and translating using a conversion rule at the semantic category level (Shirai, Yokoo, Uchino, Matsuo, "Japanese-English Conversion Technology and Semantic Dictionary", NTT R & D, Vol. 46, pp.1405-14
10, 1997) or a method of searching for synonyms of source language words and translating them by increasing the applicability of conversion rules (for example,
"Machine translation apparatus" is disclosed in Japanese Patent Application Laid-Open No. 5-158970.

Further, as an improved method for selecting a more appropriate translation, there is a method using statistical knowledge acquired from a corpus. As a method for selecting a translated word based on statistical knowledge, a method using a bilingual corpus and a translated word selection method using word co-occurrence information on the target language side have been proposed.

As a method of using a bilingual corpus, for example, word sense polysemy elimination (translation term) using statistical knowledge of the meaning of a word to be translated and a co-occurrence word of a source language obtained from a bilingual corpus with word correspondence is used. Selection) methods (P. Brown, S. Della Pietra, V. Della Pietra, R. Me
rcer, “Word-sense disambiguation using statistical
methods ”, Proceedings of Annual Meeting of the
Association for Computational Linguistics, pp. 264
-270, 1991).

[0005] As an example of a translation word selection method using word co-occurrence information on the target language side, the frequency of co-occurrence words having a dependency relationship in the target language corpus is collected, and using these, the dependency relationship between two sets of translation words is obtained. With the intensity of
A method of translating multi-translation verbs has been proposed (Nomiyama, "Translation word selection using knowledge of target language and its learning", Information Processing Society of Japan, NL86-8, 199
1).

After collecting frequency information of co-occurring words in the target language corpus, the selection criterion value is based on the ratio of the highest co-occurrence probability of the combination of translated words to the second highest co-occurrence probability. There has been proposed a method of selecting a translation word set whose selection reference value exceeds a statistically significant threshold value.
(Ido Dagan, Alon Itai, “Word sense disambiguatio
n using a second language monolingual corpus ”, Co
mputational Linguistics, Vol. 20, No. 4, pp. 563-59
6, 1994).

[0007]

The translation word selection method in the conventional rail-based machine translation system has the following problems.

When selecting a translation of a certain source language word, the source language word is obtained by using a rule and a dictionary to obtain a target language translation candidate, and is determined according to the priority of the translation word selection described in the dictionary and the constraint by the conversion rule. , A certain translation is selected. After that, the selected translated words are arranged side by side with respect to a plurality of input words and synthesized (element synthesis). The translation result obtained in this manner does not sufficiently consider the appropriateness of generating a translated word, and thus has a problem that unnatural translated words are likely to be arranged as a target language.

[0009] Further, there are cases where appropriate translations are not registered in the translation dictionary of the machine translation system, and the choices of target language translations are not sufficient.

[0010] Furthermore, in the translation word selection method based on the corpus, even when a bilingual corpus with word correspondence or sentence correspondence or a target language corpus is used, a corpus having inter-word dependency information accurately added manually is used. It is more difficult to work with than a corpus without dependency information. Therefore, there is a problem that it is difficult to obtain statistical knowledge comprehensively. Further, when parsing a corpus of a target language, there is a problem that erroneous dependency information due to analysis failure affects statistical knowledge obtained therefrom as erroneous information. Therefore, it is desirable to consider the use of a target language corpus without dependency information. In addition, as a property of natural language processing using a corpus, when co-occurrence information for word expressions appearing in a corpus is collected, co-occurrence phenomena of words are diluted, and the collected co-occurrence frequency information has a small value. More things to take. As a result, the reliability of the individual co-occurrence frequency information decreases, and when selecting a translation word using the co-occurrence information database constructed in this way, there is a problem that errors are likely to be mixed in the translation word candidates. Further, there is a problem that restrictions on target language translation word candidates due to rules in the source language are not taken into consideration, and a problem that co-occurrence strength of source language words is not taken into account.

SUMMARY OF THE INVENTION An object of the present invention is to solve a problem of unnatural translated word sequence output of translated word selection processing and sparse co-occurrence phenomenon in a corpus without using a corpus to which inter-word dependency information is added. It is an object of the present invention to provide a translation method, an apparatus, and a recording medium on which a machine translation program is recorded.

[0012]

A machine translation method according to the present invention uses a translation dictionary that holds a set of bilingual relations between a source language word and a target language translation word candidate, and converts each source language word in a source language word string to be translated. A translation word candidate generation step of searching for a target language translation word candidate of a language word, collating the target language translation word candidate with an entry in the co-occurrence information database, and calculating the co-occurrence strength of the target language translation word candidate pair for the source language word A co-occurrence strength detection step, and a translation word determination step of selecting a target language translation string for the source language word using the co-occurrence strength.

According to the method of constructing a co-occurrence information database of the present invention, in a sentence in a target language corpus composed of a set of sentences in a target language, a word pair and a co-occurrence frequency information of a word pair appearing simultaneously within a predetermined range are determined. Store in database.

Further, the machine translation apparatus of the present invention provides a translation dictionary that holds a set of bilingual relationships between source language words and target language translation word candidates, and uses the translation dictionary to translate each source language word in the source language word string to be translated. A translation word candidate generating means for searching for a target language translation word candidate of a language word, collating the target language translation word candidate with an entry in the co-occurrence information database, and calculating a co-occurrence strength of the target language translation word candidate pair for the source language word A co-occurrence strength detection means for performing the translation and a target language translation word string for the source language word using the co-occurrence strength.

Further, the co-occurrence information database construction apparatus of the present invention comprises a target language input means for inputting a target language corpus consisting of a set of sentences in the target language, and a sentence in the target language corpus simultaneously within a predetermined range. There is a co-occurrence information extracting means for storing the appearing word pair and its co-occurrence frequency information in a co-occurrence information database.

In the machine translation apparatus of the present invention, the translation word candidate generating means uses the translation dictionary holding a set of bilingual relations between the source language word and the target language translation word candidate, and inputs the input source language word string to be translated. The target language translation word candidate of each source language word in it is searched. The co-occurrence strength detecting means calculates the co-occurrence strength of the target language translation word candidate pair with respect to the source language word using a co-occurrence information database holding a set of entries consisting of word pairs of the target language and its co-occurrence frequency information. . The translation word determining means selects the combination of the target language translation word candidates having the highest co-occurrence strength, and selects the target language translation word sequence for the source language word.

This makes it possible to obtain a maximum likelihood translated word sequence that is likely to appear as a target language expression with respect to the source language word sequence, and to select a natural translated word sequence as the target language, which is the object of the present invention. Will be able to

The target language input means inputs a target language corpus consisting of a set of sentences in the target language, and the co-occurrence information pouring means inputs a corpus of the target language, and outputs a corpus of the target language corpus. Word pairs that appear simultaneously and their co-occurrence frequency information are stored in the co-occurrence information database. Thereby, the target language co-occurrence information database can be created from the target language corpus, and the cost and labor for preparing the co-occurrence information database can be reduced.

The target language co-occurrence information database is
A set of entries is held from the part-of-speech-tagged word pair of the target language and its co-occurrence frequency information, and the co-occurrence intensity detecting means uses the part of speech information by using the co-occurrence information database including the part of speech information. The co-occurrence strength of the target language translation word candidate pair with respect to the source language word is calculated. At this time, when the co-occurrence information database is obtained from the target language corpus, a target language corpus consisting of a set of sentences to which each part of the word is given the part of speech information of the word is input. A part-of-speech tagged language pair that appears simultaneously in the range and its co-occurrence frequency information are stored in the co-occurrence information database.
Next, the co-occurrence intensity detecting means calculates the co-occurrence intensity of the target language translation word candidate pair for the source language word using the part of speech information, using the co-occurrence information database including the part of speech information. As a result, it is possible to select the part of speech of the target language word to be collected, and to accurately compare the difference in the part of speech with the target language word in the target language translation word candidate.

Further, the co-occurrence information extracting means, when collecting a word pair with a part-of-speech tag, counts words (co-occurrence words) co-occurring with a certain key word (key word) for each part of speech, For each co-occurrence word, a co-occurrence rank for each part of speech that indicates the order of the part of speech is extracted. Store in database. The co-occurrence intensity detecting means can calculate the co-occurrence intensity in consideration of the part-of-speech co-occurrence order and the co-occurrence frequency information. This makes it possible to avoid noise such as parsing failure that is mixed when obtaining the target language corpus describing the dependency, and to collect co-occurrence information even if the target language corpus does not include the dependency. In addition, word pairs having no dependency can be excluded to some extent, and effective co-occurrence information can be obtained with a small amount of data.

Further, after inputting the field information in the field input means, the field corpus extraction means extracts a set of sentences related to the field information from the target language corpus, and the co-occurrence information extraction means outputs the field corpus input means. Collect the co-occurrence frequency information of the target language in the sentence extracted in. As a result, it is possible to tune the target language translation string of the final output result into a translation for a specified field.

The co-occurrence strength detecting means refers to the translation dictionary and the semantic category dictionary when matching each target language translation word candidate with the target language word pair in the entry of the co-occurrence information database. The co-occurrence strength is calculated by collating not only with each other but also with the target language word related words and synonyms of the target language word. As a result, words having the same meaning can be obtained irrespective of the notation, and it is possible to avoid the sparseness of co-occurrence of words, which is the object of the present invention.

Further, the translation dictionary includes rule constraint information for selecting a translated word by a language conversion rule, and the translated word candidate generating means includes a means for referring to the translated dictionary and adding rule constraint information to the target language translated word candidate. In this case, the translation determining means
It is possible to select a target language translation candidate that does not contradict the rule constraint assigned to the target language translation candidate. This makes it possible to eliminate unnecessary target language translation word candidates due to the source language rule constraint, which is the object of the present invention, thereby increasing the possibility of selecting an appropriate translation word.

The translation word candidate generating means includes means for searching for a target language translation word candidate of each source language word in the source language word string by referring to another translation dictionary in addition to the translation dictionary possessed by the machine translation system. Including. This makes it possible to increase the number of target language translation word candidates and increase the possibility of selecting an appropriate translation word.

Also, using a source language co-occurrence information database that holds a set of entries consisting of source language word pairs and their co-occurrence frequency information, based on the co-occurrence strength of the source language words in the source language. , It is possible to weight the corresponding candidate word co-occurrence strength. As a result, it is possible to select a translation word with emphasis on the co-occurrence relationship of translation word candidates of words that are likely to co-occur in the source language, and it is possible to further improve the translation word selection accuracy.

[0026]

Next, embodiments of the present invention will be described with reference to the drawings. In the embodiment described below, the source language is Japanese and the target language is English.

FIG. 1 is a basic block diagram showing the configuration of a machine translation apparatus according to an embodiment of the present invention, and FIG. 2 is a flowchart showing the processing. This machine translation apparatus is composed of a co-occurrence information database construction unit 100, a co-occurrence use translation word selection unit 200, a target language co-occurrence information database 300, a dictionary 400, and a source language co-occurrence information database 500.

The co-occurrence information database construction unit 100 includes a field input unit 110, a field corpus extraction unit 120, a target language input unit 130, and a co-occurrence information extraction unit 140. Field information is input in the field input section 110 (step 61).
0), the field corpus extraction unit 120 extracts a set of sentences related to the field information from the target language corpus consisting of a set of statements in the target language (step 620). The target language input unit 130 inputs the target language corpus (step 63).
0), the co-occurrence information extracting unit 140 accumulates, in the target language co-occurrence information database 300, word pairs and their co-occurrence frequency information that simultaneously appear in a predetermined range in the sentences of the target language corpus (step 640).

The dictionary 400 includes a translation dictionary 410 and a semantic category dictionary 420. The translation dictionary 410 holds a set of bilingual relationships between source language words and target language translation word candidates. The semantic category dictionary 420 is composed of a set of correspondences between words of the object and semantic categories representing the meanings of the words.

The co-occurrence use translation word selection section 200 includes a source language word input section 210, a translation word candidate generation section 220, and a co-occurrence strength detection section 2.
30 and a translation word determination unit 240. The translated word candidate generation unit 220 inputs a source language word string to be translated (step 650), and searches for a target language translated word candidate of each source language word in the source language word string using the translation dictionary 410 (step 660). ). The co-occurrence intensity detection unit 230
0 or using the semantic category dictionary 420 to search for related words and synonyms of each target language translation,
Then, the co-occurrence strength of the candidate word pair for the source language word is calculated using 00 (step 670). The translation word determination unit 240
A target language translation string for the source language word is selected using the co-occurrence strength (step 680).

Next, the procedure for selecting a machine-translated word in this embodiment will be described. Here, assuming that a part of the input sentence of the machine translation system is "a leading market research institution", the result of the syntax analysis results in a dependency relationship as shown in FIG. The source language word string having this dependency will be described below as an input example.

First, the target language input unit 130 inputs a target language corpus. When text is input from a one-line corpus, the co-occurrence information extraction unit 140 scans all co-occurring word pairs within a predetermined range (for example, within one read line), and searches for the entry in the co-occurrence information database 300. Are increased by one each. If the corresponding entry does not exist, it is newly added as a frequency 1 entry. Repeat this operation.

FIG. 4 shows an example of a target language corpus with a morpheme tag. Each word is composed of a “word / speech tag”. For example, "market / NN" is the singular noun "mark
et ”. The words that co-occur with the first word“ trading / NN ”are (“ trading / NN ”,“ stock-index / NN ”),
("Trading / NN", "futures / NNS"), ("trading / NN
NN ”,“ first / JJ ”), etc. By using morpheme information obtained from a tagged corpus, the co-occurrence strength detection described later can be compared with the co-occurrence information of written words without morpheme tags. In the section 230, the target language word in the target language translation word candidate can be more accurately collated.
Is a content example of the co-occurrence information database 300 constructed by this operation.

[0034]

[Table 1]

When inputting the target language corpus,
The co-occurrence information database 300 can be constructed by selecting the field of the target language corpus to be analyzed. First, field information is input by the field input unit 110. For example, a keyword such as "machine translation" is input. Next, the field corpus extraction unit 120 extracts a set of sentences related to the field information from the target language corpus. For example, by using search software that can extract text including a given keyword from a large amount of text data, a set of sentences related to a target field can be obtained. The co-occurrence information extraction unit 140 constructs the co-occurrence information database 300 using the sentence set as an input. When the co-occurrence information database 300 is used,
It is possible to tune the target language translation word string output in step (1) to one suitable for the field.

The operation up to this point is performed in the co-occurrence use translation word selection unit 2
The co-occurrence information database 300 to be used at 00
Is a pre-process performed before executing the machine translation system.

Next, the source language word input section 210 inputs a source language word string to be translated, which is extracted from the translation process of the machine translation system. FIG. 3 shows a source language word string having a dependency.

The translation word candidate generator 220 searches the translation dictionary 400 for each source language word in the source language word string, and obtains a target language translation word candidate consisting of one or more target language words. The obtained target language translation word candidate set is stored in the bilingual list. As the translation of each source language word in FIG. 3, for example, target language translation word candidates shown in Table 2 are obtained.

[0039]

[Table 2]

At this time, by referring to another translation dictionary in addition to the translation dictionary possessed by the machine translation system, the target language translation word candidates can be increased. For each source language word in FIG. 3, a bilingual translation (order a, b
Table 3 shows the result of the addition of the word ()).

[0041]

[Table 3]

The co-occurrence intensity detector 230 extracts the co-occurrence intensity using the translated word candidate list obtained by the above processing. FIG. 5 is a flowchart showing the processing of the co-occurrence intensity detecting section 230. In the following, a target language translation candidate pair for a word pair having a dependency relationship among the source language word strings will be described. (Step 701) First, a translation word pair included in one target language translation word candidate pair is selected. (Step 702) Next, for each target word of the selected target word pair, reference is made to the translation dictionary 410 which can take into account the target language word related words that are variations, derivatives, and synonyms. Search for. Table 2 expands the related word combination and selects one related word pair of the combination.
Tables 4 and 5 show a set of related words of each translation of the target language translation word candidate.

[0043]

[Table 4]

[0044]

[Table 5]

When the combination of related words has been verified, combinations of synonyms having the same semantic category as the translated word can be considered. That is, a synonym is searched using the semantic category dictionary 600 composed of a set of correspondences between words in the target language and semantics representing the meaning of the word, and combinations of the synonyms are also expanded. Choose one synonym pair. Table 6 shows examples of synonyms having the same semantic category of each translated word in Table 2.

[0046]

[Table 6]

By considering the related words of the two target language translation word candidates, the problem of the sparseness of the word co-occurrence phenomenon due to the use of the target language corpus can be solved. (Step 703) The co-occurrence information database 300 described above
, And collate with the entry of the selected related word pair, and if it exists, obtain the co-occurrence frequency. (Step 704) The obtained co-occurrence frequency is added to the frequency of the currently selected translated word pair. Here, weighting can also be performed by multiplying the frequency by a certain coefficient according to a related word or a synonym. (Step 705) If all the related words and synonyms included in the translation word pair have been checked, the process proceeds to Step 706. If not, the process proceeds to step 701. Table 7 shows an example of a translated word pair frequency list in which the translated word pair frequencies are arranged.

[0048]

[Table 7] (Step 706) If all the translated word pairs included in the target language translated word candidate pair have been checked, the process proceeds to Step 707. If not, the process proceeds to step 701. (Step 707) One target language translation word candidate pair is selected. (Step 708) The co-occurrence strength for the target language translation word candidate pair is calculated using the frequency information of all the translation word pairs included in the target language translation word candidate pair, and registered in the translation word candidate co-occurrence strength list.

An example of an approximation method for calculating the co-occurrence strength of the target language translation word candidates from the frequency information of the co-occurrence word pairs will be described below. Calculate the co-occurrence probability for each translation word pair included in the target language translation word candidate pair. The translation word candidate co-occurrence strength is set to the average value of the co-occurrence probabilities of all the translation word pairs, considering combinations of translation word pairs included in the target language translation word candidates. Instead of the co-occurrence probability, other statistics such as mutual information may be used. Table 8
FIG. 9 shows a translation word candidate co-occurrence strength list obtained by the above processing.

[0050]

[Table 8] (Step 709) If the co-occurrence strength has been calculated for all target language translation word candidate values, the processing of the co-occurrence strength detection section 230 ends. If not, the process proceeds to step 7
Shift to 07.

Next, the translated word determination unit 240 selects a final translated word sequence using the translated word candidate co-occurrence strength list. FIG. 6 is a flowchart showing the processing of the translated word determination unit 240. (Step 711) First, if there is an entry of the translated word candidate co-occurrence strength remaining in the translated word candidate co-occurrence strength list, the process proceeds to step 712. If the list is empty, the process of the translation word determination unit 240 is terminated, and the correspondence of the target language translation word candidate of the source language word is set as the final output. (Step 712) An entry having the highest value of the co-occurrence strength is searched for and acquired from the list of the translation word candidate co-occurrence strengths. (Step 713) The target language translation candidate of the selected entry is determined as the target language translation candidate of the corresponding source language word, and the correspondence between the source language word and the target language translation candidate is stored. (Step 714) The selected entry is deleted from the translation word candidate co-occurrence strength list. In addition, among the target language translation candidates of the source language word for which the target language translation candidate is determined, all entries including the target language translation candidate other than the target language translation candidate selected in step 713 are deleted from the translation word candidate co-occurrence strength list. I do. Thereafter, the process proceeds to step 711.

Here, when selecting target language translation word candidates based on the above procedure, first, (“market research”,
The “organization” pair is selected and then (“powerfu
l ”,“ organization ”) is selected, the translation word co-occurrence strength list becomes empty, and the process of the translation word determination unit 240 ends. The final correspondence between the source language word and the target language translation word candidate is as shown in FIG. Become.

In an actual machine translation system, using the final output of the present invention, the inflection and the order of each word are adjusted, and a translated sentence of the target language is generated.

Further, the co-occurrence information extracting unit 230 can exclude a word pair having a certain degree of dependency with the following ingenuity when collecting a word pair with a part-of-speech tag.

If FIG. 4 is an example of the target language corpus, the first word “trading / NN” is used as a key word (key word).
Then, words that co-occur with this (co-occurrence words) are tabulated for each part of speech, and a part-of-speech co-occurrence ranking indicating the order of each part of speech is extracted. Table 9 shows an example of the counting result.

[0056]

[Table 9]

Next, co-occurrence information is stored in the co-occurrence information database 300 for each pair of a key word and a co-occurrence word and for each co-occurrence order for each part of speech. This operation is performed for all the words in the input target language corpus, and a co-occurrence information 300 database is constructed. Table 10 shows an example of the contents of the co-occurrence information database 300 constructed as described above.

[0058]

[Table 10]

Further, the co-occurrence intensity detecting section 230 calculates the co-occurrence intensity in consideration of the part-of-speech-based co-occurrence order and the co-occurrence frequency information. For example, the co-occurrence strength can be calculated using only the first co-occurrence frequency information in the sentence end direction in the co-occurrence order.

Thus, the co-occurrence information database 30
Although the resource for constructing 0 is a target language corpus having no dependency, a word pair having no dependency can be excluded to some extent, and an effective co-occurrence relationship can be obtained even with a small amount of data.

Also, the input sentence to the machine translation system is
If “the leading market research institution expects...”, And the rule constraint information in the translation dictionary 410 is searched and matched, it can be determined that the conversion rule of FIG. 7 can be applied. In the conversion rules shown in FIG. 7, “(subject)” indicates the semantic category of the source language, and indicates that a source language word that can take the meaning of “subject” is applicable. The translated word candidate generator 220 assigns this rule constraint information to the target language translated word candidate. In other words, the translation of the word “institution” in the source language word string is restricted by the rules, and “organizatio
It is found that only the translation candidate of “n” can take the meaning of “subject”, and information of the constraint “priority” is added to “organization.” The result of adding the rule constraint information is shown in Table 11.

[0062]

[Table 11]

Next, the translation word determination unit 240 performs a process of preferentially adopting a target language translation word candidate with a rule constraint. As a result, "or"
ganization "is preferentially adopted. This makes it possible to remove unnecessary target language candidate candidates at an early stage.

An example of a method for selecting a priority translation word using a source language co-occurrence information obtained from the source language co-occurrence information database 500 while holding a set of entries including co-occurrence frequency information in the source language will be described below. . Table 12 shows an example of the source language co-occurrence information database 500.

[0065]

[Table 12]

The co-occurrence strength of the source language word pair in the source language is defined as the ratio of the co-occurrence frequency of the source language word pair to the sum of the co-occurrence strengths of the word combinations in the source language word string. The co-occurrence strength detection unit 230 sets each translated word candidate co-occurrence strength as the product of the translated word candidate co-occurrence strength calculated by the co-occurrence strength detection unit 230 and the co-occurrence strength of the corresponding source language word pair in the source language.
This is again stored in the translation word candidate co-occurrence strength list, and the translation word determination unit 240 selects the final translation word using this. Table 8
Table 13 shows the result of changing the translation word candidate co-occurrence strength list indicated by the above in accordance with the above procedure.

[0067]

[Table 13] FIG. 9 is a block diagram of a machine translation device according to another embodiment of the present invention.

The machine translation apparatus of the present embodiment has an input device 80
1, a storage device 802 to 805, an output device 806, recording media 807 and 808, and a data processing device 809.

The input device 801 is an input device such as a scanner or a keyboard for inputting a target language corpus and a source language word string. Storage devices 802, 803, 80
4 is a target language co-occurrence information database 3 in FIG.
The storage device 805 corresponding to the language co-occurrence information database 500 and the dictionary 400 is a hard disk. The output device 806 is a display, a printer, or the like to which a target language translated word string is output. Recording media 807 and 808 are FD (floppy disk), CD-ROM, MO
A co-occurrence database constructing program including the processes of the co-occurrence database constructing unit 100 shown in FIG. The selection program is recorded. The data processing device 809 stores the recording medium 80
The CPU reads a co-occurrence database construction program and a co-occurrence use translation word selection program from the storage device 805 and 7,808, respectively, and executes them.

[0070]

As described above, according to the present invention,
A word pair obtained by analyzing the corpus of the target language and its co-occurrence frequency information are stored in a co-occurrence information database, and the target language target word candidate pair for the source language word searched by the dictionary using the co-occurrence information database. After calculating the co-occurrence strength of the source language word, by selecting an appropriate target language translation word sequence for the source language word, the maximum likelihood translation word sequence for the source language word sequence can be obtained. By specifying the field of the target language corpus to be analyzed, it is possible to make the target language translation word candidates to be output suitable for the field.

Further, a variation to the target language translation word candidate,
Consideration of derivatives, synonyms, and synonyms of the same semantic category can solve the problem of sparseness of co-occurrence.

Further, the number of target language translation word candidates is increased by using a plurality of translation dictionaries, unnecessary target language translation word candidates are removed by applying rule constraints on the source language, and the source language word co-occurrence information is used. The accuracy of selecting a suitable translation can be improved.

When extracting co-occurrence information, a word co-occurring with a key word and a co-occurrence order for each part of speech are extracted and registered in the co-occurrence information database together with co-occurrence frequency information. It is possible to calculate the co-occurrence intensity by taking into account the co-occurrence rank and co-occurrence frequency information for each part of speech, so that when collecting co-occurrence information, it is possible to avoid parsing failure noise and to include dependency. In spite of the target language corpus which is not available, it is possible to exclude a word pair having no dependency to some extent.

As described above, the maximum likelihood translated word sequence for the source language word sequence can be obtained without using the analyzed corpus and suppressing the size of the co-occurrence information database.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a machine translation device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing an entire process of the machine translation device of FIG. 1;

FIG. 3 is a diagram showing an example of a source language word string having a dependency.

FIG. 4 is a diagram showing an example of a target language corpus.

FIG. 5 is a flowchart illustrating a process of a co-occurrence intensity detection unit 230.

FIG. 6 is a flowchart showing processing of a translation word determining unit 240.

FIG. 7 is a diagram illustrating an example of a conversion rule.

FIG. 8 is a diagram showing the correspondence between source language words and target language translation word candidates.

FIG. 9 is a configuration diagram of a machine translation device according to another embodiment of the present invention.

[Explanation of symbols]

REFERENCE SIGNS LIST 100 co-occurrence database construction unit 110 field input unit 120 field corpus extraction unit 130 target language input unit 140 co-occurrence information extraction unit 200 co-occurrence use translation word selection unit 210 source language word input unit 220 translation word candidate generation unit 230 co-occurrence strength detection unit 240 translation word determination unit 300 target language co-occurrence information database (co-occurrence information database) 400 dictionary 410 translation dictionary 420 semantic category dictionary 500 source language co-occurrence information database 610-680, 701-709, 711-714
Step 801 Input device 802 to 805 Storage device 806 Output device 807, 808 Recording medium 809 Data processing device

Claims (43)

[Claims] 1. A target language co-occurrence information for storing, in a co-occurrence information database, word pairs and co-occurrence frequency information of words that appear simultaneously within a predetermined range in a sentence in a target language corpus including a set of sentences of the target language. How to build a database. 2. A target language input step of inputting a target language corpus consisting of a set of target language sentences, and word pairs and co-occurrence frequency information of words in a sentence in the target language corpus that simultaneously appear within a predetermined range. 2. The method according to claim 1, further comprising the step of extracting co-occurrence information in the co-occurrence information database. 3. The target language co-occurrence information database,
2. The method according to claim 1, wherein a set of entries including a part-of-speech-tagged word pair, which is a word pair to which part-of-speech information is added, of the target language and co-occurrence frequency information thereof is maintained. 4. A target language input step of inputting a target language corpus consisting of a set of sentences in which each word in the target language is given the part of speech information of the word, and a sentence in the target language corpus includes: 4. The method according to claim 3, further comprising a co-occurrence information extracting step of storing, in the co-occurrence information database, a part-of-speech-tagged word pair and its co-occurrence frequency information that appear simultaneously in the database. 5. The co-occurrence information extracting step includes, when collecting a word pair with a part-of-speech tag, counting co-occurrence words that are words co-occurring with a key word that is a key word for each part of speech. Extracting the co-occurrence order by part of speech that indicates the nearest part of each co-occurrence word for each co-occurrence word, and co-occurrence frequency information for each of the key word and the co-occurrence word and the co-occurrence order by part of speech 5. The method according to claim 4, wherein the information is stored in an activation information database. 6. A field input step of inputting field information, and a field corpus extraction step of extracting a set of sentences related to the field information from the target language corpus, wherein the co-occurrence information extraction step further comprises: 6. The method according to claim 2, wherein co-occurrence frequency information of a target language is collected in a sentence extracted in the corpus input step. 7. A translation candidate generation method for searching for a target language translation candidate of each source language word in a source language sequence to be translated, using a translation dictionary holding a set of bilingual relationships between source language words and target language translation word candidates. And collating the target language candidate with the entry of the co-occurrence information database according to any one of claims 1 to 6.
A machine comprising: a co-occurrence strength detection step of calculating a co-occurrence strength of the target language translation word candidate pair for a source language word; and a translation word determining step of selecting a target language translation word sequence for the source language word using the co-occurrence strength. Translation method. 8. The method of claim 7, further comprising inputting said source language word sequence. 9. In the co-occurrence intensity detecting step, using the co-occurrence information database according to claim 3 including part of speech information, the co-occurrence information database is used to generate a target language translation word candidate pair for a source language word using the part of speech information. 9. The method according to claim 7 or 8, wherein the intensity is calculated. 10. The co-occurrence intensity detecting step,
The method according to claim 9, wherein the co-occurrence strength is calculated by taking into account the co-occurrence rank for each part of speech according to claim 5 and the co-occurrence frequency information. 11. The translation dictionary holds a set of correspondences between target language words and target language word-related words that are variations, derivatives, and synonyms of the target language words. When referring to the translation dictionary to match each target language translation candidate with the target language word pair in the entry of the co-occurrence information database, not only the target language words but also the target language word related to the target language word The method according to any one of claims 7 to 10, wherein the co-occurrence strength is calculated by also matching words. 12. The co-occurrence intensity detecting step,
Referring to the translation dictionary, when matching each target language translation word candidate with the target language word pair in the entry of the co-occurrence information database, not only the target language words but also the words of the target language and the meaning of the words And calculating a co-occurrence strength by referring to a semantic category dictionary composed of a set of correspondence relations with semantic categories representing the target language, and comparing the semantic categories of the target language words with each other.
The method according to any one of the preceding claims. 13. The translation dictionary, wherein the translation dictionary includes rule constraint information describing a condition for determining a target language translation word candidate to be selected for the source language word. And assigning rule constraint information to the target language translation candidate, and selecting the target language translation candidate not inconsistent with the rule constraint information assigned to the target language translation candidate in the translation word determining step. The method according to claim 1. 14. In the translated word candidate generating step,
In addition to the translation dictionary used in the translation word candidate generation step, referring to another translation dictionary holding a set of bilingual relations between the source language word and the target language translation word candidate, the source language words in the source language word string are referred to. 8. A target language translation word candidate is searched.
14. The method according to any one of claims 13 to 13. 15. In the co-occurrence intensity detecting step,
Using a source language co-occurrence information database that holds a set of entries consisting of source language word pairs and their co-occurrence frequency information,
The method according to any one of claims 7 to 14, wherein a weight of a corresponding candidate word co-occurrence is weighted based on the co-occurrence intensity of the source language word pair in the source language. 16. A target language input means for inputting a target language corpus consisting of a set of sentences in a target language, word pairs in a sentence in the target language corpus that appear simultaneously within a predetermined range and co-occurrence frequency information thereof Target language co-occurrence information database construction apparatus having co-occurrence information extraction means for accumulating the information in a co-occurrence information database. 17. The apparatus according to claim 16, wherein the target language co-occurrence information database holds a set of entries including a part-of-speech tagged word pair of the target language and its co-occurrence frequency information. 18. The target language input unit inputs a target language corpus consisting of a set of sentences in which each word of the target language is given the part of speech information of the word, and the co-occurrence information extracting unit outputs 2. The co-occurrence information database stores, in sentences in a corpus, word-word pairs with part-of-speech tags and co-occurrence frequency information that appear simultaneously within a predetermined range.
An apparatus according to claim 7. 19. The co-occurrence information extracting means, when collecting a word pair with a part-of-speech tag, counts co-occurrence words which are words co-occurring with a key word which is a key word for each part of speech. The co-occurrence order for each co-occurrence word indicating the closest to each part-of-speech is extracted. 2. A means for accumulating data in an activation information database.
An apparatus according to claim 8. 20. A field input means for inputting field information,
A field corpus extracting means for extracting a set of sentences related to the field information from the target language corpus, wherein the co-occurrence information extracting means co-occurs with the target language in the sentence extracted by the field corpus inputting means. 20. Apparatus according to any one of claims 16 to 19, comprising collecting frequency information. 21. A translation dictionary holding a set of bilingual relationships between source language words and target language translation word candidates, and using the translation dictionary to convert target language translation candidate candidates for each source language word in the source language word string to be translated. A target word candidate generating means for searching, and comparing the target language candidate candidate with an entry in the co-occurrence information database according to claim 16 to calculate a co-occurrence strength of the target language target word candidate pair with respect to a source language word. A machine translation apparatus comprising: a co-occurrence strength detection unit that performs the translation; and a translation word determination unit that selects a target language translation string for the source language word using the co-occurrence strength. 22. The apparatus according to claim 21, further comprising source language word string input means for inputting the source language word string. 23. The co-occurrence strength detection means uses the co-occurrence information database including the part of speech information to determine the co-occurrence strength of a target language translation candidate pair for a source language word using the part of speech information. Apparatus according to claim 21 or 22, which calculates. 24. The method according to claim 21, wherein the co-occurrence intensity detection unit includes a unit for calculating a co-occurrence intensity in consideration of the part-of-speech co-occurrence order and the co-occurrence frequency information. apparatus. 25. The translation dictionary holds a set of correspondences between target language words and target language word-related words that are variants, derivatives, and synonyms of the target language words. When referring to the translation dictionary to match each target language translation candidate with the target language word pair in the entry of the co-occurrence information database, not only the target language words but also the target language word related to the target language word 25. The apparatus according to any one of claims 21 to 24, further comprising means for calculating a co-occurrence strength by also matching words. 26. A semantic category dictionary comprising a set of correspondences between words in the target language and semantic categories representing the meanings of the words, wherein the co-occurrence strength detecting means refers to the translation dictionary. When matching each target language translation word candidate with a target language word pair in an entry of the co-occurrence information database, not only the target language words but also the meaning category dictionary is referred to, and the meaning of the target language word is referred to. 26. The apparatus according to any one of claims 21 to 25, further comprising means for calculating co-occurrence strength by matching categories. 27. The translation dictionary includes rule constraint information describing conditions for determining a target language translation word candidate to be selected for the source language word. And means for assigning rule constraint information to the target language translation candidate, wherein the translation word determination means selects a target language translation candidate that is consistent with the rule constraint information assigned to the target language translation candidate. 27. The device according to any one of claims 21 to 26, comprising: 28. The translation word candidate generation means refers to another translation dictionary holding a set of bilingual relationships between source language words and target language translation word candidates in addition to the translation dictionary used in the translation word candidate generation means. 28. The apparatus according to claim 21, further comprising means for searching for a target language translation word candidate of each source language word in the source language word string. 29. A source language co-occurrence information database, further comprising a source language co-occurrence information database for holding a set of entries consisting of source language word pairs and co-occurrence frequency information thereof. The apparatus according to any one of claims 21 to 28, further comprising means for weighting the corresponding candidate word co-occurrence strength based on the co-occurrence strength in the source language for the source language word pair using the source language. 30. A target language input procedure for inputting a target language corpus consisting of a set of sentences in a target language, word pairs appearing simultaneously within a predetermined range in a sentence in the target language corpus, and co-occurrence frequency information thereof Recording medium for storing a co-occurrence information extraction program for causing a computer to execute a co-occurrence information extraction procedure for accumulating the co-occurrence information in a co-occurrence information database. 31. The target language co-occurrence information database holds a set of entries consisting of a part-of-speech-tagged word pair, which is a word pair to which part-of-speech information is added, of the target language and co-occurrence frequency information thereof. 30. The recording medium according to 30. 32. A target language input procedure for inputting a target language corpus consisting of a set of sentences to which each part of the word of the target language is assigned the part of speech information of the word, and a target language corpus which is defined in the sentences in the target language corpus. A recording medium for recording a co-occurrence information extraction program for causing a computer to execute a co-occurrence information extraction procedure for accumulating in a co-occurrence information database the word pairs with part-of-speech tags that co-occur in a range. 33. The co-occurrence information extracting step, when collecting a word pair with a part-of-speech tag, collects co-occurrence words which are words co-occurring with a key word which is a key word for each part of speech. The co-occurrence rank is extracted for each co-occurrence word, indicating the number of the part-of-speech co-occurrence order. 33. The recording medium according to claim 32, wherein the recording medium is stored in an activation information database. 34. A field input procedure for inputting field information;
A field corpus extraction step of extracting a set of sentences related to the field information from the target language corpus, wherein the co-occurrence information extraction step includes a co-occurrence frequency of the target language in the sentences extracted in the field corpus input procedure. 34. The recording medium according to claim 30, which collects information. 35. Generation of a translated word candidate for searching a target language translated word candidate of each source language word in a source language word string to be translated using a translation dictionary holding a set of bilingual relations between the source language word and the target language translated word candidate 35. A procedure, said target language translation word candidate and any one of claims 30 to 34.
A co-occurrence strength detection procedure for calculating the co-occurrence strength of the target language translation word candidate pair with respect to the source language word by comparing the entry with the co-occurrence information database described in the section, A recording medium recording a machine translation program for causing a computer to execute a translated word determination procedure for selecting a target language translated word sequence for a word. 36. The recording medium according to claim 35, further comprising a step of inputting the source language word string. 37. The co-occurrence intensity detection procedure, wherein the co-occurrence information database according to claim 31 is used, wherein the co-occurrence information database is used to generate a target language translation word candidate pair for a source language word using the co-occurrence information. 4. The co-occurrence strength of the following is calculated.
37. The recording medium according to 5 or 36. 38. The recording medium according to claim 37, wherein said co-occurrence intensity detecting step calculates the co-occurrence intensity in consideration of said part-of-speech-specific co-occurrence order and said co-occurrence frequency information. 39. The translation dictionary holds a set of correspondences between target language words and target language word-related words that are variants, derivatives, and synonyms of the target language words. When referring to the translation dictionary to match each target language translation candidate with the target language word pair in the entry of the co-occurrence information database, not only the target language words but also the target language word related to the target language word The recording medium according to any one of claims 35 to 38, wherein the co-occurrence strength is calculated by collating with a word. 40. In the co-occurrence intensity detecting step, when matching each target language translation word candidate with a target language word pair in an entry of the co-occurrence information database with reference to the translation dictionary, the target language words are compared with each other. As well as a semantic category dictionary composed of a set of correspondences between words in the target language and semantic categories representing the meaning of the words,
The recording medium according to any one of claims 35 to 39, further comprising a step of comparing the semantic categories of the target language words with each other to calculate a co-occurrence strength. 41. The translation dictionary includes rule constraint information describing a condition for determining a target language translation word candidate to be selected for the source language word. 42. Rule translation information is assigned to the target language translation candidate with reference to the above, and in the translation word determination step, a target language translation candidate not inconsistent with the rule constraint information assigned to the target language translation candidate is selected. The recording medium according to any one of the preceding claims. 42. In the translated word candidate generating step, in addition to the translation dictionary used in the translated word candidate generating step, another translation dictionary holding a set of bilingual relationships between source language words and target language translated word candidates is referred to. 42. The recording medium according to claim 35, wherein a target language translation word candidate of each source language in the source language word string is searched. 43. In the co-occurrence strength detection procedure, a source language co-occurrence information database that holds a set of entries consisting of source language word pairs and their co-occurrence frequency information is used. 43. The recording medium according to claim 35, wherein a weight is assigned to the corresponding translated word candidate co-occurrence strength based on the co-occurrence strength.