JP2003085169A - Parallel translation compound word dictionary updating method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a parallel translation compound word dictionary updating method for registering a dictionary by extracting a parallel translation compound word at an inexpensive operation cost. SOLUTION: A patent document pair file is made by gathering a pair of patent documents coincident in a priority number of a Japanese patent document and a filing number of an English patent document (S01). A Japanese compound word appearing in Japanese patent documents is extracted from the respective Japanese patent documents in the patent document pair file (S02). An English compound word appearing in English patent documents is extracted from the respective English patent documents in the patent document pair file (S03). A cooccurrence degree of respective English compound words is calculated to this Japanese compound word with every Japanese compound word in the same patent document pairs (S04). The English compound word having a high cooccurrence degree to the Japanese compound word is outputted b the proper number as parallel translation compound word candidates with every Japanese compound word (S05). A parallel translation English compound word of the Japanese compound word is selected from the parallel translation compound word candidates of the English compound word outputted with every Japanese compound word (S06). A parallel translation compound word dictionary is updated thereby (S07).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the field of language processing technology between a plurality of languages, and relates to a bilingual compound word dictionary updating method for extracting a bilingual compound word and registering it in a dictionary at a low operating cost.

[0002]

2. Description of the Related Art Due to the spread of the Internet, it has become an era where various foreign language information can be easily obtained, and there is an increasing demand for users to read their foreign language information in their native language, or conversely, to read their own information in other languages. is doing. However, new words appear every day on the Internet (such as Web pages), and many new words and technical terms that are not registered in existing translation dictionaries are included, so machine translation and multilingual search are available. I often cannot. How to register these unregistered words in the dictionary quickly and inexpensively is an important issue in language processing technology.

As a measure against such a problem, a statistical method has been considered in which a bilingual translation is automatically extracted in units of words or phrases from a multilingual corpus (example sentence collection) having a bilingual relationship, but following the occurrence of a new word. Since there is no prospect of quickly obtaining the latest corpus, it has not been put to practical use.

If a page having a bilingual relationship or a page on which bilingual information is posted can be extracted from the Web, an inexpensive bilingual corpus can be acquired by periodically collecting the pages. However, it is difficult to mechanically acquire only high-quality information because low-quality information is mixed in a Web page, and a Japanese page has a bilingual translation like a European language. There are few such pages, and in the case of compound words, simply compounding the bilingual words for each word is not sufficient.

[0005]

SUMMARY OF THE INVENTION In view of such circumstances, it is an object of the present invention to provide a bilingual compound word dictionary updating method for extracting a bilingual compound word and registering it in the dictionary at a low operating cost. It is what

[0006]

The above-mentioned problems can be solved by the following means. That is, the solution means of the first invention is to create a patent document pair file that collects a pair of patent document pairs in which the priority number of the first language patent document and the application number of the second language patent document match. Patent document pair creating step, a first language compound word extracting step for extracting a first language compound word appearing therein from each first language patent document in the patent document pair file, and the above patent document From each second language patent document in the document pair file:
The second to extract the second language compound that appears in this
In the language compound word extraction step and the same patent document pair,
For each of the first language compound words, a co-occurrence degree calculating step for calculating the degree of co-occurrence of each of the second language compound words with respect to the first language compound word, and for each of the first language compound words , A parallel compound word candidate output step for outputting an appropriate number of second language compound words having a high degree of co-occurrence as bilingual compound word candidates, and for each first language compound word in the bilingual compound candidate output step. A selection step for selecting a parallel translation second language compound word of the first language compound word from the output bilingual compound word candidates of the second language compound word, and a first language compound word selected in the above selection step And a dictionary registration step for registering the bilingual second language compound word in the bilingual compound word dictionary.

According to a second aspect of the present invention, in the method for updating a bilingual compound word dictionary according to the first aspect of the present invention, the patent document pair file created in the patent document pair creating step is referred to as the first language patent document. This is a patent document pair file created by using some columns corresponding to each other in the second language patent document.

According to a third aspect of the present invention, in the method for updating a bilingual compound word dictionary according to the second aspect of the present invention, the above-mentioned partial columns are replaced by a column of the title of the invention, a column of the abstract, and a column of claims. ,
One or more columns in the detailed description of the invention or the examples.

According to a fourth aspect of the invention, in the bilingual compound word dictionary updating method according to the first to third inventions, the bilingual compound word candidate output step is not registered in the current bilingual compound word dictionary. The bilingual compound word of is extracted and output.

A fifth solution of the invention is, in the bilingual compound word dictionary updating method of the first to fourth inventions, that the patent document pair file created in the patent document pair creating step is stored in each language. It is made up of a set of files for each.

According to a sixth aspect of the present invention, there is provided a method for updating a bilingual compound word dictionary according to any one of the first to fifth inventions, wherein the first language compound word extracted in the first language compound word extracting step and The second language compound word extracted in the second language compound word extracting step is generated by morphologically analyzing each patent document and generating the extracted word in accordance with a predetermined combining rule.

According to a seventh aspect of the present invention, in the method for updating a bilingual compound word dictionary according to the first to sixth aspects of the invention, the co-occurrence degree S (J, E) has the expression S (J, E) ＝ log F (J, E) ・ 2 F (J, E) / (F (J) + F (E)) (where J and E are first language compound words and second language compound words, respectively) F (J) and F (E) are the first
The frequency of occurrence of the first language compound word in the language patent document and the frequency of occurrence of the second language compound word in the second language patent document are shown. F (J, E) is the first language compound in the patent document pair. It is the co-occurrence frequency in which a word and a second language compound word both appear. ) Is used.

The eighth means of solving the problem is, in the method for updating a bilingual compound word dictionary according to the first to seventh inventions, wherein the first language and the second language are Japanese and English, respectively. The first language patent document and the second language patent document are referred to as Japanese patent document and US patent document, respectively.

According to a ninth aspect of the present invention, in the method for updating a bilingual compound word dictionary according to the eighth invention, the combining rule for generating a Japanese compound word in the above combining rule is set to 1
Combines a series of nouns, unknown words, prefixes, independent verbs, and independent adjectives, except for non-independent nouns, pronouns, and numbers. 2 If the last word combined is a prefix or an independent adjective, it is not a compound word. For a compound word having more than three combined words, a combination of less words is generated. 4 suffix nouns combine with the previous word.
5 Katakana Unknown words are treated as one word and the above rules are applied. In addition, the combination rule for generating an English compound word combines a sequence of one adjective and noun. For a compound word in which the number of words combined by 2 is more than 3, a combination of words other than this is generated.
3 If the last word combined is an independent adjective, it is not a compound word. It was decided.

According to a tenth aspect of the present invention, in the method for updating a bilingual compound word dictionary according to any one of the first to ninth aspects, the first language patent document is preliminarily divided for each technical field by a search. This is a method of updating a bilingual compound word dictionary.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Since the present invention extracts compound word bilingual information from corresponding patent documents and updates the dictionary accordingly,
In principle, the language does not matter. Therefore, the present invention will be described here by taking as an example a method of updating a bilingual compound word dictionary from corresponding patents filed in Japan and the United States. In the present invention, the corresponding patents filed in Japan and the US based on the priority claim are used as a bilingual corpus, the Japanese-English bilingual compound words are extracted therefrom, and the bilingual compound word dictionary is updated.

FIG. 1 is a flow explanatory diagram showing the concept of the flow of processing in the parallel compound word dictionary updating method of the present invention.
In this embodiment, a Japanese patent publication database and a US patent specification database are used. Here, the first language in the claims is Japanese and the second language is English (US).
In addition, the first language patent document corresponds to the Japanese patent publication, and the second language patent document corresponds to the US patent specification.

When the process starts (step S00), a pair of patent documents in which the priority number of the Japanese patent publication and the application number of the US patent specification match in the patent document pair creating step (S01) are collected. A patent document pair file is created. It should be noted that FIG. 7 shows that the priority number and the application number (enclosed by a dotted line) have a corresponding relationship in the Japanese Patent Laid-Open Publication and the US patent specification.

FIG. 2 is an explanatory view of the flow outline of this patent document pair creating step. In step S11, the publication of the application filed for the US priority is extracted from the Japanese published patent publication database by referring to the priority claim number and the priority claiming country, and the US priority publication is published. Create a file. If there are multiple Japanese published patent publications having the same priority number due to a divisional application or the like, all of them can be targeted.

In step S12, the US patent specification having the same application number as the priority number is extracted from the US patent specification database by using the priority claim number as a key,
Create a US patent file for Japanese applications. Since these U.S. priority claim publication files and US patent files corresponding to Japanese applications are paired with each other, they are treated as patent document pair files. It may be another file.

Japanese (first language) compound word extraction step S
In 02, the Japanese (first language) compound word appearing therein is extracted from each Japanese patent publication in the patent document pair file. FIG. 3 is a flow schematic explanatory view showing the contents of the Japanese (first language) compound word extraction step.

A word is extracted from the US priority claim publication file created in step S01 by morphological analysis (step S21). Note that this word can be extracted by a morphological analysis / part-of-speech assignment program, for example, "chasen" (chasen, Nara Institute of Science and Technology, Graduate School of Information Science, Natural Language Processing Course (Matsumoto Lab), http: // chasen.aist-n
ara.ac.jp/chasen/distribution.html.ja) can be used. Further, prefixes such as "the book, each, the above" in "the system, the present invention, each process, the above processing" which are very often found in patent publications are removed by the affix processing.

In order to set a combination rule for combining Japanese compound words from the extracted words, a combination rule for Japanese compound words extracted from the bilingual corpus (patent document pair file) is set as follows.

1 Excluding non-autonomous nouns, pronouns and numbers,
Combines nouns, unknown words, prefixes, independent verbs, and independent adjective sequences in a sentence. An example of this is shown in FIG. 2 If the last word combined is a prefix or an independent adjective, it is not a compound word. An example of this is shown in FIG. 3 For compound words with more than 3 words combined,
In addition to this, less word combinations are generated. An example of this is shown in FIG. 4 The suffix noun combines with the previous word. 5. Treat unknown katakana words as one word and apply the above criteria. An example of this is shown in FIG.

Based on the above Japanese combining rules, in step S22, the words extracted in step S21 are combined to obtain a Japanese compound word file.

Next, in step S03, the English (second language) compound word appearing therein is extracted from each US patent specification in the patent document pair file. FIG. 4 is a flow schematic explanatory view showing the contents of the English (second language) compound word extraction step.

From the US patent file corresponding to the Japanese application created in step S01, the part-of-speech assignment program (step S3
Extract the word by 1). For the part-of-speech assignment program, for example, "Brill Tagger" (http: //www.cs.jhu.
edu / ~ brill / home.html) can be used.

In order to set the English combination rule for combining the English compound words from the extracted words, the English combination rule of the English compound words is set as follows.

1 Combines adjective and noun sequences. This example is shown in FIG. 2 For compound words with more than 3 combined words,
In addition to this, less word combinations are generated. 3 If the last word combined is an independent adjective, it is not a compound word. This example is shown in FIG. In the case of English, there are inversion compound words with coordinate conjunctions (for, in, of, etc.), but there are few terms that contain coordinate conjunctions in the technical term dictionary. Absent.

Based on this English combining rule, step S3
In step 2, the words extracted in step S31 are combined to create an English compound word file.

In the compound word file, the Japanese compound word and the English compound word are not yet associated with each other. For this association, the co-occurrence degree calculating step S04 and the parallel compound word candidate outputting step S05 are executed. This co-occurrence degree calculation step S04 and the parallel compound word candidate output step S0
The contents of No. 5 are shown in FIG.

Step S in co-occurrence degree calculation step S04
41, in the same patent document pair, the degree of co-occurrence of each English (second language) compound word with respect to this Japanese (first language) compound word is calculated for each Japanese (first language) compound word. To be done. The co-occurrence degree is a measure of the degree to which the Japanese compound word and the English compound word appear in parallel. For example, "weighted Dice coefficient" by Mihoko Kitamura and Yuji Matsumoto.
The following scores S (J, E) proposed as can be used (Note: Mihoko Kitamura, Yuji Matsumoto. Automatic extraction of bilingual expressions using bilingual corpus, Journal of Information Processing Society of Japan, Vol.
38, No.4, pp.727-736, 1997. ).

S (J, E) = log F (J, E) ・ 2 F (J, E) / (F (J)
+ F (E)) Here, J and E represent a Japanese (first language) compound word and an English (second language) compound word, respectively. Also, F (J) and F (E)
Represents the frequency of appearance of Japanese (first language) compound words in Japanese (first language) patent documents and the frequency of appearance of English (second language) compound words in US (second language) patent documents, respectively. Therefore, F (J, E) is the co-occurrence frequency in which the Japanese compound word and the English compound word both appear in the patent document pair.

From this equation, the co-occurrence degree for various combinations of Japanese compound words and English compound words can be obtained. In step S42, N (for example, 4 or 5) parallel translation candidate lists are output in descending order of co-occurrence (score) for each Japanese compound word. Although not shown in the figure, at this time, the compound word pairs included in the current bilingual compound word dictionary to be updated are excluded, and only unregistered compound word pairs are output as the bilingual candidate list.

When a certain Japanese compound word appears at a corresponding place in the publication, if a particular English compound word appears almost regularly or at a high frequency, this means that the Japanese compound word and its English word. It can be said that this has a close relationship with the compound word, and that there is a high possibility that the two compound words have a bilingual relationship with each other. It should be noted that outputting only an appropriate number of high co-occurrence degrees also means that the burden of the next manual selection work is reduced.

In the selection step S06 and the dictionary registration step S07, the bilingual compound word for each Japanese compound word is selected based on the bilingual candidate list output in step S42. FIG. 6 shows the contents of the selection step S06 and the dictionary registration step S07. In step S51, the parallel translation candidate list, that is, a plurality of English compound words (candidates) corresponding to each Japanese compound word is displayed (or printed out), and in step S52, a person looks at it and the Japanese compound word Select an English compound word that can be judged to correspond to the word. Selection of all Japanese compound words appearing here is completed, and the parallel translation selection list is completed.

The contents of the parallel translation selection list completed in step S53 are added to the old parallel translation compound word dictionary, and the update of the parallel translation compound dictionary is completed.

In the experiment, out of the approximately 1.75 million cases published as Japanese published patent publications during the five years from 1995 to 1999, 32,590 applications with priority claims in the United States were used. As a result of extracting US registered publications based on the priority claim number, patent publication 32,8
96 cases were obtained. The total number of extracted Japanese is 1,100,416
There were 345,291 different words, 1,233,070 words in total, and 439,020 words different. An example of the extracted parallel translation is shown in the table of FIG.

Due to the nature of updating the bilingual dictionary, the correct answer rate is 10
Unless it is 0%, automatic renewal registration cannot be performed, so it is unavoidable to perform manual correction confirmation work, but the load of manual work can be greatly reduced. This is a huge advantage in terms of operation.

In the above description, the name and the abstract of the invention are taken as an example of the corresponding parts in the publication for creating the patent document pair file, but not only this,
Each section of the claims, the detailed description of the invention, and the embodiments can be targeted. Furthermore, in the case of a compound word dictionary for each field, it is possible to perform the above-mentioned processing after the patent documents are divided in advance for each technical field by a search based on the international classification, FI, F-term, corresponding US classification, and the like. As a result, the dictionary can be updated for each specialized field.

As shown in the above embodiments, since the patent publication (specification) having a relatively well-formatted format can be used, it is possible to correspond the bilingual compound word dictionary to a compound word which appears newly every day. Further, since the patent publication (specification) is data provided by the governments of the respective countries, it can be obtained at a low cost, and the bilingual compound word dictionary can be updated at a low operating cost. In addition, the new compound word candidates are displayed (or printed out) by limiting only those that have a high probability of being translated.
Selection work becomes easy.

[0042]

According to the present invention, since a patent publication (specification) having a relatively well-formatted format can be used, it is possible to make the bilingual compound word dictionary correspond to a compound word that appears newly every day. Further, since the patent publication (specification) is data provided by the government of each country, it can be acquired at a low cost, and the bilingual compound word dictionary can be updated at a low operating cost. Further, the new compound word candidates are displayed (or printed out) by limiting only those that are highly probable as parallel translations, which has the effect of facilitating the selection work.

[Brief description of drawings]

FIG. 1 is a flow explanatory diagram showing the concept of the flow of processing in a bilingual compound word dictionary updating method according to the present invention.

FIG. 2 is a flow outline explanatory diagram of a patent document pair creation step.

FIG. 3 is a flow schematic explanatory view showing the contents of a Japanese (first language) compound word extraction step.

FIG. 4 is a flow schematic explanatory view showing the contents of an English (second language) compound word extraction step.

FIG. 5 is a flow outline explanatory diagram showing the contents of the co-occurrence degree calculating step S04 and the parallel translation compound word candidate outputting step S05.

FIG. 6 is a selection step S06 and a dictionary registration step S0.
It is a flow outline explanatory drawing which shows the contents of 7.

FIG. 7 is an explanatory diagram showing that the priority number and the application number (enclosed by a dotted line) have a correspondence relationship between the Japanese Patent Laid-Open Publication and the US patent specification.

FIG. 8 is an explanatory diagram showing an example (Japanese) of combining a sequence of nouns, unknown words, prefixes, independent verbs, and independent adjectives, excluding non-independent nouns, pronouns, and numerals.

FIG. 9 is an explanatory diagram showing an example (Japanese) in which a combined word is not a compound word when the last word is a prefix or an independent adjective.

FIG. 10 is an explanatory diagram showing an example (Japanese) in the case of generating a combination of less words for a compound word having more than three combined words.

FIG. 11 is an explanatory diagram showing an example (Japanese) in which an unknown katakana word is treated as one word.

Figure 12: Example of combining adjective and noun sequences (English)
FIG.

FIG. 13 is an explanatory diagram showing an example (English) in which a combined word is not a compound word when the last word is an independent adjective.

FIG. 14 is a table showing an extracted parallel translation example.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Shigeto Higuchi             2-4-29 Shiohama, Koto-ku, Tokyo Stock market             Company Patrice (72) Inventor Masatoshi Fukui             2-4-29 Shiohama, Koto-ku, Tokyo Stock market             Company Patrice F term (reference) 5B091 AA05 AB11 BA02 CC02 CC16

Claims (10)

[Claims] 1. A patent document pair creation step for creating a patent document pair file that collects a pair of patent document pairs in which the priority number of the first language patent document and the application number of the second language patent document match. A first language compound word extraction step for extracting a first language compound word appearing therein from each first language patent document in the patent document pair file, and each of the first language compound word in the patent document pair file, A second language compound word extracting step for extracting a second language compound word appearing therein from a second language patent document, and this first language for each of the first language compound words in the same patent document pair. A co-occurrence degree calculating step for calculating a degree of co-occurrence of each of the second language compound words with respect to the compound word; and a second language compound with a high co-occurrence degree for each of the first language compound words. Bilingual translation From the bilingual compound candidate output step for outputting an appropriate number of word candidates and the bilingual compound candidate of the second language compound word output for each first language compound word in the bilingual compound candidate output step, Parallel translation of the first language compound word: a selecting step for selecting the second language compound word; and registering the first language compound word and its bilingual second language compound word selected in the above selecting step in the bilingual compound word dictionary And a dictionary registration step, the method of updating a bilingual compound word dictionary. 2. The bilingual compound word dictionary updating method according to claim 1, wherein the patent document pair files created in the patent document pair creating step are the first language patent document and the second language patent document. A bilingual compound word dictionary updating method, which is a patent document pair file created by using some columns corresponding to each other. 3. The method for updating a bilingual compound word dictionary according to claim 2, wherein the part of the columns includes a title column of the invention, a summary column, a claims column, and a detailed description of the invention. A parallel translation compound word dictionary updating method, characterized in that it is one column or a plurality of columns among the columns or the columns of the embodiment. 4. The bilingual compound word dictionary updating method according to claim 1, wherein in the bilingual compound word candidate output step, the bilingual compound word not registered in the current bilingual compound word dictionary is used. A method for updating a bilingual compound word dictionary, characterized in that is extracted and output. 5. In the bilingual compound word dictionary updating method according to any one of claims 1 to 4, the patent document pair file created in the patent document pair creating step is a file for each language. A method for updating a bilingual compound word dictionary comprising a set. 6. The method for updating a bilingual compound word dictionary according to claim 1, wherein the first language compound word and the second language are extracted in the first language compound word extracting step. A bilingual compound word dictionary updating method characterized in that the extracted second language compound word in the compound word extracting step is generated by morphologically analyzing each patent document and the extracted word is generated according to a predetermined combining rule. . 7. The bilingual compound word dictionary updating method according to any one of claims 1 to 6, wherein the co-occurrence degree S (J, E) is expressed by the equation S (J, E) = log. F (J, E) ・ 2 F (J, E) / (F (J) + F (E)) (where J and E are the first language compound word and the second language compound word, respectively. , F (J) and F (E) are the first
The frequency of occurrence of the first language compound word in the language patent document and the frequency of occurrence of the second language compound word in the second language patent document are shown. F (J, E) is the first language compound in the patent document pair. It is the co-occurrence frequency in which a word and a second language compound word both appear. ) Is used, a parallel translation compound word dictionary updating method. 8. The method for updating a bilingual compound word dictionary according to any one of claims 1 to 7, wherein the first language and the second language are Japanese and English, respectively. The bilingual compound word dictionary updating method, wherein the patent document and the second language patent document are a Japanese patent document and a US patent document, respectively. 9. The bilingual compound word dictionary updating method according to claim 8, wherein the combining rule for generating the Japanese compound words in the combining rule is as follows: 1 except for non-independent nouns, pronouns, and numbers Noun,
Combines a sequence of unknown words, prefixes, independent verbs, and independent adjectives. 2 If the last word combined is a prefix or an independent adjective, it is not a compound word. 3 For compound words with more than 3 words combined,
In addition to this, less word combinations are generated. 4 The suffix noun combines with the previous word. 5 Treat katakana unknown words as one word and apply the above rules. I will. Further, the combination rule for generating an English compound word combines a sequence of 1 adjectives and nouns. 2 For compound words with more than 3 combined words,
In addition to this, less word combinations are generated. 3 If the last word combined is an independent adjective, it is not a compound word. I will. A bilingual compound word dictionary updating method characterized by the above. 10. The method for updating a bilingual compound word dictionary according to any one of claims 1 to 9, wherein the first language patent document is divided in advance by search into each technical field. A bilingual compound word dictionary updating method characterized by.