JP2840258B2 - Method of creating bilingual dictionary and co-occurrence dictionary for machine translation system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a machine translation method and system having a bilingual dictionary and a co-occurrence relation dictionary having a self-propagating function.

[Conventional technology]

An important component of a machine translation system is a dictionary. The dictionary contains words in the first language (source language) and second language (target language) and their attribute information (part of speech, semantic codes, case frames, etc.), and translations between words in the first language and words in the second language It contains information such as relations and co-occurrence relations of words in the first language or the second language. Hitherto, the creation of dictionaries has been left to manual labor, but there is a problem that enormous labor is required. If an automatic creation or self-reproduction function can be realized, the effect is extremely large. Dictionary information having a high possibility of automatic creation includes word co-occurrence relations. For example, Japanese Patent Laid-Open No. 232076/1987 discloses the extraction of word co-occurrence relations from sentence analysis results and accumulation in the knowledge base. A method for performing the operation is shown. The idea of extracting knowledge from sentences in this way is very useful because the ability of a machine translation system increases with the use of the system. However, the method disclosed in Japanese Patent Application Laid-Open No. 62-232076 is the first method.
This is limited to only the co-occurrence relation of words in a language, and cannot be applied to other dictionary information.

[Problems to be solved by the invention]

A first object of the present invention is to provide a machine translation method and system for acquiring bilingual knowledge.

A second object of the present invention is to provide a machine translation method and system for acquiring knowledge of co-occurrence relations of words in a second language without performing syntax / semantic analysis of the second language.

[Means for solving the problem]

In order to achieve the first object, a first feature of the present invention is to use a bilingual dictionary to provide a word-level correspondence between a sentence in a first language and a translated sentence in a second language. To identify
Among the identified bilingual relations, those not registered in the bilingual dictionary are registered in the bilingual dictionary.

In order to achieve the second object, a second feature of the present invention is that, in addition to the processing described above, a syntax
A semantic analysis is performed to extract the co-occurrence relations of the words included in the sentence, and the first language co-occurrence relation is used by using the correspondence between the words of the first language and the words of the second language obtained by the bilingual relation identification. The co-occurrence relation of the first language word obtained by the relation extraction is mapped to the co-occurrence relation of the second language word, and the obtained co-occurrence relation of the second language word is registered in the second language co-occurrence dictionary. Is to do.

[Action]

In the bilingual relationship identification process, when a sentence in the first language and a sentence in the second language, which is a translation thereof, are given, the correspondence at the word level is identified as follows. First, referring to the dictionary of the first language, it is identified that the sentence of the first language is composed of m words S (1),..., S (m). Similarly, referring to the dictionary of the second language, the sentence of the second language is composed of n words T (1),.
(N) is identified. (In the present invention, since it is fundamental to consider the bilingual relation of a sentence as a set of bilingual relations of words, it is preferable that m = n, but in practice, m = n is not always true.) After identifying the words that make up the sentences in one language and the sentences in the second language, the process proceeds to identification of the correspondence between words. For this purpose, first, the bilingual dictionary is referred to. If a set of S (i) and T (j) is included in the bilingual dictionary, it is determined that S (i) and T (j) also correspond to the bilingual sentence currently being processed. In this way,
It is assumed that the correspondence of p sets (provided that pm and pn) can be identified. If the bilingual sentence includes a correspondence that is not registered in the bilingual dictionary, p <m and p <n. Then, the remaining (mp) first language words and the (np) second words
The process proceeds to estimating the correspondence between words in the language. This estimation is not always possible. However, p = m-1
In the case of = n-1, since the remaining words are the same in both the first language and the second language, it can be determined that they correspond. Even when p = m−1 = n−1, m−p, n−p
If is small, it is often possible to identify the correspondence using information such as the part of speech as a clue. Further, when m> n, in a sentence in the first language, if a word in which a corresponding word in the second language is not identified is continuous, these words are regarded as one compound word. In some cases, the correspondence between words can be estimated. The same is true for m <n. From the above description, it can be understood that a word-level correspondence can be identified for a bilingual sentence including a bilingual relationship that is not registered in the bilingual dictionary.

In the bilingual dictionary registration process, among the word-level correspondences identified in the bilingual relationship identification process, those that have not been registered in the bilingual dictionary are registered in the bilingual dictionary. A machine translation system having the function of performing

Furthermore, the first language co-occurrence relation extraction process selects only unambiguous ones from a set of word dependencies obtained by syntactic and semantic analysis of sentences in the first language.
The co-occurrence relationship between words in the first language is extracted. The result is mapped to the co-occurrence relation of the words in the second language by the co-occurrence relation mapping processing, and is registered in the second language co-occurrence relation dictionary by the second language co-occurrence relation dictionary registration processing. In this way, a machine translation system having a function of acquiring a co-occurrence relationship between words in the second language from a bilingual sentence and accumulating the acquired words in a dictionary can be realized.

〔Example〕

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In this embodiment, it is assumed that the first language is Japanese and the second language is English. The hardware necessary for the machine translation system of the embodiment includes a central processing unit 1, an input unit 2, an output unit 3, a dictionary storage unit 4, and a text storage unit 5, as shown in FIG.

The central processing unit 1 performs a translation process, a text input / output and an update process, in addition to a process of acquiring knowledge of a bilingual relationship and a co-occurrence relationship according to the present invention. The input device 2 is used for inputting and correcting text, and the output device 3 is used for displaying text, but is not directly related to the present invention.

The dictionary storage device 4 includes a Japanese dictionary 41, a Japanese co-occurrence relation dictionary 42, a Japanese-English bilingual dictionary 43, an English dictionary 44, and an English co-occurrence relation dictionary 4.
5 is stored. Note that such division of the dictionary is logical, and the physical structure is not particularly limited, including integrating and storing a plurality of dictionaries. FIG. 3A illustrates a record of the Japanese dictionary 41. The record of the Japanese dictionary includes a Japanese word 411, a part of speech 412 as its attribute information, a meaning code 413, and a case frame 414. Although the English dictionary 44 is not illustrated, it is the same as the Japanese dictionary. FIG. 3 (b) shows an example of a record in the Japanese-English bilingual dictionary 43. A record in the Japanese-English bilingual dictionary is a pair of Japanese word 431 and English word 432. Figure 3 (c) shows the English co-occurrence dictionary 4
5 records are exemplified. The record of the English co-occurrence relation dictionary includes two English words 451 and 452 having a co-occurrence relation and a code 453 indicating a relation between them. Although the Japanese co-occurrence relation dictionary 42 is not illustrated, it is the same as the English co-occurrence relation dictionary.

The text storage device 5 has a Japanese text file 51
And the English text file 52 are stored. Each of the Japanese text and the English text is assigned a sentence number for each sentence, and sentences having the same sentence number have a bilingual relationship. Therefore,
A bilingual sentence can be searched using the sentence number as a key.

Next, according to FIG. 1, a process for identifying a correspondence at a word level from a Japanese sentence and an English sentence which is a translation thereof,
Further, a process of extracting a co-occurrence relationship between English words will be described.

The first step is to identify the words that make up the Japanese sentence. First, the sentence is divided into words while referring to the Japanese dictionary (processing 101). Since Japanese sentences do not include a space indicating a word boundary, slightly complicated processing is required. For example, a method disclosed in Japanese Patent Application No. 59-162443 may be used.
Next, a content word is selected from the words constituting the sentence (processing
102). Assume that m content words are included, and
(1),..., S (m). By this processing, functional words such as particles and auxiliary verbs are excluded from the processing for identifying the correspondence. Since the function words play an important role in the processing of the machine translation system, and the number is small, it can be considered that the description of the dictionary is completed. Also, the correspondence between languages is not as simple as content words.

The second step is to identify the words that make up the English sentence. Split sentence into words while referring to English dictionary (processing
103). Since English sentences have word boundaries shown as blanks, they can be realized by simple processing, except for the need for change meter processing. Next, a content word is selected from the words constituting the sentence (process 104). It is assumed that n content words are included, and they are T (1),..., T (n).

In the third step, referring to the bilingual dictionary, the Japanese words S (1),..., S (m) and the English words T (1),.
This is a process of identifying the correspondence between An index indicating a Japanese word is i, an index indicating an English word is j, and a mapping from i to j is σ. Also, let k be a register indicating the number of correspondences determined in the third step. The process in the third step starts by setting k to an initial value 0 (process 105). Next, i is set to the initial value 1 (process 106). Further, j is set to an initial value 1 (process 107). After this, S
(I) and σ ⁻¹ (j) are undecided (process 108) T
It is checked whether the pair of (j) is included in the bilingual dictionary (process 109). If the pair of S (i) and T (j) is included in the bilingual dictionary, σ is defined as a correspondence between i and j (process 110), and k is increased by 1 (process 111). S
If the pair of (i) and T (j) is not included in the bilingual dictionary, count up j to n (process 112) (process 113), and the pair of S (i) and T (j) The process of checking whether the entry is included in the bilingual dictionary is continued. The processing of 107 to 113 is continued while counting up (processing 115) until i becomes m (processing 114). By the above processing, the Japanese words S (1),..., S (m) and the English words T (1),.
Among the correspondences between (n), those included in the bilingual dictionary are identified.

The fourth step is an estimation of the correspondence that could not be identified in the third step. In the present embodiment, this is performed only when the estimation can be made most easily. That is, this is the case where the number m of Japanese words and the number n of English words match (process 116), and (m-1) correspondences are identified in the third step (process 117). In this case, since there is one j for which σ (i) has not been determined and one j for which σ ^-1 (i) has not been determined, this is searched (process 118). If the corresponding i, j is i ₀ , j _0, it is assumed that σ (i ₀ ) = j ₀ (process 119), and S
A pair of (i ₀ ) and T (j ₀ ) is registered in the bilingual dictionary (processing 12
0).

The fifth step is a process of performing syntactic and semantic analysis on Japanese sentences to extract co-occurrence relations of words (processing
121). That is, the dependency relationship between the content words S (1),..., S (m) obtained as a result of the first step is analyzed, and only the unambiguous one is selected.

From a Japanese sentence, one co-occurrence relation [S (i _p ), S
(I ' _p ), R _p ] (p = 1,..., L) is extracted.
Here, S (i _p ) and S (i ′ _p ) co-occur, and R _p is a code representing the relationship between them.

The sixth step is processing for mapping the co-occurrence relation of Japanese words obtained in the fifth step to the co-occurrence relation of English words.
Since the correspondence between Japanese words and English words is represented by σ, the co-occurrence of Japanese words [S (i _p ), S
(I ′ _p ), R _p ] is _defined as the co-occurrence relation of English words [T (σ
(I _p )), T (σ (i ′ _p )), R _p ] (processing 12
3). Thereafter, this is registered in the English co-occurrence dictionary (process 124). The above processing is performed on the index P indicating the co-occurrence relationship.
Is repeated from the initial value 1 (process 122) to 1 (process 125) while counting up (process 126).

The processing for acquiring the bilingual relations and co-occurrence relations of words has been described above with reference to FIG.

FIG. 4 shows an example in which a bilingual relation and a co-occurrence relation of words are obtained from a bilingual sentence. For bilingual sentences: ・ Update the document file.

・ Update the document file. This bilingual sentence is processed using the dictionary shown in FIG. The Japanese words obtained by the first step are as shown in FIG. The English words obtained by the second step are as shown in FIG. 4 (b). The correspondence between words obtained in the third step is as shown in FIG. 4 (c). Since the Japanese-English bilingual dictionary of FIG. 3 (b) does not include a pair of "update" and "update", this correspondence has not been identified. This is the fourth
And the word correspondence shown in FIG. 4 (d) is obtained. In the Japanese-English bilingual dictionary, "update" and "up
The date pair is registered and has the contents shown in FIG. 4 (e). Further, in the fifth step, the co-occurrence relation of Japanese words shown in FIG. 4 (f) is obtained. This is mapped to the English word co-occurrence relation shown in FIG. 4 (g) in the sixth step, and is registered in the English co-occurrence relation dictionary. The content of the English co-occurrence dictionary changes from FIG. 3 (c) to FIG. 4 (h).

In the present embodiment, the processing in the fourth step is described only in the case where the correspondence can be estimated most easily. Here, it is shown that the ability to estimate the correspondence can be improved by making some contrivances.

For example, even if the number m of Japanese words and the number n of English words are the same, a plurality of correspondences may not be identified in the third step. Now, the Japanese-English bilingual dictionary in FIG.
Suppose that it does not include a pair of “document” and “document”. At this time, in the example of FIG. 4, the result of the third step is as shown in FIG. 5 (c). In other words, "update" and "upda
This means that the correspondence between two sets of "te", "document", and "document" has not been identified. In such a case, the correspondence may be estimated using the part of speech of the word. That is,
Both "document" and "file" are nouns, and "document file" can be considered a noun phrase. Also,
"Document" and "file" are both nouns, and "docume
nt file "can be considered a noun phrase. Here, since the correspondence between "file" and "file" has been identified in the third step, the correspondence between "document" and "document" can be estimated. Thus, the result of the fourth step is as shown in FIG.

Next, an example of how to deal with the case where the number m of Japanese words and the number n of English words are not the same is illustrated in FIG. The bilingual sentence here is:-terminal controller-terminal controller. The results of the first and second steps will typically be as shown in FIGS. 6 (a) and (b).
That is, there are three Japanese words and two English words.
In the third step, only the correspondence between "terminal" and "terminal" is identified, and the result shown in FIG. 6 (c) is obtained. Here, "control" and "apparatus", for which the correspondence could not be identified, are adjacent to each other, and if this is regarded as one compound word, the number of words in Japanese and English will be the same. Think like). In this way, the result of the third step is modified as shown in FIG. 6 (e), and the result of FIG. 6 (f) can be obtained in the fourth step. That is, the correspondence between the “control device” and the “controller” can be estimated.

In the present embodiment, the process of identifying the correspondence between words using the bilingual dictionary includes (1) identification of words constituting a Japanese sentence, (2) identification of words constituting an English sentence, 3) The processing is performed in the order of processing for searching pairs of Japanese words and English words from the bilingual dictionary. However, it does not have to be done in that order. For example, (1) identification of words constituting a Japanese sentence, (2) processing for obtaining a candidate for a translated word of a Japanese word by referring to a bilingual dictionary, and (3) candidate of a translated word in an English sentence It is also possible to perform the search in the order of processing.

Further, in the present embodiment, the co-occurrence relation of English words is registered in the co-occurrence relation dictionary. However, since the purpose of utilizing the co-occurrence relationship of English words is to select a translation at the time of translation, it is not necessary to register all of them. For each Japanese word, a bilingual dictionary is constructed so as to include information indicating the priority of the bilingual word, and the co-occurrence relation consisting of the first-rank bilingual word is registered in the English co-occurrence relation dictionary. It may be excluded from the target. Thereby, the capacity of the English co-occurrence relation dictionary can be reduced.

Further, in the present embodiment, when a Japanese compound noun exists, it is determined that there is a co-occurrence relationship between the words constituting the English noun phrase corresponding to the compound noun, and this is registered in the English co-occurrence dictionary. Will be. However, since compound nouns in Japanese and noun phrases in English can be identified by part-of-speech sequence patterns, it is conceivable to register their pairs in a Japanese-English bilingual dictionary. According to this method, a compound noun can be treated as one unit in the process of translating between Japanese and English in translation, so that the effect of reducing the load of translation processing can be obtained.

〔The invention's effect〕

According to the present invention, a function of learning a bilingual relation and a co-occurrence relation of words from a bilingual sentence can be realized. In a machine translation system, when a translated sentence obtained by performing post-editing on a translation result is considered as a pair with an input sentence, the translated sentence can be used constantly. Therefore, according to the present invention, a machine translation system having a self-propagating function can be realized.

[Brief description of the drawings]

FIG. 1 is a flowchart of a process of identifying a word correspondence from a bilingual sentence and a process of extracting a co-occurrence relationship of words in a second language according to an embodiment of the present invention, and FIG. 2 implements the present invention. FIG. 3 is a diagram showing an example of a dictionary record, FIG. 4 is a diagram showing an example of processing for a bilingual sentence, FIG.
FIG. 6 and FIG. 6 are explanatory diagrams of a modification of the processing for estimating the correspondence between words. 101-102... Identification processing of words constituting the first language sentence, 10
3 to 104: Identification processing of words constituting sentences in the second language, 105
~ 115 ... Processing for identifying correspondence between words in the first language sentence and words in the second language with reference to the bilingual dictionary, 116-120 ... Correspondence between words in the first language and words in the second language Estimating the relationship and registering it in the bilingual dictionary, 121... Extracting the co-occurrence relationship of words from sentences in the first language, 122-126. Registration processing for mapping and second language co-occurrence dictionary.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-62-297972 (JP, A) JP-A-63-5470 (JP, A) JP-A-59-165178 (JP, A) JP-A-63-297 44276 (JP, A) (58) Field surveyed (Int. Cl. ⁶ , DB name) G06F 17/20-17/28 JICST file (JOIS)

Claims (9)

(57) [Claims] 1. A bilingual dictionary creation method for a machine translation system having a first language dictionary, a second language dictionary, and a first language and a second language bilingual dictionary, wherein the first language dictionary is used. A first step of identifying a word constituting a sentence of the first language, and using a dictionary of the second language, a word constituting a sentence of the second language having a bilingual relationship with the sentence of the first language. A second step of identifying, and, for the combination of the first language word and the second language word identified in the first and second steps, respectively, determine whether a bilingual relationship can be established between the first language and the second language. A third step of determining using a bilingual dictionary to identify word-level correspondence, and examining whether any words for which correspondence could not be identified exist, and if so, the correspondence between those words And estimating in the fourth step The method of creating bilingual dictionary machine translation systems, characterized in that it comprises a fifth step of registering the correspondence relationship to the bilingual dictionary, a. 2. The method according to claim 1, wherein the step of estimating the correspondence between words is performed in a sentence of the first language or the second language, the correspondence of which is undetermined. If there is a continuous list of words, they are regarded as one compound word, and a step of estimating the correspondence between words is included. 3. The method for creating a bilingual dictionary of a machine translation system according to claim 1, wherein said first step and said second step include a step of selecting only a content word from the identified words. The third step is a method for creating a bilingual dictionary of a machine translation system, wherein the method is performed only on the selected content word. 4. A bilingual dictionary creation method for a machine translation system having a first language dictionary, a second language dictionary, and a first language and a second language bilingual dictionary, wherein the first language dictionary is used. And a first step of identifying words forming a sentence of the first language, and using the bilingual dictionary to select a candidate of a bilingual word in the second language for the word in the first language identified in the first step. A second step of searching; and a third step of searching for a bilingual word candidate obtained in the second step from a sentence in a second language having a bilingual relationship with the sentence in the first language to identify a word-level correspondence. A fourth step of examining whether or not there is a word for which a correspondence could not be identified in the third step, and estimating a correspondence between the words when there is, a correspondence estimated in the fourth step; Register relationships in the bilingual dictionary Bilingual dictionary creating a machine translation system which comprises 5 steps and. 5. A method according to claim 4, wherein the step of estimating the correspondence between words is performed in a sentence of the first language or the second language, the correspondence of which is undetermined. , A method for creating a bilingual dictionary for a machine translation system, which includes a process of estimating the correspondence between words when there are consecutive words. 6. A method for creating a bilingual dictionary for a machine translation system according to claim 4 or 5, wherein said first step includes a step of selecting only content words from the identified words,
A bilingual dictionary creation method for a machine translation system, wherein the second step and the third step are performed only on the selected content word. 7. A method for creating a co-occurrence relation dictionary of a machine translation system having a bilingual dictionary of a first language and a second language and a co-occurrence relation dictionary of a second language, comprising the steps of: A first step of identifying a word-level correspondence between a sentence in the first language and a sentence in the second language, which is a translation thereof, using a dictionary; Performing a second step of extracting co-occurrence relations of words included in the sentence, and the second step of utilizing the correspondence between the words of the first language and the words of the second language obtained in the first step. The first extracted in
A third step of mapping the co-occurrence relations of the words of the language to the co-occurrence relations of the words of the second language; and storing the co-occurrence relations of the words of the second language mapped in the third step in a second language co-occurrence dictionary. Registering a co-occurrence relation dictionary for a machine translation system. 8. A method for creating a co-occurrence relation dictionary for a machine translation system according to claim 7, wherein said bilingual dictionary stores information on the priority between bilingual words when one word has a plurality of bilingual words. Wherein the fourth step excludes a co-occurrence relationship obtained as a first-ranked pair of bilingual words from the registration target. . 9. A method for creating a co-occurrence relation dictionary for a machine translation system according to claim 7, wherein the co-occurrence relation of a second language word obtained corresponding to a compound word in a sentence of the first language is the first language. A method for creating a co-occurrence relation dictionary for a machine translation system, comprising registering in a bilingual dictionary of a first language and a second language together with the compound word of the language.