JP3045832B2 - Relevant word selection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a machine translation system that executes a translation process using a computer. The present invention relates to a suitable translation word selection method for selecting a suitable translation word most suitable as a translation word from candidates.

[0002]

2. Description of the Related Art In recent years, many machine translation systems have been commercialized. How to select and translate an appropriate translation word from a plurality of translation word candidates greatly affects the quality of a translation.
The method of selecting a suitable word can be broadly divided into two methods: a method of judging from semantic information such as a subject, an object, and a predicate, and a method of translating according to the field of the text to be translated. It is thought that the former is effective for novels and letters, and the latter is effective for manuals and scientific and technical sentences.

However, the current machine translation targets (targets) are mainly manuals and scientific and technical sentences, so it is important to enhance the latter method in order to improve the quality of translated sentences.

By the way, in the latter method, a method of preparing a translation word for each field in a dictionary and designating the field and translating the data is generally used. However, there is a problem that the field must be specified in advance, There was a problem that the translation was changed to a translation for each field, even if the translation was good.

To solve this problem, Japanese Patent Laid-Open Publication No. Sho 63-29880 proposes a method of translating an appropriate word using a thesaurus dictionary. In this method, for example, when translating the word “Hane” into English in a Japanese-English translation, if the translation word “airplane” was used before, based on a thesaurus dictionary, a translation candidate of “Hane” is obtained. ("Feat
her ”,“ wing ”...),“ wing ”, which is a lower word of the“ airplane ”, is selected as a suitable translation.

[0006]

However, in the case of the conventional method using a thesaurus dictionary, a great deal of labor is required for constructing the thesaurus dictionary, and at the same time, knowledge of a specialized field is required, and thesaurus dictionary construction is required. , The production cost of the machine translation system becomes expensive due to the time and labor required.

[0007] Further, unless words to be handled in the thesaurus, such as upper words, lower words, and synonyms, do not appear, the thesaurus dictionary itself cannot be effectively used, and as a result, an appropriate translated word is selected. There was a possibility that the performance would be reduced.

The present invention has been made in view of the above circumstances, and can solve the problem that the production cost of a machine translation system becomes expensive due to the labor required for dictionary construction and the difficulty of dictionary construction work. Moreover, even if a semantic related word such as a high-order word or a low-order word does not appear in the text to be translated, it can be appropriately translated into the appropriate word, thereby improving and stabilizing the translation quality. The purpose is to provide a suitable word selection method.

[0009]

In a machine translation system that executes a translation process using a computer, a suitable translation word selection method according to the present invention includes a plurality of translation word candidates for an input original phrase. In this case, a suitable translation word most suitable as a translation word is selected from the plurality of translation word candidates.

Specifically, the machine translation system includes:
A co-occurrence dictionary is provided, and a proper translation word selection unit that executes a proper translation word selection process for selecting a proper translation word from a plurality of translation word candidates using the co-occurrence dictionary is provided.

In addition, the co-occurrence dictionary sets, as a co-occurrence word, a phrase that appears together with the head word with respect to the word in the target language set as the head word, and Are described for each co-occurrence word, indicating an evaluation value K indicating the degree of appearance of the word together with the translated sentence.

Further, in the translation word selection processing in the translation word selection unit, a co-occurrence word for the translation word candidate is checked for each translation word candidate using the co-occurrence dictionary, and the co-occurrence word itself is further translated. It is checked whether or not the co-occurrence-related word is present in the original text to be translated. If a co-occurrence related word exists, the co-occurrence word is described in the co-occurrence dictionary in terms of the appearance position and the number of appearances of the co-occurrence related word with respect to the phrase Tj in the original text to which the appropriate translation is to be added. In consideration of the evaluation value K of the co-occurrence word, the evaluation value Vi indicating the degree to which the translation word candidate is used as the appropriate translation word is calculated for each translation word candidate i.

Then, from among a plurality of translation word candidates for the phrase Tj in the original text, a translation word candidate i having the maximum evaluation value Vi is selected as a suitable translation word for the phrase Tj.

[0014]

According to the suitable word selection method of the present invention, a suitable word selection process for selecting the best suitable word as a translation from a plurality of candidate words uses a co-occurrence dictionary and does not use a thesaurus dictionary.

Moreover, the co-occurrence dictionary itself can be created by simply arranging the appearance positions and the number of appearances without taking into account the semantic relationship between the words appearing in the sentence in the target language. .

Therefore, as compared with the conventional method using a thesaurus dictionary, it is possible to solve the problem that the production cost of the machine translation system becomes expensive due to the labor required for the dictionary construction and the difficulty of the dictionary construction work. Can be translated into appropriate translations even if no semantically related words such as high-level words and low-level words appear in the text to be translated, thereby improving the translation quality and stabilizing the translation. It becomes possible.

[0017]

FIG. 5 is a block diagram showing an apparatus configuration of a machine translation system which employs an embodiment of a suitable word selection system according to the present invention. As shown in the figure, this machine translation system includes an input unit 1, a translation unit 2, a suitable word selection unit 3, an output unit 4,
It comprises a word dictionary 5, a co-occurrence dictionary 6, and the like.

The input unit 1 is for inputting a text to be translated (original text) by an OCR, a keyboard, or the like, and sends the text input to the input unit 1 to the translation unit 2.

The translation unit 2 and the appropriate word selection unit 3 are units that execute predetermined translation processing in the computer main body (CPU). The translation unit 2 translates the input sentence into a target language using the word dictionary 5, and sends a result and a candidate word to the appropriate word selection unit 3. In addition, the translation word selection unit 3
In the part for executing the suitable word selection process according to the present invention, the translated word of the translation unit 2 is corrected to the suitable word using the co-occurrence dictionary 6, and the result (translated sentence) is sent to the output unit 4.

The output unit 4 outputs a translation to a CRT, a printer, or the like. The word dictionary 5 describes information on words in the source language and the target language, and the co-occurrence dictionary 6 includes a word in the target language (entry word in FIG. 2) and another word (co-occurrence word in FIG. 2). Are described as the evaluation value K (see FIG. 2) for each co-occurrence word. The word dictionary 5 and the co-occurrence dictionary 6 are used by being arranged on an external storage device or a main memory managed by a CPU.

Next, the structure close to one embodiment will be described in detail.

First, the configuration of the co-occurrence dictionary 6 will be described. The co-occurrence dictionary 6, as shown in FIG.
It is composed of a co-occurrence word 9 and an evaluation value K for the co-occurrence word 9. Here, the headword 8 is an arbitrary word in the text in the target language, and the co-occurrence word 9 is a word that appears on the text together with the headword 8. Headword 8 and co-occurrence 9
May be extracted within a single paragraph (paragraph) of a series of sentences, or the entire series of sentences (composed of multiple paragraphs).
By using a plurality of sentences (a large amount of sentences) that are common in the technical field as a series of sentences to create a co-occurrence dictionary, it is conceivable to improve the reliability of the appropriate word selection process. The words set as the headword 8 and the co-occurrence word 9 may be limited to a specific part of speech, or may be all parts of speech.

The evaluation value K indicates the degree to which both the word set as the headword 8 and the word set as the co-occurrence word 9 appear in one sentence. 1
/ (The number of words between the headword and the co-occurrence word + 1)}. In other words, the evaluation value K is calculated for all words appearing in a certain sentence in the target language.
The appearance degree is obtained according to the appearance position, the number of times, and the like.

The co-occurrence dictionary having the above-described structure is obtained by simply rearranging the appearance positions and the number of appearances of words appearing in a sentence in a target language without taking into account semantic association. Can be created. Therefore, as compared with the case of a thesaurus dictionary which has to be created in consideration of the semantic relationship between words, the labor required for creating the dictionary is reduced and the creation processing work is facilitated.

In the co-occurrence dictionary 6 shown in FIG. 2, the sentence “Machine translation system... Greatly affects the quality of the translated sentence. Thus, the parts of speech of the words to be the headword 8 and the co-occurrence word 9 are limited to nouns and verbs.

FIG. 3 shows the structure of a word dictionary 5 for English-Japanese translation in a certain technical field. The word dictionary 5 has the same configuration as that of the conventional word dictionary, and includes a headword 11, a part of speech 12, a candidate word 13 for translation, and the like.

FIG. 4 shows a co-occurrence dictionary 6 created by analyzing in advance a Japanese sentence in a technical field (or a translation that has been previously translated) in order to translate an English sentence in a technical field into a Japanese sentence. This is an example. As in FIG. 2,
It is composed of a headword 8, a co-occurrence word 9, an evaluation value K, and the like, and the corresponding technical field is common to the word dictionary 5 of FIG.

Next, "The solar system ... by comput
er system. ”(In addition, there are 10 words in the“ ... ”part of the English sentence on the left, and those 10 words include“ solar ”,“ system ”,“ computer ”, etc. The processing operation of the machine translation system shown in FIG. 5 will be described by taking as an example the case of translating an English sentence (which assumes that no word is used) into a Japanese sentence.

For translation of the English sentence, the word dictionary 5 shown in FIG. 3 and the co-occurrence dictionary 6 shown in FIG. 4 are used.

FIG. 1 shows a series of translation processing procedures in the machine translation system of FIG. 5, and based on FIG. 1, a suitable word selection method of one embodiment will be described in detail.

First, the input unit 1 sends the input English sentence to the translation unit 2 (step 100). Then, the translation unit 2 uses the word dictionary 5 of FIG. 3 to select a translated word for each word (original language) in the input English sentence,
Perform translation processing to generate candidate translations for each source language,
The processing result is output to the appropriate word selection unit 3 (step 20).
0).

FIG. 6 shows the result of the translation processing in the translation section 2. In FIG. 6, an original word 15 is a word in the input sentence to be translated (the above-described sentence), a translated word 16 is a translated word assigned by the translation unit 2 based on the word dictionary 5, and a translated word candidate 17 is a word dictionary for the original word 15. This is the translation word candidate described in No. 5.

In the case of this specific example, in the translation processing in the translation unit 2, "solar" contains "sun" and "syst" of the translation word candidate 1.
For “eme”, the “organization” of the translation candidate 1 is also selected as a translation. That is, in the processing in the translation unit 2, the original text is translated as "the solar organization is a computer organization ...".

The processing after step 300 in FIG. The appropriate word selection unit 3 performs a process of selecting an appropriate word from candidate words for each word in the original sentence. First, an initial value 1 is set to the parameter j (step 300). The parameter j indicates the order of the word in the original sentence to be processed. For example, if j = 1, the word Tj to be processed is “so” located first in FIG.
lar ”.

Next, the parameter i is set to an initial value 1 (step 301). The parameter i is used to select a translation word from among translation word candidates of the processing word Tj specified previously. When i = 1, the first word in the translation word candidate field shown in FIG.
This means that the second translated word (translated word candidate 1) has been selected. Therefore, if j = 1 and i = 1, the translation selected at that time is “sun”.

Next, the word T set in the above process is set.
The co-occurrence dictionary 6 shown in FIG. 4 is searched for the translated word candidate i of j (step 302). In this case, "solar"
As a result of searching the co-occurrence dictionary 6 for the “sun” of the translation word candidate 1, “battery (evaluation value K = 6.3)”
And “system (evaluation value K = 4.0)”.

Next, for each co-occurred word, it is checked whether or not an original word (hereinafter referred to as a co-occurrence related word) having the word as a candidate for a translated word appears in the same paragraph (or paragraph). The evaluation value Vi of the translation word candidate i is calculated from the number of times, the appearance position, and the like (step 303).

In this example, a co-occurrence-related word having the co-occurrence word “battery” as a translation candidate does not appear in the original text. But,
As a co-occurrence-related word having the co-occurrence word “system” as a candidate for translation, “system” appears in two places in the original text.

The evaluation value Vi of the translation word candidate i is K, the machine evaluation value of a co-occurrence word for the translation word candidate i, a co-occurrence-related word having the co-occurrence word as a translation word candidate, and the processing word Tj.
, The individual evaluation value Pn = {K / (Z + 1)} is calculated for each co-occurrence-related word appearing in the original sentence, and the sum of the calculated individual evaluation values Pn and I do. However, the subscript n in the individual evaluation value Pn is a parameter indicating the order of appearance of co-occurrence related words.

In this example, there is no co-occurrence related word with "battery" as a candidate for translation, but as described above, "system" is twice as a co-occurrence related word with "system" as a candidate for translation. Since it has appeared, the above-described individual evaluation value p is calculated for each “system”, and the sum thereof is obtained, whereby the evaluation value Vi for the translation word candidate “sun” is obtained.

The individual evaluation value P1 for the first appearing co-occurrence related word "system" is Z = 0 and K = 4.0, so that P1 = 4 / (0 + 1) = 4.

The individual evaluation value P2 for the second co-occurrence related word "system" is Z = 13 and K = 4.0, so that P2 = 4 / (13 + 1) = 0.29.

Therefore, the evaluation value Vi of the translation word candidate “sun” of the processing word “solar” is given by Vi = P1 + P2 =
It becomes 4.29.

When the calculation of the evaluation value Vi of the translation word candidate i is completed, it is determined whether or not there is another translation word candidate for the word Tj based on the processing result of the translation unit 2. If so, 1 is added to the value of the parameter i to set the next candidate word, and the process returns to step 302 (steps 304 and 305).

In the case of this example, by setting i = 2, “solar” of the translation word candidate 2 for the original word “solar” is to be processed in step 302. In the case of the translation word candidate 2 “solar”, as a result of searching the co-occurrence dictionary 6, “car (evaluation value K = 3.1)” is obtained as a co-occurrence word. Then, in the following step 303, the existence of a co-occurrence related word having “car” as a candidate word is checked, and the evaluation value Vi for the candidate word “solar” is determined according to the presence / absence. In this example, Since there is no word in the original sentence with “car” as a translation candidate, the evaluation value Vi of the translation candidate “solar” for the original “solar” is Vi = 0.

If it is determined in step 304 that the processing of steps 302 and 303 has been completed for all the translation word candidates of the processing word Tj, the evaluation values Vi of the respective translation word candidates of the processing word Tj are compared with each other. The translation word candidate having the maximum evaluation value Vi is set as a suitable translation word m (step 306).

In the case of the above example, the evaluation value Vi of the translation word candidate 1 (sun) is 4.29 and the evaluation value Vi of the translation word candidate 2 (solar) is 0 for the original word "solar". “Sun” of translation candidate 1 with the maximum Vi is the appropriate translation m
Will be recognized.

Next, the translation result of the translation unit 2 is checked, and the translated word is translated into the original word "solar" in step 30.
It is determined whether or not the appropriate word m obtained in step 6 is used (step 307). If the appropriate word m is not used, the translated word is changed to the appropriate word m (step 308).

Next, if the processing word Tj still remains, the value of the parameter j is incremented by 1 and step 301
(Steps 309, 310).

In this example, by executing steps 301 to 308 with j = 2, the original language “syste
The translation for "m" can be corrected to the appropriate translation. Similarly, every time the value of j is further increased by 1, the steps 301 to 301 are executed.
By repeating 308, the translated words processed by the translating unit 2 are sequentially corrected to appropriate translated words.

If it is determined in step 309 that the processing in steps 301 to 308 has been completed for all the processed words Tj, the translated sentence corrected with the appropriate word m is output to the output unit 4. Translation word selection part 3
Is completed (step 40).
0).

Incidentally, a supplementary explanation of the above-mentioned specific example is as follows. In the suitable word selection processing for the first "system" in the original text, the following result is obtained. "Organization" of translation candidate 1
The co-occurrence word for "reform (evaluation value K = 14.4)"
And "reform (evaluation value K = 12.8)", co-occurrence-related words for these co-occurrence words do not appear in the original text. Therefore, the evaluation value Vi for the translation word candidate “organization” is 0.

The co-occurrence words of the translation word candidate 2 for “system” are “translation (evaluation value K = 7.7)”, “development (evaluation value K = 6.6)”, and “computer (evaluation value K). =
5.2) ", and since" computer "exists as a co-occurrence term for the co-occurrence word" computer ",
The evaluation value Vi for the translation candidate “system” is expressed as Vi =
5.2 / (11 + 1) = 5.2 / 12 = 0.43.

The co-occurrence words of the translation word candidate 3 for “system” are “sun (evaluation value K = 4.0)” and “galaxy (evaluation value K =
3.9) ”, and since“ solar ”exists as a co-occurrence-related word for the co-occurrence word“ sun ”, the evaluation value Vi of the translation candidate“ system ”is Vi = 4.0 / 1 = 4.00.

Then, by comparing the evaluation values Vi of the respective translated word candidates with each other, “system” is selected as a suitable translated word for the first “system”.

In the case of “by” and “computer”, there is only one translation word candidate, and there is no choice. Therefore, the translation word given by the translation unit 2 is left as a proper translation word.

In the case of the second "system", the first "sy"
In the same processing procedure as in the case of “stem”, the translation word candidate “system (evaluation value Vi = 5.
2) "is selected as the appropriate word.

From the above processing results, "The sun is
The translation sent by the computer system is transmitted to the output unit 4.

In the suitable word selection method of one embodiment described above, to select the best suitable word as a translation from a plurality of candidate words, first, the co-occurrence dictionary 6 is used to select each suitable word. First, the co-occurrence word corresponding to the candidate word is checked, and it is further checked whether or not a co-occurrence related word having the co-occurrence word itself as a candidate word exists in the original text to be translated.

When a co-occurrence related word exists,
In consideration of the appearance position and the number of appearances of the co-occurrence-related word in the original text and the evaluation value K of the co-occurrence word with respect to the phrase Tj in the original text to which the appropriate translation word is to be given, the translation candidate is used as the appropriate translation word. An evaluation value Vi indicating the degree is obtained for each translated word candidate i.

Then, from among a plurality of translation word candidates for the phrase Tj in the original text, a translation word candidate i having the maximum evaluation value Vi is selected as a suitable translation word for the phrase Tj.

In other words, the suitable word selection method of one embodiment does not use a conventional thesaurus dictionary, but uses a co-occurrence dictionary and uses evaluation values K and Vi representing the degree of co-occurrence. The co-occurrence dictionary prepared in the machine translation system in advance is used to select the translation word. The co-occurrence dictionary itself simply determines the location of the words appearing in the sentence in the target language without considering the semantic relation. And by organizing the number of appearances. Therefore, as compared with the case of a thesaurus dictionary which has to be created in consideration of the semantic relationship between words, the labor required for dictionary construction is reduced and the dictionary construction processing operation is facilitated.

Therefore, as compared with the conventional method using a thesaurus dictionary, it is possible to solve the problem that the production cost of the machine translation system becomes expensive due to the labor required for the dictionary construction and the difficulty of the dictionary construction work. Can be translated into appropriate translations even if no semantically related words such as high-level words and low-level words appear in the text to be translated, thereby improving the translation quality and stabilizing the translation. You can also.

[0064]

According to the suitable word selection method according to the present invention, a suitable word selection process for selecting a suitable word as a target word from a plurality of candidate words uses a co-occurrence dictionary and does not use a thesaurus dictionary. Moreover, the co-occurrence dictionary itself can be created by simply rearranging the appearance positions and the number of appearances without taking into account the semantic relationship between words appearing in a sentence in the target language.

Therefore, as compared with the conventional method using a thesaurus dictionary, it is possible to solve the problem that the production cost of the machine translation system becomes expensive due to the labor required for the dictionary construction and the difficulty of the dictionary construction work. Can be translated into appropriate translations even if no semantically related words such as high-level words and low-level words appear in the text to be translated, thereby improving the translation quality and stabilizing the translation. It becomes possible.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an appropriate word selection process in one embodiment of the present invention.

FIG. 2 is a configuration example of a co-occurrence dictionary according to an embodiment of the present invention.

FIG. 3 is a diagram illustrating a configuration of a word dictionary according to an embodiment of the present invention.

FIG. 4 is a diagram illustrating the configuration of a co-occurrence dictionary according to an embodiment of the present invention.

FIG. 5 is a schematic configuration diagram of a machine translation system that executes an embodiment of the present invention.

FIG. 6 is an explanatory diagram of a translation process according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input part 2 Translation part 3 Appropriate translation word selection part 4 Output part 5 Word dictionary 6 Co-occurrence dictionary 8,11 Headword 9 Co-occurrence word K, Vi Evaluation value 12 Part of speech 13,17 Translation candidate 15 Original language 16 Translation word

────────────────────────────────────────────────── ─── Continuation of front page (56) References JP-A-3-15764 (JP, A) JP-A-3-37763 (JP, A) JP-A-3-206558 (JP, A) Masaki Matsudaira, 2 Name, "Creation of word semantic network using co-occurrence data", Information Processing Society of Japan 42nd (Early 1991) National Convention, p. 3-202-3-203 (58) Field surveyed (Int. Cl. ⁷ , DB name) G06F 17/28 JICST file (JOIS)

Claims (1)

(57) [Claims] In a machine translation system that executes a translation process using a computer, when there are a plurality of candidate translations for an input word of an original sentence, a translation is selected from the plurality of candidate translations. A translation word selection method for selecting the most suitable translation word, wherein the machine translation system uses the co-occurrence dictionary and a translation word selection process of selecting a translation word from a plurality of translation word candidates. The co-occurrence dictionary is equipped with a suitable translation word selection unit to execute, and for a phrase in the target language set as a headword, a word that appears together with the headword is set as a co-occurrence word, In addition, the headword and its co-occurrence word have a configuration in which an evaluation value K indicating the degree of appearance in the translated sentence is described for each co-occurrence word. , Previous The co-occurrence dictionary is used to check for a co-occurrence word for each translation word candidate, and whether a co-occurrence-related word whose co-occurrence word itself is a translation word candidate exists in the original text to be translated. If there is a co-occurrence related word, if the co-occurrence related word exists, it is described in the co-occurrence dictionary, where the co-occurrence related word appears in the original sentence with respect to the word Tj in the original text to which the appropriate translation is to be added. In consideration of the evaluation value K of the co-occurring word, the evaluation value Vi indicating the degree to which the translation word candidate is used as the appropriate translation word is calculated for each translation word candidate i, and a plurality of translation word candidates for the phrase Tj in the original text are calculated. From among them, the translation word candidate i having the maximum evaluation value Vi is represented by the phrase Tj
A suitable word selection method characterized by selecting as a suitable word.