JPH0668134A - Translation system using translation usage base - Google Patents

Abstract

PURPOSE:To attain natural translation comparatively simply and to improve the translation accuracy of even a routine sentence whose translation is difficult or a conversational sentence having many omission. CONSTITUTION:Part of a basic usage database 18 is modified and a modification usage database 20 having a modification identifier is generated for the part. The modification usage database 20 and an input sentence from an original text input section 12 are collated by a collation processing section 22. When the collation of the both is failed, a translation main processing section 26 translates the input sentence by the transfer system. On the other hand, when the collation is successful, a modified usage in which other parts than the modification identifier are coincident with the input sentence is extracted from the modification usage database 20 and fed to a usage base translation generation processing section 24. A variable identifier equivalent part is translated by a translation main processing section 26 and the translation of the variable identifier equivalent part obtained in this way is imbeded in the usage translation and the resulting translation is outputted from a translation output section 28.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a translation example-based translation method, and more particularly to a translation example-based translation method used in a machine translation system for translating conversational sentences and business documents.

[0002]

2. Description of the Related Art In a conventional machine translation system, all input sentences are translated through each process of analysis processing, conversion processing and sentence generation processing.

[0003]

However, if such a process is performed on all sentences, it is difficult to translate greetings and fixed phrases, and unnatural translated sentences are generated, resulting in an overall translation accuracy. Was causing the drop. For example, a sentence like (1) below should be translated into a translated sentence like (2), but in the past, it would be translated into a translated sentence like (3).

English: Give me extension 2318. (1) Translated text: Extension 2318 (2) Translated text: Give me the extension 2318. (3) Therefore, a method of preparing a translation template may be considered in order to improve the translation accuracy, but enormous work is required to create the template.

Further, as shown in Japanese Patent Laid-Open No. 3-276367, examples that can be collected are stored in an example database, and the example that is most similar to the input original sentence is searched from the example database. Some translate the input language according to the translation of the example. However, in this technique, similarity judgment is difficult, it is difficult to deform in response to such tense, also it is difficult to determine may be collecting what text attracted much the examples collection, further, similar There was a problem that the search time for judging the degree of time becomes huge.

Therefore, a main object of the present invention is to provide a translation example-based translation method that can be searched relatively easily.

[0007]

SUMMARY OF THE INVENTION The present invention includes a basic example database and a modified example database having a variable identifier in a part of the basic example database and having a variable identifier in the part, and collates an input sentence with the modified example database. Then
A modified example in which other parts of the modified example database other than the variable identifier match the input sentence was taken out from the modified example database, the variable identifier part was translated, and the translated sentence was generated by assigning it to the variable identifier. It is an example-based translation method.

[0008]

Operation: A modified example database is generated from the basic example database. Then, the input sentence is collated with the transformation example database, and when there is no transformation example that matches the input sentence, the input sentence is translated by the transfer method. On the other hand, if there is a modified example that matches the input sentence, the part corresponding to the variable identifier is extracted from the input sentence, that part is translated by the transfer method, and the obtained translated word is embedded in the translated sentence of the modified example and translated. Complete the sentence. If the grammatical attribute of the variable identifier is described, the grammar attribute is also used as the matching condition, and the grammatical attribute is taken into consideration for translation by the transfer method.

[0009]

According to the present invention, by using the modified example database, natural translation can be performed relatively easily, and translation accuracy can be improved even for fixed sentences that are difficult to translate and conversational sentences that are often omitted. You can The above-mentioned objects, other objects, features and advantages of the present invention will become more apparent from the following detailed description of the embodiments with reference to the drawings.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Referring to FIG. 1, an English-Japanese machine translation system 10 of this embodiment adopts a translation example-based translation method and includes an original text input section 12 for inputting text.
The text input to the source text input unit 12 is the text processed by the translation processing unit 1.
4 is given to the one-sentence cutout unit 16 and is cut out for each sentence.

Further, the translation processing unit 14 includes a basic example database 18 for storing basic example data such as parallel translation example collections and past translation results in a hard disk, for example, and a modification example including a modification identifier by a modification example generation function. A modified example database 20 for storing data is generated. The basic example database 18 and the modified example database 20 are recorded in a state in which the correspondence between Japanese and English sentences is clear. The transformation example generation function can automatically convert numbers, proper nouns, and the like into transformation identifiers, but it is a manual task for languages that are difficult to convert.

Then, the text for each sentence from the one-sentence cut-out unit 16 and the transformation example database 20 are collated by the collation processing unit 22, and if the collation fails, the translation main processing unit 26 is activated. After the morpheme analysis / syntactic analysis / semantic analysis analysis and conversion processing by the transfer method, a translated sentence is generated, and the translation output unit 28
Is output from. A translation engine 36 including an original sentence / syntactic analysis dictionary 30, a translation conversion dictionary 32, and a translated sentence generation dictionary 34 is used for the translation processing in the translation main processing unit 26. On the other hand, if the collation processing unit 22 succeeds in collation, the example-based translation generation processing unit 24 is activated, and with the help of the translation main processing unit 26 for the modified identifier, an appropriate translation sentence corresponding to the input sentence is generated. .

With respect to such an English-Japanese machine translation system 10, a modification example generation function for creating a modification example database 20 from the basic example database 18 will be described. First, in order to generate a modified example, English text and corresponding Japanese text are extracted for each sentence. Here, consider the case where there is a basic example as shown in FIG.

First, in order to find the correspondence between English and Japanese words, attention is paid to the numerical expression as the first step. That is, it is detected whether or not a numerical expression is present in a given English text. If there is a numerical expression in the English text, in most cases, the value corresponding to that numerical value should also exist in the Japanese text. Search the Japanese text and find the correspondence. Find out.

In the example shown in FIG. 2, as shown in FIG.
Correspondence can be made between two numerical expressions of "44" and "66.89". Since such a numerical expression is not used elsewhere in the Japanese text, it can be judged that this correspondence is established. The modification example database 20 is automatically generated based on this correlation information. Since the numerical expression is considered to be variable, the modified example database 20 replaces the numerical expression as a variable identifier. Here, the variable identifiers are replaced with # Num1, # Num2, ... In the order of appearance so that the variables can be identified even if they exist. Therefore, in this example, the modified example of FIG. 4 can be obtained from the basic example of FIG.

Here, the numerical expression has been described as the first step, but similarly, with respect to proper nouns and pronouns as well, the correlation correspondence between English and Japanese is checked,
Replace those that can be dealt with with variable identifiers,
The modification example database 20 is registered. The above processing is performed on all the basic example databases 18 to obtain the modified example database 20 for each sentence. When there is no part to be replaced with the variable identifier, the basic example is a modified example.

Next, the translation processing operation will be described with reference to FIG. First, in step S1, the collation processing unit 22 collates the input sentence with the modification example database 20. At this time, if the grammatical attribute of the variable identifier is described as in a translation example frame described later, the matching of the grammatical attributes is set as the matching establishment condition.

Then, in step S3, the input sentence is collated with the modification example database 20. If there is no transformation example in which the input sentence matches the entire sentence or the portion except the conversion identifier, the entire input sentence is translated by the transfer method in step S5, and the process ends. On the other hand, when there is a modified example that matches the input sentence in step S3, the variable identifier equivalent part corresponding to the variable identifier of the modified example is extracted from the input sentence in step S7, and the variable identifier equivalent part is extracted in step S9. Is translated by the transfer method. That is, in step S11, analysis processing such as morphological analysis, syntactic analysis and semantic analysis is performed. At this time, if the syntactic form condition of the variable identifier is described in the translation example frame, the syntactic analysis is executed according to the condition. Then, after the conversion process is performed in step S13, a translated sentence generation process is performed in step S15. At this time, if the connection example condition of the variable identifier is described in the translation example frame, the translated word is transformed according to the condition. Then, in step S17, the translated word of the obtained variable identifier equivalent portion is embedded in the translated text of the modified example, and the process ends.

Here, consider the case where the following English text is input. Total amount is JPY 7,800, in USD 60, tax and serv
ice charge included. First, in step S1, the modification example database 2
Match 0 with the entered English text. Here, by adopting a matching method in which the numerical expression and #Num match the pattern match, the input English text and the transformation example database 20 are successfully matched, and in step S3, match the English text. It is determined that there is a modified example to be performed. In this case, as shown in FIG. 6, since the whole sentence is compared, the numerical expression can be recognized as a variable identifier,
Proceeding to step S7, the variable identifier corresponding part is 7,80.
Extract 0 and 60. Here, the numerical expression is not limited to a mathematical number, and may be a number spelled out in English. Then, in step S9, the translation main processing unit 26 is activated to translate 7,800 and 60, and the translated word is obtained. Then, the translation results of 7,800 and 60, which are variable identifier equivalent parts, are substituted into the Japanese generation pattern shown in FIG. 4 to finally complete the translated sentence (output translated text) as shown below.

Total: 7800 yen (60 US $) tax,
Including service charge In this way, the operation of the English-Japanese translation system 10 is in a form in which the processing using the modified example and the translation processing by the transfer method are integrated. A translation process using a modified example is executed for a sentence requiring free translation, and a translation process is executed for a general sentence.

Next, a grammar attribute is added to the variable identifier, and the variable identifier is compared with the grammatical attribute of the variable identifier equivalent part of the input sentence.
The case of generating a translated sentence will be described. First, consider the following example sentence translation process. Example sentence 1) English: May I speak to Mr. Yamato, please? Translated by: Mr. Yamato.

Example sentence 2) English sentence: This is Howard Smith. Translated sentence: Yes, this is Howard Smith. Example sentence 3) English: Would you mind waiting for a few min
Translate: Would you mind waiting for a while? All the example sentences are colloquial expressions, and it is difficult to properly translate them by the transfer-based translation process. Therefore, consider the use of the following translation example.

Example 1) May I speak to ~, please? ⇒ Thank you. Example 2) This is ~. ⇒ Yes, it is ~. Example 3) Would you mind ~ ing? ⇒ Could you do it? By comparing this example original sentence with the input sentence, translating only the part corresponding to in the corresponding example original sentence, and embedding the translated word in the example translated sentence, appropriate translation can be easily performed.

However, for example, the pattern of Example 2 matches a sentence such as This is a book. This is a book. ⇒ Yes, it is a book. There is a risk that it will be translated like. Therefore, in this case, the condition that the part of ~ is a person's name is necessary.

Furthermore, in Example 3, the word form change is required when embedding the translated word in the portion of. In other words, considering the case of example sentence 3, it would be Would you mind [waiting for a few minutes]? Just by translating the part corresponding to ~ normally and embedding it in the example translation.
[Wait for a while] Could you do that? Will result in a translation like. Therefore, in this case, the part of ~ is the ing form of the verb, and it is necessary to add the connection particle "te" when embedding the translated word.

Therefore, in the present invention, a translation example is described by a frame structure or the like. The translation example frame is basically a set of attribute name and its value pairs, as shown below. (Example frame for translation) (English: text) (Translation: text) (Variable identifier attribute frame) (Attribute name 1: value) (Attribute name 2: value) Describe. The parallel translation pattern can include a variable identifier. If the grammatical condition for the variable identifier is required, describe it in the "variable identifier attribute frame". The variable identifier attribute frame expands according to the number of variable identifiers. If the grammatical condition is unnecessary, the variable identifier attribute frame can be omitted. Therefore, the above-mentioned translation example is described by this framework as the following translation example frame.

(Translation example frame 1) (English: May I speak to # 1, please?) (Translation: Please indicate # 1.) (Variable identifier attribute frame) (Identifier: # 1) (English syntax form: (Noun) (Semantic Feature: Human) (Translation Example Frame 2) (English: This is # 1.) (Translation: Yes, # 1.) (Variable Identifier Attribute Frame) (Identifier: # 1) (English Syntax: (Noun) (Semantic feature: human) (Translation example frame 3) (English: Would you mind # 1?) (Translation: Can I have # 1?) (Variable identifier attribute frame) (Identifier: # 1) (English syntactic form) : Verb ing) (Japanese conjunctive form: continuous form) (Japanese conjunctive particle: te) Here, referring to FIG. 5, the grammatical attribute of the variable identifier is the matching condition between the input sentence and the modified example database 20. The operation in the case will be described.

First, in step S1, the input sentence is compared with the original example sentence. At this time, the matching error is prevented by using the grammatical attribute of the variable identifier described in the translation example frame as the matching condition. For example, in the case of the translation example frame 2 described above, this is a book. Does not match due to the condition that the variable identifier has the semantic feature = human. This is because the semantic feature of "book" is not human. Therefore, in this case, it is determined that there is no matching example in step S3 shown in FIG. 5, and the whole sentence is translated by the transfer method in step S5.

On the other hand, if there is a matching example in step S3, the variable identifier corresponding part is extracted from the input sentence in step S7, and the variable identifier corresponding part is translated by the transfer method in step S9. Here, if necessary, the grammatical attribute of the variable identifier described in the translation example frame is referred to. For example, the above example sentence 3
When translating the English sentence of, in the translation of waiting for a few minutes, wait for a while ⇒ wait for a while, depending on the condition The transformation is performed.

Finally, in step S17, the translation is completed by embedding the translation result of the obtained variable identifier in the example translation. The translation result of each of the above example sentences is as follows. However, the translated text a is according to the present invention, the translated text b is the one in which [] is independently translated and embedded in the example translated text, and the translated text c is based on the transfer method.

Example sentence 1) English sentence: May I speak to [Mr. Yamat
o], please? Translated text a: [Mr. Yamato] Translation b: [Mr. Yamato] Translated c: Can I speak with Mr Yamato? Example sentence 2) English sentence: This is [Howard Smith]. Translation a: Yes, [Howard Smith].

Translated text b: Yes, [Howard Smith]. Translation c: This is [Howard Smith]. Example sentence 3) English sentence: Would you mind [waiting for a few m
Translated text a: [Wait a few minutes] Could you please? Translation b: [Wait a few minutes] Can you please?

Translated c: Do you mind waiting for a few minutes? In this embodiment, the description content can be intuitively understood, the descriptiveness of the translation example frame that can deal with the exchange of words and phrases is realized, and the stored grammatical attribute is used as the matching condition between the input sentence and the example original sentence. Mismatching can be prevented. Furthermore, according to the stored grammatical attribute, the variable identifier equivalent part extracted from the input sentence is translated and the word form is changed by a transfer method or the like, and embedded in the example translated sentence, so that the unnaturalness of connection of the embedded translated sentence can be eliminated.

In the above embodiment, the case of English-Japanese translation has been described, but it goes without saying that other language pairs may be used.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is an illustrative view showing one example of a basic example database.

FIG. 3 is an illustrative view showing a correspondence relationship of numerical expressions.

FIG. 4 is an illustrative view showing one example of a modification example database.

FIG. 5 is a flowchart showing the operation of this embodiment.

FIG. 6 is an illustrative view showing a comparison result.

[Explanation of symbols]

 10 ... English-Japanese machine translation system 12 ... Original sentence input unit 14 ... Translation processing unit 18 ... Basic example database 20 ... Modified example database 22 ... Collation processing unit 24 ... Example base translation generation processing unit 26 ... Translation main processing unit 28 ... Translation output Part 36 ... Translation engine

Claims (2)

[Claims] 1. A basic example database and a modification example database having a variable identifier in a part of the basic example database and having a variable identifier in the part, the input example is collated with the modification example database, and the modification example database is included. The modified example in which the other part except the variable identifier matches the input sentence is taken out from the modified example database, the part of the variable identifier is translated, and then the translated sentence is generated by allocating to the variable identifier. , Translation example-based translation method. 2. A grammatical attribute is determined for the variable identifier, and when a part of the input sentence corresponding to the variable identifier matches the grammatical attribute, a corresponding modified example is retrieved from the modified example database. The translation example-based translation method according to claim 1.