JP2004199519A - Mechanical translation method, mechanical translation device, and mechanical translation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To output a proper translation result by determining a translation or syntax with a viewpoint of the whole original document or a part where ambiguity is solved. <P>SOLUTION: In this mechanical translation program, the inputted original document is analyzed on the basis of a predetermined rule, the analysis result including ambiguity information of meaning and modification relation is added to the original document as a tag to form an intermediate document (1), which is then stored in a storage device. The document is further converted to a document of an intended language in the state where the tag is included in the document of the intended language on the basis of a predetermined rule in reference to the original document and the intermediate document (1) and stored as an intermediate document (2) in the storage means, and a translation having ambiguity information included in the intermediate document (2), or an intended language document obtained by selecting/editing the translation in reference to the original document and the intermediate document 1 is outputted, whereby a further proper translation result can be outputted. The analysis/conversion generation result is made into an intermediate document of standard XML format, whereby the man-power of program development can be reduced. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to machine translation, in particular, to input a source sentence described in a source language to be translated, analyze the input source text using a source language dictionary, grammar rules, and the like, and recognize by analysis. The present invention relates to a technique for converting a source language in a stepwise manner and outputting a target language sentence with ambiguity resolved.
[0002]
[Prior art]
With the current internationalization of information, the use of machine translation devices is increasing, and there is a strong demand for improved translation quality. In the past, if the meaning of the original text could be read, the purpose of use was often achieved, or it was often used as an auxiliary tool for translators, but now it is used for various purposes such as using the translation result without processing it as it is A form is required.
[0003]
Research on machine translation began around 1945, when computers appeared, and various basic natural language processing techniques were proposed, and morphological analysis, which decomposes sentences into words registered as dictionary headings, sentence Parsing to analyze the structure of a sentence, expressing grammatical rules in the form of a logical expression, using that to analyze the sentence, preparing grammatical information in a vocabulary item, and Techniques related to various natural language processing techniques such as lexical analysis for analysis are described in publications (for example, see Non-Patent Document 1).
[0004]
A conventional machine translator incorporating such a basic technology includes, for example, an input unit 1 for inputting a translated word or a translated word specifying original sentence specifying a translated word, a CPU and a memory 2 for performing a machine translation process as shown in FIG. The CPU and the memory 2 have a morphological analysis unit that analyzes morphemes (conjugated endings, prefixes, suffixes, etc.) as units smaller than words and idioms that are basic units forming sentences to identify words. 21; a syntactic / semantic analysis unit 22 that performs a syntax analysis from the morphological analysis result; a conversion generation unit 23 that converts a source language recognized in the analysis stage to generate a sentence described in a target language; The output unit 24 outputs the post-edited target language sentence as needed.
[0005]
The morphological analysis unit 21 decomposes the inputted translated word specifying original text into morphemes and analyzes the morpheme using the dictionary 25. The morphologically analyzed original text is subjected to syntax analysis and semantic analysis by the syntax / semantic analysis unit 22. "Syntactic analysis" is to extract the syntactic structure of a sentence from the interrelationship between words and idioms. "Semantic analysis" is mainly to check the semantic category of nouns and the case control of verbs, and Select the meaning uniquely and extract the semantic structure of the source language. The analysis result of the syntax analysis and the semantic analysis is input to the conversion generation unit 23, and the source language recognized in the analysis stage is converted to generate a sentence described in the target language (for example, refer to Patent Document 1). 1).
[0006]
[Patent Document 1]
JP-A-6-12448 (page 2-3, FIG. 5)
[0007]
[Non-patent document 1]
Nomura Hirosato, "Basic Technology of Natural Language Processing," 2nd Edition, The Institute of Electronics, Information and Communication Engineers, May 25, 1989 [0008]
[Problems to be solved by the invention]
However, in the conventional machine translation apparatus as described above, since the translation technique for each sentence has been fundamental, even if it analyzes a certain degree of context, it is possible to completely remove the ambiguity of the sentence. It was difficult to respond, and there were many problems that the translated words and translations were not appropriate. For example, there is an original text "I am an eel.""Givemeeel" and "I am an eel" are candidates for translation. If you are translating one sentence at a time as in the past, the ambiguity cannot be resolved and you will have to select either "give me eel" or "I am an eel", so that often the translation will not be appropriate. there were.
[0009]
However, somewhere in the document, there is a sentence with a keyword such as "I finished eating and left the store", and "give me eel" is a suitable translation if I take the whole original document into consideration during translation. You should understand. The present invention has been proposed in view of the above circumstances, and determines a translation word and a syntax in view of the entire original document or a part where ambiguity has been resolved, and outputs an appropriate translation result. It is an object.
[0010]
[Means for Solving the Problems]
FIG. 1 shows an overall configuration diagram of Embodiment 1 of the present invention. The machine translation program 20 of the present invention is a program executed on the machine translation device 2 which is a computer. When the original document 31 to be translated is specified through the input unit 1, the analysis unit 21 is instructed to input. The original document 31 is analyzed based on a predetermined rule, and an analysis result including the ambiguity information of the meaning and the dependency relation added to the original document as a tag is stored in the storage device 3 as the intermediate document (1) 32. The conversion generation means 22 refers to the original document 31 and the intermediate document (1) 32 and converts a document in the target language including the tag into a document in the target language based on a predetermined rule based on a predetermined rule. The intermediate document (2) 33 is stored in the storage device, and the translation completion means 34 translates the translation or the translated sentence having the ambiguity information included in the intermediate document (2) 33 into the original document 31 and the intermediate document ( ) 32 with reference to select and edit, and is obtained by the outputs of the desired word document by removing the tag, it is possible to increase the accuracy of the translation resolving the ambiguity.
[0011]
The analysis means 21 analyzes morphemes (conjugated endings, prefixes, suffixes, etc.) as units smaller than words or idioms, which are basic units for forming sentences, similarly to a conventional machine translation program, and performs morphological recognition of words. It consists of an analysis unit and a syntactic / semantic analysis unit that performs syntactic analysis from the morphological analysis results.In the present invention, the analysis results were used as internal text information for processing by a machine translation program. It is characterized in that it is converted into a standard XML (eXtensible Markup Language) format in which analysis results are added as tags for each morpheme in order to facilitate the processing.
[0012]
The conversion generation unit 22 converts the source language recognized in the analysis stage and generates a sentence described in the target language in the same manner as in a conventional machine translation program. In the present invention, the conversion generation unit 22 generates the sentence. There is a feature in that the intermediate document (1) 32 in the XML format is converted including the tag, and the intermediate document (2) 33 to be generated is also converted to the XML format with a tag. By doing so, the development of the analysis means 21, the conversion generation means 22, and the translation completion means 23 is facilitated, and the accuracy of translation is improved, and the program is modified according to the characteristics of the original text used by the user. In addition, a general user can easily add a new analysis unit 21 and a conversion generation unit 22.
[0013]
By using a plurality of analysis units 21 and conversion generation units 22, the analysis unit 21 and the conversion generation unit 22 may execute parallel processing as separate programs for parts where no inconsistency occurs even if parallel processing is performed. By selecting the minimum conversion generating means 22 according to the characteristics of the original text used by the user, it is possible to reduce the throughput of the entire machine translation process.
[0014]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 shows an overall configuration diagram of Embodiment 1 of the present invention. The machine translation program 20 of the present invention is a program executed on the machine translation device 2 which is a computer. When the original document 31 to be translated is specified through the input unit 1, the analysis unit 21 is instructed to input. The original document 31 is analyzed based on a predetermined rule, and an analysis result including the ambiguity information of the meaning and the dependency relation added to the original document as a tag is stored in the storage device 3 as the intermediate document (1) 32. The conversion generation means 22 refers to the original document 31 and the intermediate document (1) 32 and converts a document in the target language including the tag into a document in the target language based on a predetermined rule based on a predetermined rule. The intermediate document (2) 33 is stored in the storage device, and the translation completion means 34 translates the translation or the translated sentence having the ambiguity information included in the intermediate document (2) 33 into the original document 31 and the intermediate document ( ) 32 with reference to select and edit, and is obtained by removing the tag to output the object document.
[0015]
The input device 1 instructs the machine translation device 2 such as a keyboard to perform translation, and the storage device 3 stores an original document 30 to be translated and an intermediate document (1) which is an analysis result of the analysis unit 21. 32, an intermediate document (2) 33 converted and generated by the conversion generating means 22, and a translated document 34 output by the translation completing means 23 are stored. The storage device 4 stores a dictionary 41 used by the analysis unit 21 and the conversion generation unit 22 for morphological analysis, semantic analysis, syntax analysis, translation word selection, and the like. The storage devices 3 and 4 need not be different external storage devices, and may be the same external storage device or a storage device such as a memory.
[0016]
FIG. 2 is a flowchart showing a processing flow of the machine translation program 20 according to the first embodiment of the present invention. In this embodiment, processing from analysis to generation of a translated sentence when translating from Japanese to English based on Example 1 of the original document to be translated shown in FIG. 5 will be described, but the present invention translates from Japanese to English. The translation is not limited to the case but can be performed from any language to any language. First, the machine translation program 20 takes out the original document to be translated specified via the input unit 1 and stores it on the storage device 3 (S201). In the present embodiment, the storage device 3 stores “I saw an elephant at the zoo. The elephant had a long nose” as an original document to be translated.
[0017]
Next, the machine translation program 20 performs a predetermined morphological analysis, separates the original document 31 stored in the storage device 3 and refers to a dictionary 41 (in this processing, a Japanese grammar analysis dictionary) to determine the part of speech of each word. The determination result is added to the original document as a tag with the analysis result to which the grammar information is added, and the result is stored in the storage device 3 as the intermediate document (1) 32. (S202).
[0018]
In the present embodiment, when the sentence “I saw an elephant at the zoo” is subjected to conventional morphological analysis of machine translation, “I (noun) is (particle) zoo (noun) and (particle) elephant ( Noun) is read as (verb) (verb), and information indicating a sentence or a main part / predicate is added as grammatical information as a tag. FIG. 6 shows an example of the intermediate document 1 generated by this processing. A tag is enclosed by “<” and “>”, and is a symbol for explaining a word, a sentence, a document, etc. sandwiched therebetween, for example, as a pair of <document> and </ document>. Each word has a tag indicating a part of speech, a main part / predicate, or the like. The tag addition processing itself is the same processing as that of a general markup language, and will not be described.
[0019]
Next, the machine translation program 20 performs syntax / semantic analysis, and stores the result of analyzing the special syntax in the storage device 3 as the intermediate document (1) 32 (S203). FIG. 7 shows an example of an analysis result in this processing. Here, the dual subject problem of "elephant is nose" is solved. That is, if "elephant" has "nose", it is solved using the semantic analysis dictionary.
[0020]
Next, the machine translation program 20 refers to the dictionary 41, performs word translation processing, and stores the processing result in the storage device 3 as an intermediate document (2) 33 (S204). Note that the word translation process is also performed on tags displayed in Japanese and added to the intermediate document (1) 32. FIG. 8 shows an example of the intermediate document (2) 33 generated by this processing. Here, conversion to a target language is basically performed by dictionary lookup of words and idioms.
[0021]
Next, the machine translation program 20 performs word order conversion processing with reference to the dictionary 41, and stores the processing result in the storage device 3 as the intermediate document (2) 33 (S205). FIG. 9 shows an example of the intermediate document (2) 33 generated by this processing. Next, the machine translation program 20 performs special syntax processing with reference to the dictionary 41, and stores the processing result in the storage device 3 as the intermediate document (2) 33 (S206). FIG. 10 shows an example of the intermediate document (2) 33 generated by this processing.
[0022]
Here, special processing for the "owner" tag is performed.
"<Owner><article> an / the </ article><noun> elephant </ noun></owner><subject><article> an / the </ article><noun> nose / trunk (elepant) </ noun></subject><verb> be </ verb><adjective> long </ adjective> ”, and this part is converted to <subject-part><subject><article> an / the </ article><noun> elephant </ noun></subject></subject-part><verb> have / has / had / had </ verb><adjective> long </ adjective><article> a / the </ article><noun> nose / trunk (elepant) </ noun>"and" have "as verbs.
[0023]
Next, the machine translation program 20 performs an option determination process, and stores the processing result in the storage device 3 as an intermediate document (2) 33 (S207). FIG. 11 shows an example of the intermediate document (2) 33 generated by this processing. In this processing, the article and the polysemy are resolved. Since words that can be selected multiple times are indicated by “/”, for example, “nose / trunk”, an appropriate word is selected based on a tag surrounding the word or information obtained from sentence analysis before and after.
[0024]
Next, the machine translation program 20 removes the tag from the intermediate document (2) 33 whose ambiguity has been resolved, and stores it in the storage device 3 as the translated document 34 (S208). FIG. 12 shows an example of the translated document 34 generated in this processing.
As the analysis / conversion generation process proceeds, not only the immediately preceding processing result but also all output results processed up to that time may be required. It is preferable to stack the results of all processes in one document, or to use a plurality of documents so that each output information can be corresponded. Note that the analysis / conversion generation processing can process a plurality of sentences in parallel. In addition, processing of special syntax and processing of removing ambiguity such as polysemy are performed in parallel for each problem, so that the processing time can be kept within an allowable range.
[0025]
Also, in many cases, it is necessary to select a translated word or a translated sentence depending on the context, such as a question of an article such as whether to take a or the. Therefore, it is necessary to embed not only grammatical information but also context information obtained by semantic analysis as a tag in a document. As an example of performing such processing, FIG. 14 shows an example of an intermediate document (1) 32 obtained by performing time-series analysis on the example of the original document of FIG.
[0026]
<Scene> is a tag indicating that the beginning of a sentence, a change in time, a change in location, or a change in subject has been detected. When a related keyword is found, the analysis is performed. <Focus> sets the first sentence when the same word appears in multiple sentences. <Action> is attached to a sentence whose subject is a person or a personified noun. <Description> is a sentence whose subject is a noun that refers to something other than a person. These are used to detect missing words such as the subject and provide clues for context analysis.
[0027]
FIG. 3 is a correspondence table with the conversion generation unit 22 that is in charge of processing the tag as the analysis result of the analysis unit 21. When the conversion generation means 22 is composed of a plurality of programs, the correspondence between the conversion generation means 22 and the name of the conversion generation program to be converted and generated using the tag output by the analysis means 21 is stored. For example, the <document> and </ document> tags indicating a document can be processed by the conversion generation program 1, and the <noun> and <verb> indicating the part of speech can be processed by the conversion generation program 2. . The unsolved flag is ON in the initial state, and is turned OFF when the conversion generation program completes the processing of the tag, and indicates whether or not the processing of the corresponding tag has been resolved. This table is provided in advance to a machine translation program provided as a product, and is added or updated when a user adds a new program as the analysis unit 21 or the conversion generation unit 22 or updates an existing program. .
[0028]
FIG. 4 is a flowchart illustrating a flow of a process of calling another conversion generation program from the conversion generation program that operates as the conversion generation unit 22. Each conversion generation program processes the tag as the analysis result of the analysis unit 21 to perform the conversion generation process. However, since the conversion generation process itself is the same as the conventional machine translation program, the details of the process are omitted. Here, a description will be given focusing on a part that calls another conversion generation program from the conversion generation program.
[0029]
Each conversion generation program performs conversion generation based on the intermediate document (1) generated by the analysis unit 21, and generates an intermediate document (2) based on the processing result. The conversion generation program scans the intermediate document (1) from the beginning and detects a tag (S401), and refers to the tag / conversion generation correspondence table to determine whether the tag can be processed by the own conversion generation program. (S402). If the tag cannot be resolved by the self-conversion generation program, the processing for that tag is skipped, and the flow returns to the tag extraction processing of S401.
[0030]
If the extracted tag can be processed by the self-conversion generation program, the tag is processed (S403). For the actual tag processing, refer to the description of the processing flow of the machine translation program 20 according to the first embodiment of the present invention in FIG. When the processing of the tag is completed, the unresolved flag of the corresponding tag in the tag / conversion generation correspondence table shown in FIG. 3 is turned off (S404). The above processing is repeated until the processing of all tags is completed (S405). When the processing of all tags is completed, the unresolved flag is turned ON with reference to the tag / conversion generation correspondence table shown in FIG. It is determined whether or not there is any tag (S406), and a conversion generation program that is in charge of processing the tag whose unresolved flag is ON is called (S407).
[0031]
As described above, the analysis / conversion generation processing is performed in multiple stages, and the subsequent processing can be performed by referring to all the processing results before that, so that a more appropriate translation result can be obtained. It becomes possible. Further, the processing is divided into a plurality of stages, and the processing result is transmitted as an intermediate document (1) and an intermediate document (2) in the XML format without using the internal interface, so that it can be solved by one complicated processing. In comparison, the processing at each stage is simplified and development is facilitated. Furthermore, it is possible to customize the machine translation program provided as a product according to the characteristics of the original text to be translated by the user, and to add a new analysis / conversion generation process. Can also be easily developed.
[0032]
【The invention's effect】
It is possible to output a more appropriate translation result, and it is possible to reduce the number of program development steps by using the analysis / conversion generation result as an intermediate document in a standard XML format.
[Brief description of the drawings]
FIG. 1 is an overall configuration diagram of Embodiment 1. FIG. 2 is a machine translation program process in Embodiment 1. FIG. 3 is a tag / conversion generation correspondence table. Call processing of conversion generation program [Figure 5] Example 1 of original document to be translated
FIG. 6 is an example 1 of the intermediate document (1).
FIG. 7 is an example 2 of the intermediate document (1).
FIG. 8 is an example 1 of the intermediate document (2).
FIG. 9 shows an example 2 of the intermediate document (2).
FIG. 10 shows an example 3 of the intermediate document (2).
FIG. 11 is an example 4 of the intermediate document (2).
FIG. 12: Translation result example 1
FIG. 13 is an original document example 2 to be translated.
FIG. 14 is an example 3 of the intermediate document (1).
FIG. 15 shows an example of a conventional machine translation apparatus [Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Input means 2 Machine translation device 3 Storage device 4 Storage device 20 Machine translation program 21 Analysis means 22 Conversion generation means 23 Translation completion means 31 Original document 32 Intermediate document (1)
33 Intermediate Document (2)
41 dictionaries

Claims (3)

Input the source document described in the source language, analyze the input source document, convert the parsed source language to generate a target language document, and post-edit the target language document as necessary A machine translation method for outputting
When the input original document is analyzed based on a predetermined rule, a result obtained by adding the analysis result including the meaning and the ambiguity information of the dependency relationship to the original document as a tag is stored in the storage device as an intermediate document (1). One or more analysis steps;
With reference to the original document and the intermediate document (1), based on a predetermined rule, a document in a target language including the tag is converted into a document in a target language as an intermediate document (2). One or more transformation generating steps to be stored in a storage device;
A translation completion step of selecting and editing a translated word or a translated sentence having ambiguity information included in the intermediate document (2) with reference to the original document and the intermediate document (1), and outputting an object word document;
A machine translation method, comprising: Input means for inputting an original document described in a source language;
When the input original document is analyzed based on a predetermined rule, a result obtained by adding the analysis result including the meaning and the ambiguity information of the dependency relationship to the original document as a tag is stored in the storage device as an intermediate document (1). One or more analysis means;
With reference to the original document and the intermediate document (1), based on a predetermined rule, a document in a target language including the tag is converted into a document in a target language as an intermediate document (2). One or more conversion generating means to be stored in a storage device;
Translation completion means for selecting and editing a translation word having ambiguity information or a translation included in the intermediate document (2) with reference to the original document and the intermediate document (1) and outputting an object word document;
A machine translation device comprising: When the input original document is analyzed based on a predetermined rule, a result obtained by adding the analysis result including the meaning and the ambiguity information of the dependency relationship to the original document as a tag is stored in the storage device as an intermediate document (1). One or more analysis means;
With reference to the original document and the intermediate document (1), based on a predetermined rule, a document in a target language including the tag is converted into a document in a target language as an intermediate document (2). One or more conversion generating means to be stored in a storage device;
A computer as translation completion means for selecting and editing a translated word having ambiguity information or a translated sentence included in the intermediate document (2) with reference to the original document and the intermediate document (1) and outputting an object word document A machine translation program that works.

Images