JP3281746B2 - Machine translation equipment - 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 translator for translating an input sentence written in a first language into a desired second language and outputting the translated sentence.

[0002]

2. Description of the Related Art Conventionally, in a machine translation apparatus, when a plurality of types of syntax structures are obtained, that is, when a plurality of types of translation results are obtained, it is necessary to determine a priority among them. Means for determining this priority include (1) semantic analysis, (2) co-occurrence relation, and (3) description order of grammar rules.

As a specific method of the semantic analysis of (1), semantic category information is written in a word dictionary,
When determining the dependency between words by syntactic analysis, a syntactic structure in which the semantic categories do not match between the words to be changed is not adopted or the priority is lowered.

The use of the co-occurrence relation in (2) can be said to be a kind of semantic relation. However, for a certain word, the word of the destination of the dependency is specified and described in a dictionary, and the relation between the words is analyzed by syntactic analysis. When the reception is determined, a syntax structure in which the word of the dependency destination does not match the dictionary description is not adopted or the priority is lowered.

The description order of the grammar rules in (3) uses rules for parsing. By adjusting the order in which grammar rules are described in memory, for example, in the order in which they are to be prioritized, it is possible to change the ease of application of each grammar rule. This will make it easier to create syntactic structures.

Also, by learning the syntax structure,
Techniques exist to improve the legitimacy rate of machine translation. In Japanese Unexamined Patent Publication No. Hei 2-148261, "Machine Translation Apparatus with Syntax Learning Function", the user corrects the analysis result output by the apparatus, and in subsequent translation instructions, performs an analysis process using this correction history to obtain the same This avoids the modification.

Further, Japanese Patent Laid-Open Publication No. Hei 6-195377, "Syntax Priority Learning Machine Translation System" performs an analysis process in which the syntax structure of an input sentence is regarded as a phrase structure, and gives priority to the entire phrase structure that the user considers appropriate. This is considered to be the structure to be considered, and in the subsequent translation processing, the priority of the analysis result is interpreted using the priority of the phrase structure.

However, the technique disclosed in Japanese Patent Application Laid-Open No. 2-148261 has a precondition that the user can appropriately point out an error in the analysis processing. there were. The task of correcting an error in the parsing process requires a great deal of effort unless one is familiar with the grammatical structure of the language.

In the technique disclosed in Japanese Patent Application Laid-Open No. 6-195377, a priority rule limited to a local structure is used while trying to capture a global syntax structure. Therefore, the unit of application of each learning result is extremely limited. Not universal. In addition, there is a problem that it is not possible to extract a difference in local structure, which is an essential cause of a difference in priority between a plurality of parsing results. Furthermore, in any of the above technologies,
The target of learning is limited to the entire input sentence or phrases that are continuous in the input sentence.For languages where the appearance of modifiers is not always uniquely determined, such as Japanese,
It did not work well.

[0010]

THE INVENTION try to challenge and solutions In general, in the translation processing unit, the analysis processing for the input sentence is carried out, semantic processing, etc. by that dependency pruning is performed in one or more of The result of parsing is obtained. It is not optimal to capture the syntax structure obtained as a result of the syntax analysis with a phrase structure, at least for Japanese.

[0011] For example, the English sentence "I book books yesterda"
y. 2 shows an example of an analysis result when the syntactic structure of “” is captured by a phrase structure. At this time, “I”, “bout”, “book”
s "" yesterday "". In order to generate a sentence expressing the same meaning by combining each morpheme component of "
A sentence that uniquely forms the phrase structure of FIG. 2 must be generated. Therefore, it can be said that this syntactic structure and sentence correspond one-to-one.

However, for example, if the syntax structures of three Japanese sentences “I bought Japan yesterday”, “I bought a book yesterday”, and “I bought a book yesterday” are represented by phrase structures, respectively. (A), (b), and (c), the sentence is analyzed as having a different syntactic structure. However, since the sentence of these three sentences is almost the same, the syntactic structure and the sentence have a one-to-one correspondence. It cannot be said that it corresponds to.

For this reason, in Japanese sentences, a method of expressing a syntactic structure by a dependency structure is often used, in which the modification / modification relationship between morpheme components is regarded as a dependency relationship on a modified word after modification. If the above three sentences also take the syntactic structure as a dependency structure, they are all aggregated in the form of FIG. 3 (d), and the syntactic structure and sentence can correspond one-to-one. It will be in a form that makes it easy to understand the structure. In addition, "noun phrase""verbphrase""mainclause" in FIG. 2 and FIG.
Terms such as "sentences" are names given for convenience in order to facilitate understanding of the explanatory diagram.

In the conventional translation processing apparatus, when a plurality of analysis results are obtained, each analysis result is given some priority, and a translated sentence is generated using the most appropriate analysis result. Function to present translations based on all analysis results in a form that allows the user to clearly indicate the order in which they are considered appropriate, and by explicitly specifying the dependency between each morpheme component to achieve the desired analysis result. In some cases, the local sentence structure is learned by storing the contents of the dependency, but in the case of a sentence with a complicated sentence structure, the output translated sentence is compared to the input sentence. User can understand that the output is inappropriate,
In many cases, the cause cannot be accurately pointed out from the results of parsing. In particular, if the user is not familiar with Japanese, it is almost impossible to specify the dependency, and such a learning function may not be available. .

In the conventional translation processing apparatus, when learning the syntax structure, the user must always specify which word string should be given priority to a specific syntax structure. On the other hand, the learning is not reflected unless a continuous word string is reproduced. Therefore, when the input sentence is Japanese with few restrictions on the insertion position of the modifier, the learning effect is hard to appear.

For example, as a result of parsing “my tennis racket”, a solution in which “my” modifies “tennis” (FIG. 4A) and “my” modifies “racquet” It is assumed that there is a solution (FIG. 4 (b)), and when the latter analysis result is prioritized, an insertion phrase (in this case, "important" in the form of "my important tennis racket") ), The learning function could not be used because it was not recognized that the sentence had the same syntactic structure.

Although there is a translation processing apparatus having a function of displaying the syntax structure of the result of the syntax analysis, the syntax analysis result is given priority over the entire syntax analysis result by giving priority to any syntax structure. There was no function to explicitly show the operator.

Furthermore, even if some parsing rules that determine the priority of the syntactic structure are considered to be inappropriate for the user, there is no means for the operator to correct them. .

In the above example of "my tennis racket", the morpheme sequence of "[noun 1] [noun 2] [noun 3]" is changed to [noun 1] and [noun 2] both [noun 3] (FIG. 4D), this learning pattern is erroneously applied to the sentence “Scratch my racket” (FIG. 4E). In order to finely control the application conditions of the pattern, it is not appropriate to perform uniform processing by matching the surface patterns such as “A, B, C” in order to cope with various language phenomena.

Further, in the above example, when "my tennis racket" is input to the morpheme sequence of "[noun 1] [noun 2] [noun 3]", [noun 1] and [noun 1] [Noun 2] determines that it is correct to modify both [Noun 3], and when " my wound of racket" is input, [Noun 1] becomes [Noun 2] and [Noun 2] becomes [Noun It is contradictory to judge that it is correct to modify each of [3], but there is no means to confirm that they are mutually counterexamples, so there is a possibility that inconsistent learning may be performed. there were.

An object of the present invention is to solve such a problem, and to optimize a plurality of parsing results output by the apparatus, or a translated sentence based on the parsing results. By selecting one of the items, the operator automatically extracts the erroneous part of the parsing result and has the function of automatically learning the correct structure,
In addition, it has a function that presents the data that caused the error to the user and allows the user to correct it, and a function to learn the local syntax structure that can assign detailed application conditions such as part of speech and meaning. , Can show what input the learned data was created for,
An object of the present invention is to provide a translation processing device having a syntax structure learning function.

[0022]

According to the present invention, there is provided a translation processing apparatus comprising: an input unit for inputting a sentence described in a first language;
The input sentence input by the input unit is decomposed into morphemes,
The morphological analysis processing unit that performs the analysis and the dependency
In the morphological component to be formed, the relative position in the input sentence,
Dependency relationships according to priority conditions such as attribute conditions of the lyrics
Priority score data that quantifies the connection depth of
Syntax structure priority storage unit that is stored together with the writing priority condition
And each morphologically analyzed by the morphological analysis processing unit
Analyzed the syntactic structure between morphemes and obtained multiple
Based on the priority condition,
To obtain the priority score of the syntax structure.
A parsing processing unit for determining a priority order, and the parsing processing
Parsing that holds the plurality of analysis results output from the unit
The parsing result storage unit and the parsing result in the order of the priority.
Transforms the analysis result held in the result storage unit into a translated sentence of the second language.
A conversion generation processing unit that performs conversion generation processing; and the syntax analysis processing.
Output unit that outputs the processing results of the conversion unit and the conversion generation processing unit
And the plurality of analysis results obtained by the syntax analysis processing unit
Analysis results with the highest priority from among the
Compare the analysis results judged to be appropriate, and
In the syntax structure of the analysis result, different parts are
Analysis result difference extraction unit to extract as the analysis result difference
The local syntax structure extracted by the
The priority condition for the structure and the priority score are output to the output unit.
User to correct the priority score for the priority condition.
And a syntactical structure priority correcting unit .

Further, the analysis result difference extraction unit extracts
Priority for a given local syntax structure
Scores and items about the words that make up the local syntactic structure
Learning data such as words and meanings are output from the output unit, and
A learning data creation support unit that allows the user to modify and change
Further comprising a learning data storage unit for holding the training data, prior to
The notation parsing processing unit calculates a priority score for the analysis result.
And referring to the learning data storage unit,
If there is learning data for the syntax structure of the analysis result,
Add the score given to the training data to the priority score
It shall be the feature of the that. Further, the translation processing device according to the present invention
The input unit for inputting a sentence described in the first language;
The input sentence input by the input unit is decomposed into morphemes,
A morphological analysis processing unit for performing analysis, and the morphological analysis processing unit
/ Modification relation between each morpheme analyzed by morpheme
Dependency structure representing the dependency of a modifier on a modifier
Analyze the syntax structure composed of trees and obtain multiple
A predetermined priority is given to the analysis result of the syntactic structure
Priority score of the syntax structure based on the degree condition
A parsing processing unit that determines the priority according to the magnitude of the sum;
The plurality of analysis results output from the notation syntax analysis processing unit.
The syntax analysis result storage unit to be held, and in the order of the priority,
The analysis result held in the syntax analysis result storage unit is described in a second language.
A conversion generation processing unit that performs conversion generation processing on a translation of a word;
Output the processing results of the
A machine translation device comprising:
Excellent among the plurality of analysis results obtained by the sentence analysis processing unit
The analysis result with the highest priority and the user is most appropriate
Of the analysis results
Branches in the dependency structure tree that constructs the syntax structure
Solve, compare each analysis result for each branch unit,
And reconstruct the remaining part as a dependency tree
An analysis result difference extraction unit for extracting a difference syntax structure,
For the difference syntax structure extracted by the analysis result difference extraction unit
The analysis result selected by the user
A learning data generator for generating syntax structure learning data;
A learning data storage unit for holding the learning data.
The notation parsing processing unit, based on the learning data,
It is characterized in that the priority score of the structure is changed.

Thus, by simply selecting the most appropriate sentence from the output target language sentence by the user of the apparatus, it is possible to extract the local syntax structure that needs to be corrected from the syntax analysis result. Yes, it is possible to automatically create learning data for always prioritizing one of the structures to be prioritized, and provide the user with data that is the basis for prioritizing the syntax analysis result deemed most appropriate by the device. A local syntax structure learning function that allows the user of the device to directly change data for determining the priority of the syntax structure in addition to being able to specify, and allows the user to freely assign application conditions such as part of speech and meaning. It has a data creation support function, and when displaying the results analyzed using the learning data, to show the basis of the previous learning, the syntax structure is created when the learning data is created. Those having a function to be displayed in conjunction with a word that had forms.

[0025]

According to the present onset Ming configuration, when the translation processing unit has output a translation corresponding to each based on a plurality of analysis results, determination in their translation, the system the highest priority There is a translation that is considered appropriate from the translated sentences, and when the user selects it, the local syntax structure is automatically extracted from the two analysis results and presented to the user. Change
In the extracted local syntax structure, the analysis result to be prioritized
Automatically generate learning data for. In addition, the system
Priority conditions that are the basis for determining the priority of the syntactic structure
And the priority score data are presented to the user. Also, priority order
In order to reverse the position (use the user-selected syntax
User to change the above data
Can adjust the priority. Enter more priority
The learning data for replacement is based on applicable terms such as part of speech and meaning.
It is possible to arbitrarily register the matter.

[0026]

[0027]

[0028]

DESCRIPTION OF THE PREFERRED EMBODIMENTS As an embodiment of the present invention, a Japanese-to-English translation process from Japanese to English with Japanese as the input language and English as the target language will be described in detail with reference to the drawings.

FIG. 1 shows an example of the configuration of a machine translation apparatus according to the present invention. 1 is an input unit, 2 is a morphological analysis processing unit, 3 is a syntax analysis processing unit, 4 is a syntax analysis result storage unit, 5 is a conversion generation processing unit, 6 is an analysis result difference extraction unit, 7 is an output display unit,
8 is a learning data generation unit, 9 is a learning data storage unit, 10 is a syntax structure priority storage unit, 11 is a syntax structure priority correction unit,
Reference numeral 12 denotes a learning data creation support unit.

The operation of the machine translator according to the present invention will be specifically described with reference to the flowchart of FIG.

First, in the input unit 1, one sentence of Japanese sentence data is extracted and sent to the morphological analysis processing unit 2 (S50).
1). Then, the morphological analysis unit 2 extracts all the sentences or <br/> et extractable word inputted, extracting as much as possible constructed clauses as the combination of these words, this clause column To obtain a combination of phrases that can constitute the entire input sentence (S502). Next, a syntax analysis process is performed by the syntax analysis processing unit 3 on the set of clauses that are the result of the morphological analysis (S503). The syntax analysis processing will be described with reference to the flowchart in FIG.

A judgment as to whether or not the syntax structure existing before in the input sentence can be related to the syntax structure existing after it is made as an intermediate result of the syntax analysis including a set of all clauses obtained as a result of the morphological analysis. It is confirmed whether the processing has been performed for all the obtained syntax structures (S601), and if not, the following processing is performed.

First, with respect to any two syntactic structures (including phrases resulting from morphological analysis), the dependency rules based on the parts of speech and meanings of the words that form the clause located at the top of the dependency structure tree that constitutes each structure For the syntax structure pair determined to be dependable based on the syntax structure reference, the syntax structure priority storage unit 10 refers to the syntax structure priority storage unit 10 and assigns the priority score given to the corresponding priority condition to each syntax structure. To the priority score (the phrase in the morphological analysis result does not have to have a score, or an initial score can be given in order to have a priority). A new syntax structure tree to be included is generated (S602).

Subsequently, it is checked whether the generated syntax structure is a learned syntax structure by referring to the learning data storage unit 9 (S603). If there is any learning data, the learning data is given to the learning data. The score is added to the priority score given in S602 (S604).

A description will be given of the parsing process when the example sentence "I lent a new book to him" is input. FIG. 8 shows an example of a phrase string extracted as a result of morphological analysis of this sentence.
(A). Here, the extracted bunsetsu sequence, the part-of-speech name such as "pronoun" or "adjective" or the inflected form name such as "termination form" are not uniquely determined by the grammar system on which they are based. The uncertainties do not affect the operation of the present invention.

FIGS. 8C and 8D show a process in which syntactic analysis is performed on these clauses extracted in the morphological analysis and a priority score is given to the syntactic structure. For example, if the phrase [I] is analyzed to be able to modify the phrase [new], the syntactic structure priority storage unit 8 is searched for a priority condition matching this dependency structure. FIG. 8B shows an example of priority conditions and priority score data described in the syntax structure priority storage unit 8. In the case of this example, the priority condition (condition 1 in FIG. 8 (b)) that gives priority to the dependency having a short phrase distance corresponds, and corresponds to [“phrase distance” × (−1)] (−1).
The points are added as the priority score given to the syntax structure [[I] → [New]], and the sum of the original [I] and [New] scores is 0. No), and the priority score is (-1) after all (see FIG. 8).
(C)).

Here, [[I] → [New]] is
Dependency structure that [I] modifies [New] [[Him] →
8D, the sum of the priority scores of the original syntax structure is 0, and the priority condition applied (condition 1 in FIG. 8B). Since the addition of the scores according to (4) is 0 in total, the priority score for the syntax structure [[[new] → [book]] / [to him] → [lended]] is 0 in total. Here, [[[New] → [Book]]
/ [To him] → [lent]] is [[new] → [book]]
And [to him] qualify [lent].

Among the syntax analysis results stored in the syntax analysis result storage unit 4, the solution with the highest priority score is sent to the conversion generation processing unit 5, and a conversion generation process is performed (S504). Is output to the output display section 7 (S50).
5).

Here, if the operator of the apparatus does not instruct the output of a translated sentence (hereinafter, referred to as "another interpretation") based on the analysis result of the next candidate, the process ends (S506). Similarly, when there is no further interpretation, the process ends (S50).
7).

When the operator instructs another interpretation output and another interpretation exists, a solution having the next highest priority score from the syntax analysis result storage unit 4 is sent to the conversion generation processing unit 5, and the conversion generation processing is performed. Then, the translated sentence generated from the highest priority solution is written together with the translated sentence and output to the output display unit 7 (S50).
9). If the operator instructs further interpretation output,
The processing after S507 is repeated (S510). If the operator determines that the translated sentence of the different interpretation is correct and does not perform the learning of the syntax analysis, the process ends (S51).
1).

On the other hand, when the operator learns,
A parsing result difference extracting unit 6 parses the parsing result of the top priority solution and the parsing result of generating a translated sentence determined by the operator to be correct.
The analysis result difference extraction unit 6 compares the two syntax analysis results, extracts a local syntax structure of the difference (hereinafter, referred to as a “difference structure pair”), and outputs it to the output display unit 7 (S5).
12).

Since there may be a plurality of the differential structure pairs,
While there is a difference structure pair for which the learning process has not been completed, the steps of the learning process are repeated (S514), but if not, the process ends (S513). Details of the learning process will be specifically described with reference to the flow of FIG.

When automatically creating learning data from the difference structure pair, the syntax structure of the difference structure pair which should be given priority is sent to the learning data automatic generation unit 8 (S701). , Generates learning data to which data necessary to indicate that the sent syntax structure has priority over the other syntax structure of the difference structure pair is stored in the learning data storage unit 9, and the learning process ends. (S702). At this time, if the parsing method that gives priority to the parsing result in which the sum of the priority scores given to the local syntactic structure is large is used, the additional score that exceeds the score of the other syntactic structure of the differential structure pair The data corresponds to data necessary to indicate the priority of the learning data.

If the operator does not select the automatic learning in step S701, the difference structure pair to be subjected to the learning process and the priority score given to each syntax structure are output and displayed together with the priority condition giving the score. Displayed on the part 7 (S70
3). At this time, if one of the syntax structures has already been learned, the words at the time of learning and the priority score added by learning are also displayed. Here, in order to reverse the priority of the difference structure pair, the priority condition held in the syntax structure priority storage unit 10 is corrected or deleted, or the priority score for the priority condition is changed. Performs the following processing, otherwise proceeds to S706 (S706)
704). The syntax structure priority correction unit 11 rewrites the priority condition or the priority score data stored in the syntax structure priority storage unit 10 according to the correction data input by the operator, and the learning process ends (S705).

If the priority condition and the priority score are not corrected in S704 and the following learning process is not performed, the process ends (S706). The learning application range is limited to the difference structure pair, but not limited to the combination of the input words forming the local syntax structure that the operator wants to prioritize, but also the part of speech or meaning to which the word belongs. In order to perform the learning process of expanding the application range to the section, the syntactic structure of the one of the difference structure pairs to be prioritized is sent to the learning data creation support unit 12,
The learning data creation support unit 12 extracts the syntax structure itself and attribute information such as the part of speech and semantic classification of each clause constituting the syntax structure as initial data for creating the learning data, and displays them on the output display unit 7 (S707). ). When the operator gives a condition to each attribute to give a specific instruction on the applicable range of learning, and further gives an additional score to the learning data, the learning data creation support unit 12 accordingly applies the learning data application condition to each clause. Then, learning data is created by adding the score of the priority score to the input word, and the learning data is stored in the learning data storage unit 9, and the learning process is terminated (S708).

For example, FIG. 9B shows an example of a translation result based on the analysis result of FIG. 9A. At this time, a process in which the operator of the apparatus judges that the translated sentence for the analysis result 2 is more appropriate and learns the syntax analysis result will be described. First, in order to obtain a difference structure pair between the syntax structure of the analysis result which has been given the highest priority by the apparatus and the one selected by the operator, the branching part of the dependency structure tree is solved, and the remainder obtained by removing the matching branch is used as the dependency structure. Rebuild on a tree. Thus, even when the operator cannot point out an error in the syntax structure, a difference structure pair is extracted as shown in FIG. 9C.

When learning the syntax structure automatically, learning data in which the learning application conditions are limited to the input word is automatically generated. At this time, the learning data automatic generation unit 8 performs the processing shown in FIG.
Learning data to which scores (two points) necessary for reversing the difference (one point) of the priority scores between the first and second syntax structures 1 and 2 are generated as shown in FIG. The application range of the learning data shown in FIG. 12 is limited to a syntax structure consisting of four clauses “I”, “new”, “book” and “lent”, but “I gave him two new books yesterday. The learning data is applied even if there are other modifiers such as "I lent a book."

In the case where the operator corrects the defined priority score itself, the difference score is presented to the operator together with the priority score for the difference structure pair and each syntax structure together with the priority condition for giving the score. The user can modify or delete the priority condition, or change the priority score according to the condition. In FIG. 10, these display examples and the priority score for the condition that the operator modifies the verb with a noun or a pronoun are respectively 2 to 10 points.
This shows that the priority score of the entire syntax structure is reversed by raising the score to a point. Changing the priority score in this way affects not only a specific syntax structure but also an extremely wide range of input.

Further, when learning a syntax structure having a clear scope of application without changing the priority score,
Create initial data for learning that stores the syntax structure to be learned and the attribute information of the words that make up each clause belonging to it.
It is presented to the operator as shown in FIG. The operator sets learning application conditions based on the presented attribute information.

FIG. 13 shows an example in which the operator creates learning data. The symbol * (asterisk) in the figure indicates that any element is permitted in the item, for example, [central word = *], [attached word =], [part of speech = noun], [meaning =
*] Indicates that the condition is "any noun with the attached word""." Learning data shown in FIG. 1. 2 with respect to the input sentence "I have lent a new book to him", but has the effect of always giving priority to the specified syntax structure of the subsequent operator, the operator has expanded the scope The learning data of the example in FIG. 13 can cope with more various input sentences than the learning data in FIG. 12 of the example automatically generated by the device. For example, in the dependency tree of the result of parsing the sentence "I envy watching him read a difficult English book," the structure shown in FIG. Since the data structure is included, the designation of the syntax structure by the learning process is reflected, and a learning effect is obtained.

Next, an example of preventing the operator from performing learning inconsistent with past learning will be described. If it is assumed that the learning data of FIG. 13 has been applied when analyzing the sentence "She gave her clothes", the interpretation that [She] modifies [She gave] has priority. In order to prevent the operator from canceling the training data by thinking that [She] should give priority to the interpretation that modifies [I want],
A word in the case where the application of the learning data actually registered at the time of learning is appropriate is presented as shown in FIG. On this screen, the user gives priority to this interpretation,
Further, the registration is performed by limiting the elements. In the example of FIG.
The central word is limited to [want], and the verb is limited to those that have the conjunctive word "ta", which has a semantic meaning. Thus, learning data can be added so that [she] modifies [want] without losing the learning effect of the past.

[0052]

As described above in detail, according to the present invention, among the plurality of translation sentences presented from one input sentence, the translation sentence converted and generated from the analysis result regarded as the highest priority by the device is obtained. If it is not appropriate, the analysis result is compared with the analysis result that generated the target language sentence determined by the operator to be the most appropriate, and the part with a different structure is extracted from each analysis result as a local syntax structure By providing a result difference extraction unit, it is possible to automatically extract places that are considered inappropriate from the results of the syntax analysis processing, so that the operator can judge the suitability of the translated sentence, but point out the inappropriate places in the syntax analysis results. Is difficult, it is possible to present a local syntax structure to which the assigned priority is not appropriate.

Further, as means for preventing an error in the extracted syntax analysis processing from being repeated, a learning data automatic generation unit for automatically generating syntax structure learning data for giving priority to a translation result considered appropriate by the operator thereafter. By providing a learning data storage unit for holding the learning data, the apparatus automatically creates learning data of a syntax structure to be prioritized without performing complicated registration work on the structure to be learned by the operator. Thus, for the reproduction of the same syntactic structure by the same word, the translation result desired by the operator is output with the highest priority without performing the correction again.

Further, priority score data for determining the magnitude of the priority given to each local syntax structure is expressed as follows:
By providing a syntax structure priority storage unit that holds the priority condition along with the score and an output display unit that presents these, the operator can specify the basis of the priority given to the syntax structure by the operator. Can be shown.

Further, by providing a syntax structure priority correction unit for correcting and registering the priority data held in the syntax structure priority storage unit in accordance with the instruction of the operator, a general structure having a syntax structure given as empirical knowledge is provided. The priority can be changed, and the operator can make a general-purpose adjustment to the syntax analysis performed by the device, not limited to the syntax structure composed of specific words.

For words constituting a local syntax structure whose priority is to be changed, data necessary for creating learning data to which conditions such as parts of speech and meaning are given to an operator, and the intention of the operator is provided. By providing a learning data creation support unit that creates learning data having application conditions to be applied and a learning data storage unit that holds the learning data, based on attribute information of input words presented from the learning data creation support unit, When the operator can indicate in detail the components of the syntax structure to be prioritized, the detailed intention can be fully reflected,
It is possible to create learning data having a wide application range.

Furthermore, by writing words that were components of the syntax structure at the time of learning in the learning data storage unit, the basis of the learning data created in the past can be presented.
It is possible to prevent creation of inconsistent learning data and improve work efficiency.

As described above, it is possible to select a learning means according to the operator's preference, use, or proficiency in the language, and to improve the operator's work efficiency for correcting translation results and creating learning data. There is.

[Brief description of the drawings]

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

FIG. 2 is a structural diagram of an example in which a syntax structure of an English sentence is represented by a phrase structure tree.

FIG. 3 is a structural diagram of an example in which a syntax structure of a Japanese sentence is expressed by a phrase structure tree and a dependency structure tree.

FIG. 4 is a diagram illustrating a case where there are a plurality of analysis results and learning that contradicts each other is repeated.

FIG. 5 is a flowchart showing the operation of one embodiment of the present invention.

FIG. 6 is a flowchart showing a process in a syntax analysis process (S503 in FIG. 5).

FIG. 7 is a flowchart showing a process in a learning process (S514 in FIG. 5).

FIG. 8 is a diagram illustrating a priority score given to a syntax structure.

FIG. 9 is a diagram for explaining difference extraction of a syntax structure.

FIG. 10 is a diagram illustrating a change in priority score for a rule used for prioritization in syntax analysis.

FIG. 11 is a diagram illustrating initial data for learning.

FIG. 12 is a diagram illustrating an example of automatically generated learning data.

FIG. 13 is a diagram illustrating an example of learning data specified by an operator.

FIG. 14 is a diagram illustrating a case where learning data is applied to another sentence.

FIG. 15 is a diagram illustrating an example in which the basis of past learning is presented when correction of learning data is required.

FIG. 16 is a diagram illustrating creation of learning data consistent with past learning.

[Description of sign]

 DESCRIPTION OF SYMBOLS 1 Input part 2 Morphological analysis processing part 3 Syntax analysis processing part 4 Syntax analysis result storage part 5 Conversion generation processing part 6 Analysis result difference extraction part 7 Output display part 8 Learning data automatic generation part 9 Learning data storage part 10 Syntax structure priority Storage unit 11 Syntax structure priority correction unit 12 Learning data creation support unit

Claims (3)

(57) [Claims] An input for inputting a sentence described in a first language
And the input sentence input by the input unit are decomposed into morphemes
In the morphological analysis processing unit that performs the analysis and the morphological component
Priority conditions such as relative position and part-of-speech attribute conditions in
The numerical value of the depth of the dependency relationship
It is possible to hold previous score data together with the priority condition.
Sentence structure priority storage unit and each form morphologically analyzed by the morphological analysis processing unit
Analyzes the syntactic structure between elements, and obtains multiple
The structure analysis result is configured based on the priority condition.
Calculate the priority score of the sentence structure, and order the priority according to the size of the calculated sum
A parsing processing unit for determining a position, and the plurality of analysis results output from the parsing processing unit
And a syntax analysis result storage unit that holds the same in the order of the priority.
The generated analysis result into a second language translated sentence
A conversion generation processing unit, and processing results of the syntax analysis processing unit and the conversion generation processing unit.
An output unit for outputting, and the plurality of analysis results obtained by the syntax analysis processing unit.
From the analysis result with the highest priority and the most appropriate for the user
The analysis results determined to be
In the syntactic structure of the result, different parts are
And a local syntax structure extracted by the analysis result difference extraction unit.
To calculate the priority condition and the priority score for the syntax structure.
Output to the output unit and calculate the priority score for the priority condition.
A syntax structure priority correction unit for allowing the user to correct the
And a machine translation device. 2. The station extracted by the analysis result difference extraction unit.
Priority score for the syntactic structure,
Part-of-speech and meaning of words that make up
Outputs learning data such as taste from the output unit, and
A learning data creation support unit for correcting and changing the learning data; and a learning data storage unit for holding the learning data.
The parsing processing unit obtains a priority for the analysis result.
While obtaining points, referring to the learning data storage unit,
If there is learning data for the syntax structure of the analysis result,
The score given to the training data is added to the priority score.
2. The machine translation apparatus according to claim 1, wherein the calculation is performed. 3. An input for inputting a sentence described in a first language.
And the input sentence input by the input unit are decomposed into morphemes
And a morphological analysis unit for performing the analysis, and each form subjected to morphological analysis by the morphological analysis unit.
Modification / modification relation between elements depends on modifiers
Analyzes the syntactic structure composed of dependency structure trees representing existence relationships
And the parsing results
Of the syntax structure based on a predetermined priority condition.
Find the score and determine the priority according to the sum
Parsing processing unit, and the plurality of analysis results output from the parsing processing unit
And a syntax analysis result storage unit that holds the same in the order of the priority.
The generated analysis result into a second language translated sentence
A conversion generation processing unit, and processing results of the syntax analysis processing unit and the conversion generation processing unit.
An output unit for outputting the result of the parsing, wherein the
From the analysis result with the highest priority and the most appropriate for the user
And the analysis results determined to be
Branching in the Dependency Tree Constructing the Resulting Syntactic Structure
Solve the part, compare the analysis results for each branch unit,
The remaining part after removing the matching branch is re-
Analysis result difference extraction unit that extracts the constructed difference syntax structure
And the difference syntax structure extracted by the analysis result difference extraction unit.
In order to prioritize the analysis result selected by the user
A learning data generating unit that generates the syntactic structure learning data, and a learning data storage unit that holds the learning data, wherein the syntactic analysis processing unit is configured based on the learning data.
Machine translation characterized by changing the priority score of sentence structure
Translation device.