JP2015146216A - machine translation apparatus, machine translation method and machine translation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation apparatus capable of handling diversity of source language sentence at low cost.SOLUTION: The machine translation apparatus is an apparatus for translating a source language sentence of a first language into a target language sentence of a second language. The machine translation apparatus includes: source language sentence conversion means; translation means; and proposition sentence conversion means. The source language sentence conversion means analyzes the language of the first source language sentence on the basis of the original language propositional sentence; extracts tense, phase, and mode, or voice as expression features; separates the expression features from the source language sentence; and converts source language proposition sentence excluding the expression features. The translation means translates the source language propositional sentence into a target language proposition sentence of the second language. The proposition sentence conversion means converts the target language propositional sentence into the target language sentence of the second language based on the expression features.

Description

Embodiments described herein relate generally to a machine translation apparatus, a machine translation method, and a machine translation program.

2. Description of the Related Art In recent years, machine translation devices that translate a source language sentence expressed in a first language into a target language sentence expressed in a desired second language have been developed with the progress of natural language processing technology. Machine translation devices include a data-driven type that translates based on parallel translation examples consisting of source language sentences and target language sentences that are in translation relation, and a rule base type that translates based on rules such as grammar rules and translation rules. . In particular, the two methods described above are widely used in practice. The data driven type has the advantage that the translation result is a natural expression, and the rule-based type has the advantage that the translation is highly consistent.

However, in order to cope with the diversity of source language sentences using these methods, it is necessary to prepare a large number of parallel translation examples for the data-driven type, and various rules for the rule-based type, resulting in high development costs. There was a problem.

JP 2007-328483 A JP 2009-217360 A Japanese Patent Laid-Open No. 5-35784

The problem to be solved by the invention is to develop a machine translation apparatus capable of dealing with the diversity of source language sentences at a low cost.

The machine translation apparatus according to the present embodiment is a machine translation apparatus that translates a source language sentence in a first language into a target language sentence in a second language, and includes a source language sentence conversion unit, a translation unit, and a proposition sentence conversion unit. . The source language sentence conversion means extracts a tense, phase, law or state as an expression feature by analyzing the source language sentence in the first language, separates the expression feature from the source language sentence, Convert to a source language proposition sentence that does not include. The translating means translates the source language proposition sentence into a target language proposition sentence in the second language. The proposition sentence conversion means converts the target language proposition sentence into a target language sentence in the second language based on the expression feature.

1 is a block diagram illustrating a machine translation apparatus according to a first embodiment. The figure which shows the hardware constitutions of the machine translation apparatus of embodiment. The figure which shows an example of the analysis candidate set of embodiment. The figure which shows an example of the morpheme dictionary of embodiment. The figure which shows an example of the translation candidate set of embodiment. The flowchart of the machine translation apparatus of embodiment. The figure which shows an example of the modification expression information of embodiment. The block diagram which shows the machine translation apparatus of a modification. The block diagram which shows the machine translation apparatus of a modification.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
The machine translation apparatus according to the first embodiment is an apparatus that translates a source language sentence expressed in a first language into a target language sentence expressed in a desired second language. In this embodiment, the first language is English and the second language
Although the case where the language is Japanese will be described, the target language is not limited to these two languages.

FIG. 1 is a block diagram showing a machine translation apparatus 100 according to the first embodiment. The machine translation apparatus according to the present embodiment includes an acquisition unit 101 that acquires a source language sentence expressed in English, extracts an expression feature from the source language sentence, and converts the source language sentence into a source language proposition sentence that does not include the expression feature. A source language sentence conversion unit 102 for conversion, a translation unit 103 for translating a source language proposition sentence into a Japanese target language proposition sentence, and a combination of a target language proposition sentence having the highest translation score in the translation unit 103 and its expression feature Is selected based on the feature feature edited by the feature editing unit 105, the feature editing unit 105 that edits the expression feature selected by the maximum likelihood candidate selection unit 104, A proposition sentence conversion unit 106 that converts the target language proposition sentence selected by the unit 104 into a Japanese target language sentence; and a presentation unit 107 that presents the Japanese target language sentence to the user.

Here, the expression feature represents the subjective recognition and the utterance attitude with respect to the proposition of the speaker in the source language sentence. In this embodiment, tense, phase, law, state, and the like are used. A proposition sentence is a sentence representing an objective thing that does not include an expression feature. The source language proposition sentence is a proposition sentence expressed in English, and is an expression in which diversity is removed compared to the source language sentence. The target language proposition sentence is a Japanese proposition sentence obtained by translating an English source language proposition sentence.

The machine translation apparatus of this embodiment extracts an expression feature from a source language sentence to be translated, and converts the source language sentence into a source language proposition sentence that does not include the expression feature. Then, the source language proposition sentence is translated into the target language proposition sentence by the translation section. Since the source language proposition sentence has been removed from the diversity, the development cost of the translation section that translates the sentence can be reduced.

The machine translation apparatus according to the present embodiment converts the target language proposition sentence into the target language sentence based on the edited expression feature. As a result, it is possible to generate a target language sentence that reflects the diversity of the source language sentence and is accompanied by an expression desired by the user.

(Hardware configuration)
The machine translation apparatus according to the present embodiment is configured by hardware using a normal computer as shown in FIG. 2, and includes a control unit 201 such as a CPU (Central Processing Unit) that controls the entire apparatus, various data, ROM (Read Only Memory) and R for storing various programs
A storage unit 202 such as an AM (Random Access Memory) and an external storage unit 2 such as an HDD (Hard Disk Drive) or a CD (Compact Disk) drive device that stores various data and various programs.
03, an operation unit 204 such as a keyboard and a mouse that accepts user instruction input, a communication unit 205 that controls communication with an external device, a microphone 206 that acquires user utterances, and a sound waveform that is reproduced by reproducing an audio waveform And a display 20 for displaying video
9 and a bus 208 for connecting them.

In such a hardware configuration, the following functions are realized when the control unit 201 executes various programs stored in the storage unit 202 such as the ROM or the external storage unit 203.

(Input section)
The acquisition unit 101 acquires a source language sentence expressed in English. The user can input a source language sentence via the keyboard of the operation unit 204. Further, the user's voice acquired by the microphone 206 may be recognized to acquire the source language sentence. In addition, the source language sentence may be acquired by handwritten character recognition, or the source language sentence may be acquired from an external device connected via the communication unit 205.

(Source language sentence conversion part)
The source language sentence conversion unit 102 extracts a representation feature from the source language sentence acquired by the acquisition unit 101, and converts the source language sentence into a source language proposition sentence that does not include the expression feature. The source language sentence conversion unit 102
Analyzes source language sentences using morphological analysis technology, syntax analysis technology, anaphora analysis technology, etc. Then, using this analysis result, a plurality of expression features are extracted from the source language sentence, and the source language sentence is converted into a plurality of source language proposition sentences. Here, the morphological analysis technology includes an analysis method based on a connection cost, an analysis method based on a statistical language model, and the syntax analysis technology includes a CYK method, a generalized L
There is the R method.

In the present embodiment, tense, phase, law, state, and the like are extracted as expressive features, and a sentence obtained by separating these from the source language sentence is used as a source language proposition sentence. Here, the source language proposition sentence is an expression in which diversity is removed as compared with the source language sentence. Thereby, the development cost of the translation part 103 which translates a source language proposition sentence can be reduced.

FIG. 3 is a diagram illustrating an example of the representation feature and the source language proposition sentence in the source language sentence conversion unit 102.
The source language sentence conversion unit 102 according to the present embodiment outputs a plurality of combinations of expression features and information related thereto (expression information) and source language proposition sentences not including this expression feature as an analysis candidate set. In FIG. 3, three combinations 301 to 303 are generated from the source language sentence 309 “Shall I have him call you back when returns?”. In the combination 301, 304 represents a source language proposition sentence, and 305 represents expression information. The expression information 305 includes an expression feature 308, an identifier 306 indicating which morpheme of the source language proposition sentence 304 is associated with the expression feature 308, and a morpheme 3 of the source language proposition sentence indicated by the identifier 306.
07. The identifier 306 represents the position of the morpheme when the leading morpheme is 1 in the source language proposition sentence 304. In this example, the expression feature 308 “(current) (use (target he
)) (Proposal (Subject I)) (Question) ”is associated with morpheme 307“ calls ”.

The source language sentence conversion unit 102 extracts expression features based on the morpheme dictionary and the syntax information as shown in FIG. For example, a sentence “analyzed” can be analyzed as “analyze / can / are /” by referring to the dictionary in FIG. 4, so a proposition sentence “analyze” and “(passive) ( "Past)" can be generated. For example, `` Shall I have him call you
In the English source language sentence “back when returns?”, it can be analyzed that “Shall I” is equivalent to “Shall N”, so the expression feature (proposition (subject I)) is “have him call” is “have
Since it can be analyzed if it corresponds to “NV”, the expression feature (use (target he)) can be extracted. N represents a noun and V represents a verb.

(Translation Department)
The translation unit 103 translates an English source language proposition sentence into a Japanese target language proposition sentence. Translation Department 1
For the translation processing performed by 03, a transfer method that is a general rule-based translation method, an example-based method that is a data-driven translation method, a statistics-based method, or the like can be used.

The translation unit 103 according to the present embodiment executes a translation process on all source language proposition sentences belonging to the analysis candidate set generated by the source language sentence conversion unit 102, and translates each source language proposition sentence. And get its translation score. Then, a translation candidate including a combination of the source language proposition sentence, the expression information, the target language proposition sentence, and the translation score is generated.

Here, the translation score is an index representing the translation quality. In the case of the example-based method, the similarity between the input character string and the example is used. If it is a type translation system, a value based on the likelihood of the syntax and the priority of the rule used can be used.

FIG. 5 shows an example of a set of translation candidates (translation candidate set) output by the translation unit 103. In this figure, there are three translation candidates 501 to 503. Translation candidate 501, 50
Reference numeral 4 denotes a translation score, and 506 denotes a target language proposition sentence obtained by translating the source language proposition sentence 304. Expression information extracted by the source language sentence conversion unit 102 is added to each translation candidate.

The translation unit 103 according to the present embodiment translates a source language proposition sentence from which diversity has been removed, so that the development cost can be reduced. If it is a data-driven method, the amount of parallel translation examples to be collected can be reduced, and if it is a rule-based method, a rule to be described can be limited to knowledge about a source language proposition sentence.

(Maximum likelihood candidate selection unit)
Based on the translation score calculated by the translation unit 103, the maximum likelihood candidate selection unit 104 selects a combination having the highest translation score from combinations of expression information and target language proposition sentences belonging to the translation candidate set. The expression information and the target language proposition sentence in the selected combination are called a maximum likelihood expression feature and a maximum likelihood target language proposition sentence, respectively.

(Feature Editor)
The feature editing unit 105 edits the maximum likelihood expression feature. The feature editing unit 105 can add, delete, or change an expression feature in accordance with a user instruction from the operation unit 204. The expression feature after editing is called a modified expression feature.

As described above, the feature editing unit 105 according to the present embodiment edits the expression feature according to a user instruction. Thereby, in the proposition sentence conversion part 106 mentioned later, the target language sentence unified by the sentence style which a user desires can be produced | generated.

(Proposition sentence conversion part)
The proposition sentence conversion unit 106 converts the maximum likelihood target language proposition sentence into a Japanese target language sentence based on the revised expression feature. The proposition sentence conversion unit 106 of the present embodiment converts the maximum likelihood target language proposition sentence into the target language sentence based on the generation grammar, but other widely used language generation methods can also be used. Details of the proposition sentence conversion unit 106 will be described later.

As described above, the proposition sentence conversion unit 106 according to the present embodiment converts the maximum likelihood target language proposition sentence into the target language sentence based on the modified expression feature. As a result, it is possible to generate a target language sentence that reflects the diversity of the source language sentence and is accompanied by an expression desired by the user.

(Output part)
The presentation unit 107 presents the Japanese target language sentence generated by the proposition sentence conversion unit 106 to the user. The presentation unit 107 can display the target language sentence on the display 209 or print it out with a printer connected by the communication unit 205. In addition, the target language sentence can be converted into a speech waveform by speech synthesis and reproduced from the speaker 207.

(flowchart)
The process of the machine translation apparatus according to the present embodiment will be described using the flowchart of FIG.

First, in step S1, the acquisition unit 101 acquires an English source language sentence S. In the present embodiment, it is assumed that the source language sentence 309 “Shall I have him call you back when returns?” In FIG. 3 is acquired.

In step S2, the source language sentence conversion unit 102 analyzes the source language sentence S and extracts an analysis candidate set Cs having a combination of the expression information F and the source language proposition sentence Ps as elements. 301 in FIG.
Reference numeral 303 denotes an analysis candidate set Cs.

Here, the source language proposition sentence Ps is an expression from which diversity is removed as compared with the source language sentence S. Thereby, the development cost of the translation part 103 which translates a source language proposition sentence can be reduced. In other words, if it is a data-driven method, the amount of parallel translation examples to be collected can be reduced, and if it is a rule-based method, the rules to be described can be limited to knowledge about the source language proposition sentence. .

In step S3, the translation unit 103 translates the source language proposition sentence Ps to produce the target language proposition sentence Pt.
And its translation score V. Then, a set of translation candidates (translation candidate set Ct) including a combination of the source language proposition sentence Ps, the expression information F, the target language proposition sentence Pt, and the translation score V is generated. Reference numerals 501 to 503 in FIG. 5 represent translation candidates included in the translation candidate set Ct.

In step S4, the maximum likelihood candidate selection unit 104 obtains the target language proposition sentence Pt having the highest translation score and its expression information F from the translation candidate set Ct, respectively, and the maximum likelihood target language proposition sentence Pp.
t and maximum likelihood expression information Fp are selected. In the example of FIG. 5, the translation score 504 is 0.95.
The maximum likelihood target language proposition sentence Ppt is 304, and the maximum likelihood expression information Fp
305 are respectively selected.

In step S5, the feature editing unit 105 edits the maximum likelihood expression information Fp to change the modified expression information Fe.
Get. The feature editing unit 105 can edit the maximum likelihood expression information Fp in accordance with a user instruction. Further, the feature editing unit 105 may automatically set a preset expression feature. For example, when the source language sentence S is given as a document, an appropriate expression feature can be added in order to unify the entire document.

FIG. 7 is a diagram illustrating an example of the modified expression information Fe. In this example, as a new expression feature,
An expression feature “(Polite)” shown in 703 and an expression feature “(subject he)” shown in 704 are added. The representation feature 703 has an identifier 702 of the associated morpheme “
1-5 ". That is, the expression feature of “(Polite)” influences the entire maximum likelihood parallel translation proposition sentence Ppt. In addition, the expression feature 704 is a morpheme
”And“ he ”is supplemented as the“ subject ”of the morpheme.

In step S6, the proposition sentence conversion unit 106 converts the maximum likelihood target language proposition sentence Ppt into a Japanese target language sentence T based on the modified expression information Fe. Here, the target language sentence T can be said to be the result of converting all of the source language proposition sentence Ps and the modified expression feature Fe generated from the source language sentence S.
In FIG. 7, based on the modified expression information Fe 706, the maximum likelihood target language proposition sentence Ppt “705” is displayed.
He calls back. Return. "Is converted into a target language sentence T of 701" If he returns, will I call him back? "

In the present embodiment, the proposition sentence conversion unit 106 generates a target language sentence by operating as a reverse conversion of processing in the source language sentence conversion unit 102. For example, in FIG. 7, for the second word “calls” of the source language proposition sentence “He calls you back. Returns”, the expression feature “(present) (use (subject he)) (proposed (subject I)” ) (Question) ”. Here, assuming that the second word “calls” is translated as “call”, the proposition sentence conversion unit 106 uses the expression feature “(present)” to “call” this. ”And the same` `(
“Use (subject he))” to “call”, “(suggest (subject I))” to “call”, “(question)” to “call” Respectively. Furthermore, the expression feature “(Polite)” given to the entire source language proposition sentence is used to generate “Let's call”. Similarly, the fifth word “Returns” in the source language proposition is converted in the same manner to generate “if he returns”.

By the way, in the generation of the target language sentence in the proposition sentence conversion unit 106, in addition to the method described here, a natural language generation technique using a generation grammar, a statistical natural language generation technique using a Markov model, or the like is used. It may be used.

Finally, in step S7, the presentation unit 107 causes the target language sentence T generated in step S6.
To the user.

(Modification)
It is also possible to change the machine translation apparatus of this embodiment to the configuration as shown in FIG. 8 or FIG.

The machine translation apparatus 800 of FIG. 8 is different from the machine translation apparatus 100 of FIG. 1 in that the maximum likelihood candidate selection unit 104 and the feature editing unit 105 are not included. In this machine translation device 800, the translation unit 103 outputs only one translation candidate having the maximum likelihood translation score, whereby the same processing as the machine translation device 100 can be executed.

The machine translation apparatus 900 of FIG. 9 is different from the machine translation apparatus 100 of FIG. 1 in that the feature editing unit 105 is not included. Also in this case, the translation unit 103 selects 1 translation candidate having the most likely translation score.
The feature editing unit 105 can execute the same processing as the machine translation device 100 by outputting only one of them or by using the expression information in the translation candidate having the maximum likelihood translation score as the editing target.

(effect)
The machine translation apparatus of this embodiment extracts an expression feature from a source language sentence to be translated, and converts the source language sentence into a source language proposition sentence that does not include the expression feature. Then, the source language proposition sentence is translated into the target language proposition sentence by the translation section. Since the source language proposition sentence has been removed from the diversity, the development cost of the translation section that translates the sentence can be reduced.

The machine translation apparatus according to the present embodiment converts the target language proposition sentence into the target language sentence based on the edited expression feature. As a result, it is possible to generate a target language sentence that reflects the diversity of the source language sentence and is accompanied by an expression desired by the user.

Note that some or all of the functions in the present embodiment described above can be realized by software processing.

Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

100, 800, 900 Machine translation device 101 Acquisition unit 102 Source language sentence conversion unit 103 Translation unit 104 Maximum likelihood candidate selection unit 105 Feature editing unit 106 Proposition sentence conversion unit 107 Presentation unit 201 Control unit 202 Storage unit 203 External storage unit 204 Operation Unit 205 communication unit 206 microphone 207 speaker 208 bus 301 to 303 combination of expression information and source language proposition sentence 304 source language proposition sentence 305 expression information 308, 703, 704 expression feature 306, 702 identifier 501 to 503 translation candidate 504 translation score 506 Target language proposition sentence 701 Target language sentence 706 Modified expression information 705 Maximum likelihood target language proposition sentence

Claims (5)

A machine translation device that translates a source language sentence in a first language into a target language sentence in a second language,
The linguistic analysis of the source language sentence in the first language extracts the tense, phase, law or state as the expression feature, separates the expression feature from the source language sentence, and converts it into a source language proposition sentence that does not include the expression feature. Source language sentence conversion means for conversion;
Translation means for translating the source language proposition sentence into a target language proposition sentence in the second language;
A proposition sentence conversion means for converting the target language proposition sentence into a target language sentence in a second language based on the representation feature;
A machine translation apparatus comprising: A feature editing means for editing the expression feature;
The machine translation device according to claim 1, wherein the proposition sentence conversion means converts the target language proposition sentence into a target language sentence in a second language based on the feature feature edited expression feature. The source language sentence conversion means converts the source language sentence into a plurality of source language proposition sentences,
The translating means translates the plurality of source language proposition sentences into a plurality of target language proposition sentences in the second language;
The machine translation device according to claim 1, wherein the proposition sentence conversion unit converts the target language proposition sentence having the highest translation score in the translation unit into a target language sentence in a second language. A machine translation method for translating a source language sentence in a first language into a target language sentence in a second language,
The linguistic analysis of the source language sentence in the first language extracts the tense, phase, law or state as the expression feature, separates the expression feature from the source language sentence, and converts it into a source language proposition sentence that does not include the expression feature. Converting, and
Translating the source language proposition sentence into a target language proposition sentence in the second language;
Converting the target language proposition sentence into a target language sentence in a second language based on the representation feature;
A computer-implemented machine translation method comprising: A machine translation device that translates a source language sentence in a first language into a target language sentence in a second language,
The linguistic analysis of the source language sentence in the first language extracts the tense, phase, law or state as the expression feature, separates the expression feature from the source language sentence, and converts it into a source language proposition sentence that does not include the expression feature. The ability to convert,
A function of translating the source language proposition sentence into a target language proposition sentence in the second language;
A function of converting the target language proposition sentence into a target language sentence in a second language based on the representation feature;
Machine translation program for realizing

Images