JP3403715B2 - 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 translation device provided with a paraphrasing mechanism for both languages.

[0002]

2. Description of the Related Art Conventional automatic editing, automatic rewriting, etc. (hereinafter referred to as paraphrasing) for improving the translation quality of machine translation of difficult-to-translate expressions and syntaxes are mechanically If the machine translation mechanism is equivalent to or a part of the machine translation mechanism of the related art and a plurality of paraphrasing results are possible, it is necessary for a human operator to select one from the results. In a translation device having only such paraphrasing mechanism, an operator has knowledge of a source language and a target language (for example, in Japanese-English translation, the source language is Japanese and the target language is English), and the ability to select the result of the paraphrase. Therefore, there is a problem that users of the translation device are limited.

Alternatively, the condition of knowledge (hereinafter referred to as paraphrase knowledge) which is the basis of paraphrase operation must be set in advance in advance so that a plurality of paraphrase results cannot be obtained. In a translation device having only such a paraphrasing mechanism, a part of the load of development and preparation of knowledge that is the basis for language conversion in a translation mechanism in a narrow sense other than the paraphrasing mechanism (hereinafter referred to as language conversion) It was passed on to the load of preparation and preparation, and the performance of the translation device remained largely dependent on the knowledge and ability of the developer. Therefore, there has been a problem that development of a machine translation device requires a large amount of resources.

The object of the present invention is to solve the above problems, and in a machine translation device provided with a paraphrasing mechanism for both languages, it can be used by an operator who does not have knowledge of the source language and the target language,
It is to provide a machine translation device capable of reducing the resources required for development (for example, time, labor, and human resources).

[0005]

A machine translation device according to a first aspect of the present invention is a machine translation device for translating a source language sentence into a target language sentence, analyzes the source language sentence, and extracts words or words in the source language sentence. Extract the constituent elements of the source language sentence that is a phrase, refer to the paraphrasing operation of the source language consisting of a pair of constituent elements before and after the paraphrase of the paraphrasing operation of the source language, for the constituent elements of the source language sentence,
A first score indicating the degree of approximation of the pre-paraphrase component of the source language paraphrasing operation is calculated, and the original score including the pre-paraphrase component of the source language paraphrasing operation having the maximum first score is calculated. Select a language paraphrasing operation, paraphrase the source language sentence into a post-paraphrase source language sentence using the selected source language paraphrase operation, and select before and after processing of the selected source language paraphrase operation. The source language paraphrase processing unit that generates the difference information indicating the difference and the language conversion knowledge that is a pair of the source language constituent element and the corresponding target language constituent element are referred to A language conversion processing unit that performs language conversion into a target language sentence, and extracts the constituent elements of the target language sentence that are words or phrases of the target language sentence by analyzing the target language sentence, and based on the difference information, the target language An eye consisting of a pair of components before and after the paraphrase operation A second score indicating the degree of approximation of the component before the paraphrase of the paraphrase operation of the target language, which performs the paraphrase operation opposite to the difference information, with respect to the component of the target language sentence, with reference to the paraphrase operation of the language. And selecting a paraphrase operation of the target language that includes a component before paraphrase operation of the paraphrase operation of the target language having the maximum second score, and using the paraphrase operation of the selected target language to select the objective And a target language paraphrasing processing unit that paraphrases the language sentence into the target language sentence after paraphrasing.

A machine translation device according to the second invention is
In a machine translation device for translating a source language sentence into a target language sentence, the source language sentence is analyzed to extract the constituent elements of the source language sentence that are words or phrases of the source language sentence, and paraphrase of the paraphrase operation of the source language A first score indicating a degree of approximation of a component before paraphrase of the paraphrase operation of the source language with respect to a component of the source language sentence, with reference to a paraphrase operation of a source language composed of a pair of front and rear components And selecting a paraphrasing operation of the source language that includes a component before paraphrasing the paraphrasing operation of the source language having the maximum first score, and using the paraphrasing operation of the selected source language to select the source language With reference to the source language paraphrasing processing unit that paraphrases the language sentence into the source language sentence after paraphrase, and the language conversion knowledge that is a pair of the source language component and the corresponding target language component,
A language conversion processing unit that converts the source language sentence after the paraphrase into a target language sentence, and extracts the constituent elements of the target language sentence that are words or phrases of the target language sentence by analyzing the target language sentence,
On the basis of the difference information, referring to the paraphrasing operation of a predetermined target language consisting of a pair of components before and after paraphrasing the paraphrasing operation of the target language, the paraphrasing operation of the target language for the constituent elements of the target language sentence Calculating a second score indicating the degree of approximation of the pre-paraphrase component, and selecting the paraphrase operation of the target language including the preparaphrase component of the paraphrase operation of the target language having the maximum second score. And a target language paraphrasing processing unit that paraphrases the target language sentence into a target language sentence after paraphrasing using the paraphrasing operation of the selected target language, and the language conversion processing unit is a source of the language conversion knowledge. A third score indicating the degree of approximation of the source language component of the language conversion knowledge with respect to the source language sentence after the paraphrase is calculated with reference to the language component, and the calculated third score value Is higher than a given threshold The language conversion knowledge including the source language component of the language conversion knowledge having the maximum third score is selected, and the source language sentence after the paraphrase is converted into the target language sentence using the selected language conversion knowledge. While converting the language to
When the calculated value of the third score is equal to or less than the threshold value, the source language constituent element of the source language of the language conversion knowledge having the maximum third score is selected as the translatable sentence information, and the source language paraphrasing process is performed. The source language paraphrasing processing unit outputs, to the source language paraphrase processing unit, the degree of approximation of the pre-paraphrase component of the paraphrase operation of the source language with respect to the component of the source language sentence, and the paraphrase of the source language with respect to the translatable sentence information. A first score indicating the degree of approximation of the component after the paraphrase of the operation is calculated, and the paraphrase operation of the source language including the components before and after the paraphrase of the paraphrase operation of the source language having the maximum first score is performed. It is characterized in that the source language sentence is paraphrased into the source language sentence after paraphrase by using the paraphrase operation of the selected source language.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of a speech translation synthesis system including a machine translation device 3 according to an embodiment of the present invention. This speech translation synthesis system comprises a speech recognition device 2, a machine translation device 3 and a speech synthesis device 4, and the machine translation device 3 has a certain source language sentence or object in order to solve the problems of the prior art. As shown in FIG. 2, a paraphrasing mechanism for both languages that automatically selects one from a plurality of possible paraphrases for a source language sentence translated into a language is provided on both sides of the language conversion processing unit 12. Source language paraphrasing processing unit 11
And (A) the source language paraphrasing processing unit 11 is provided in the target language paraphrasing processing unit 13 to generate difference information indicating the difference between before and after the processing of the paraphrasing operation of the selected source language. Refers to the paraphrasing operation of the target language consisting of a pair of constituent elements before and after paraphrasing the paraphrasing operation of the target language based on the difference information, and the opposite of the difference information to the constituent element of the target language sentence. A second score indicating the degree of approximation of the pre-paraphrase component of the paraphrasing operation of the target language that performs paraphrasing operation is calculated, and the pre-paraphrase component of the paraphrasing operation of the target language having the maximum second score is calculated. And selecting the paraphrase operation of the target language including (B) the language conversion processing unit 12 with reference to the constituent elements of the source language of the language conversion knowledge, Source language of translation knowledge Is calculated, and when the calculated third score value is less than or equal to the threshold value, the source language component of the language conversion knowledge having the maximum third score. As the translatable sentence information and returning it to the source language paraphrasing processing unit 11 for output, and the source language paraphrasing processing unit 11 outputs the component before the paraphrase of the paraphrase operation of the source language with respect to the component of the source language sentence. And the degree of approximation of the constituents after paraphrase of the paraphrasing operation of the source language with respect to the translatable sentence information are calculated, and the source language having the maximum first score is calculated. Selecting a paraphrasing operation of a source language including components before and after paraphrasing the paraphrasing operation, and paraphrasing the source language sentence to the source language sentence after the paraphrase by using the paraphrasing operation of the selected source language. I am trying.

In FIG. 1, the speech translation synthesis system is
Microphone 1, voice recognition device 2, machine translation device 3
And a voice synthesizer 4 and a speaker 5. Here, the voice recognition device 2 and the machine translation device 3 configure a voice translation system, and the machine translation device 3 and the voice synthesis device 4 configure a translated voice synthesis system.

The uttered voice of the source language sentence input to the microphone 1 is converted to an electric signal and then input to the voice recognition device 2, which is input using a known voice recognition method. The voice signal of the language sentence is recognized by voice and converted into a character string of the source language sentence and output to the machine translation device 3. Then, the machine translation device 3 machine-translates the input character string of the source language sentence into a character string of the target language sentence by using a method described in detail later, and then outputs the character string to the speech synthesis device 4. Furthermore, the speech synthesizer 4 uses a known speech synthesis method,
The character string of the input target language sentence is voice-synthesized into a voice signal, and a synthesized voice is output from the speaker 5.

FIG. 2 is a block diagram showing a detailed configuration of the machine translation device 3 of FIG. Machine translation device 3 of this embodiment
Is a processing control unit 10, and a source language paraphrasing processing unit 11 and a language conversion processing unit 12 respectively connected to the processing control unit 10.
And a target language paraphrasing processing unit 13. The processing control unit 10 outputs the source language sentence input from the speech recognition device 2 to the source language paraphrase processing unit 11, and the source language paraphrase processing unit 11 paraphrases the source language sentence into an expression and a syntax suitable for machine translation. Then, the source language paraphrase text is generated, and the source language paraphrase text and the difference information of the paraphrase knowledge related to the paraphrase are output to the processing control unit 10. As a result of this paraphrasing, the meaning of words in the source language sentence, the structure of the parallel syntax, etc. are specified, and thereby the ambiguity of the source language sentence is eliminated. Next, the processing control unit 10 inputs the source language paraphrase sentence as a translation source sentence into the language conversion processing unit 12, and the language conversion processing unit 12 generates a target language translation sentence by performing the language conversion of the translation source sentence into a target language. Processing control unit 1
0, or when the language conversion is not possible, translatable sentence information is generated and output to the processing control unit 10. The processing control unit 10 stores the generated translatable sentence information and the corresponding source language sentence in the source language paraphrasing unit 11.
Then, the source language paraphrasing processing unit 11 refers to the translatable sentence information and executes a more appropriate paraphrase for generating the target language translated sentence. Next, the processing control unit 10 generates a target language translation sentence and difference information opposite to the generated difference information, and outputs the difference information to the target language paraphrasing processing unit 13. In order to eliminate the difference between the source language sentence and the source language sentence generated by the language paraphrasing process, the target language translation sentence is paraphrased into the target language sentence after paraphrase, using the opposite difference information. It is output to the processing control unit 10. Finally, the processing control unit 10 outputs the target language sentence to the speech synthesizer 4.

FIG. 3 is a block diagram showing a detailed configuration of the source language paraphrasing processing unit 11 of FIG. In FIG. 3, the source language paraphrase processing unit 11 includes an approximate collation processing unit 112, a source language analysis processing unit 111, a source language paraphrase sentence generation unit 113, and a paraphrase knowledge memory 114, which are connected to the approximate collation processing unit 112, respectively. The history information memory 115 connected to the approximate collation processing unit 112 and the source language analysis processing unit 111.

The source language analysis processing unit 111 morphologically analyzes the source language sentence by a known morpheme analysis method used in a known machine translation device, and each component of the source language sentence (for example, a certain word or phrase). ) Is extracted and the source language sentence is transformed into a format that can be collated with the paraphrase knowledge stored in the paraphrase knowledge memory 114 to obtain the source language sentence analysis result. Here, the paraphrase knowledge is paraphrasing operation which is information in which a constituent element (or character string) before paraphrase of the source language and a constituent element (or character string) after paraphrase are paired, and the constituent element before paraphrase. And the difference information indicating the difference between the component after the above paraphrase. Next, when the translatable sentence information generated by the language conversion processing unit 12 is not input, the approximate matching processing unit 112 matches the source language sentence analysis result with the paraphrase knowledge paraphrase operation, and uses a predetermined score calculation method. Then, a first score indicating the degree of approximation between each constituent element of the source language sentence analysis result and the constituent element before the paraphrase related to the paraphrase operation applied thereto is calculated, and the maximum first score is calculated.
And selecting the paraphrase knowledge describing the paraphrase operation applicable to the source language sentence analysis result, and the selected paraphrase knowledge,
The source language sentence analysis result and the source language paraphrase sentence generation unit 113 are output. The calculation of the first score can be performed by comparing various likelihoods among a plurality of paraphrase candidates, for example, by using an appearance probability in a separately prepared monolingual corpus. The source language paraphrase sentence generation unit 113 paraphrases the source language paraphrase sentence result into a source language paraphrase sentence by using a paraphrase operation of paraphrase knowledge, and outputs the result. Here, the known method is, for example, a means such as a morphological analysis means, a conversion means, a generation means or the like used in a known machine translation device, except that the source language and the target language are the same language. ,
The source language paraphrase sentence generation unit 113 is equivalent to or a part of the mechanism of a known machine translation device. Further, the source language paraphrase text generation unit 113 outputs the paraphrase knowledge difference information indicating the difference before and after the processing of the selected paraphrase operation. The source language paraphrasing unit 11 stores in the history information memory 115 each source language sentence to be input and the information of each item shown in the following table corresponding thereto for reference.

[0014]

[Table 1] ―――――――――――――――――――――――――――――――――――― Source language sentence Morphological analysis result of source language sentence Corresponding paraphrase knowledge Source language paraphrase sentence ――――――――――――――――――――――――――――――――――――

On the other hand, when the translatable sentence information generated by the language conversion processing unit 12 is input, the approximate collation processing unit 112 collates the translatable sentence information, the source language sentence analysis result and the paraphrase knowledge paraphrase operation. Then, the degree of approximation between each component of the source language sentence analysis result using a predetermined score calculation method and the component before paraphrase relating to the paraphrase operation applied thereto,
A first score indicating the degree of approximation between the translatable sentence information and the components after the paraphrasing operation is calculated, and has a maximum first score, and is applicable to the source language sentence analysis result. A paraphrase knowledge describing a simple paraphrase operation is selected, and the selected paraphrase knowledge and the source language sentence analysis result are combined into the source language paraphrase sentence generation unit 113.
Output to. However, other operations of the source language paraphrasing processing unit 11 are the same as when the translatable sentence information is not input.

FIG. 4 is a block diagram showing the detailed arrangement of the language conversion processing unit 12 shown in FIG. The language conversion processing unit 12
The source language collation processing unit 121 and the source language collation processing unit 12
1 and an approximate matching determination unit 122 and a language conversion knowledge memory 124, and a target language sentence generation unit 123 connected to the approximate matching determination unit 122. The source language collation processing unit 121 includes the source language paraphrase text input as the source translation text from the processing control unit 10 and the language conversion knowledge memory 124.
The source language information of the language conversion knowledge stored in
A third score indicating the degree of approximation between the translation source sentence and the source language information of the language conversion knowledge applied thereto is calculated using a predetermined score calculation method, and is shown by the calculated third score value. The linguistic conversion knowledge having a higher possibility of being the basis for linguistic conversion in the linguistic conversion processing, which is higher than a certain threshold, and the translated original sentence are output to the approximate matching determination unit 122. Third above
For the calculation of the score of, for example, a publicly known method is used to quantify the syntactical similarity between the translation source sentence and the source language information of the language conversion knowledge, and thereby the score is calculated for each source language information. . The language conversion knowledge is stored in the language conversion knowledge memory 12
4 is information in which the constituent elements of the source language (that is, the source language information) and the constituent elements of the target language (that is, the target language information) are stored as a pair, whereby the language conversion knowledge memory 124 operates as a bilingual dictionary. To do. Here, when the language conversion knowledge having the third score higher than the threshold value is not found, the translation by the machine translation device 3 is not executed. The approximate matching determination unit 122 determines whether or not to perform the language conversion by sequentially matching the translation source sentence and the source language information of the language conversion knowledge, and when performing the language conversion, the language conversion having the maximum third score. Knowledge is selected and output to the target language sentence generation unit 123 together with the translated original sentence. Further, when the values of the third scores are all equal to or less than a predetermined threshold value and no language conversion is performed, the closest approximation to the translated original sentence is made. That is, the source language information of the language conversion knowledge having the maximum score is output as translatable sentence information. The approximate matching determination unit 122 determines whether or not to perform the language conversion by gradually relaxing the approximate matching based on the number of times of the approximate matching with respect to the same translated original sentence, the elapsed time of the translation process, and the like. The target language sentence generator 123
For example, a translation source sentence is language-converted into a target language translation sentence by a publicly known method utilized in a publicly known machine translation device, and is output. The target language sentence generator 123 is equivalent to or part of the mechanism of a known machine translation device.

FIG. 5 is a block diagram showing a detailed configuration of the target language paraphrasing processing unit 13 of FIG. The target language paraphrasing processing unit 13 includes an approximate matching processing unit 132 and a translation sentence analysis processing unit 131 connected to the approximate matching processing unit 132,
Target language paraphrase sentence generation unit 133 and paraphrase knowledge memory 134
And is configured. The translation sentence analysis processing unit 131 morphologically analyzes the target language translation sentence by using, for example, a known morphological analysis method used in a known machine translation device, and analyzes each component of the target language translation sentence (for example, a certain word). Or phrase)
Is extracted and transformed into a format that can be collated with the paraphrase knowledge stored in the paraphrase knowledge memory 134 to obtain a translation sentence analysis result. Here, the paraphrase knowledge is paraphrasing operation which is information in which a component (or character string) before paraphrase of the target language and a component (or character string) after paraphrase are paired, and before paraphrase by the paraphrase operation. Is a set of data including difference information indicating a difference between the constituent element and the constituent element after the paraphrase. Approximate matching processing unit 13
2 collates the difference information input from the processing control unit 10, the translated sentence analysis result, and the paraphrase operation of paraphrase knowledge and the difference information stored in the paraphrase knowledge memory 134, and uses a predetermined score calculation method. , Calculating a second score indicating the degree of approximation of the pre-paraphrase component included in the paraphrase operation in which the opposite difference information generated by the processing control unit 10 is reflected for each component of the translated sentence analysis result Then, the paraphrase knowledge having the maximum second score and describing the paraphrase operation applicable to the translation sentence analysis result is selected, and the selected paraphrase knowledge and the translation sentence analysis result are selected as the target language paraphrase sentence generation unit. It outputs to 133.
The calculation of the second score is performed by comparing various likelihoods among a plurality of paraphrase candidates by using, for example, an appearance probability in a separately prepared monolingual corpus, as in the method of calculating the first score. can do. The target language paraphrase sentence generation unit 133 uses a known method to paraphrase the translation sentence analysis result into the target language paraphrase sentence by using the paraphrase operation of the paraphrase knowledge and outputs it. Here, the known method is, for example, a morphological analysis unit used in a known machine translation device,
The target language paraphrase sentence generation unit 1 is a means such as a conversion means and a generation means, except that the source language and the target language are the same language.
The numeral 33 is equivalent to or a part of a mechanism of a known machine translation device.

In the machine translation device 3 of the present embodiment, the source language paraphrasing processing unit 11 and the language conversion processing unit 12 are used to automatically select from a plurality of paraphrasing operations that are applicable in the source language paraphrasing. As a result, by providing a mechanism for automatically selecting one, one is automatically selected from a plurality of possible paraphrasing results. In addition, through the configuration of the source language paraphrasing processing unit 11, the language conversion processing unit 12, and the target language paraphrasing processing unit 13, in the paraphrase of the target language, one of a plurality of paraphrasing operations that can be applied is automatically performed. As a result, by providing the mechanism for selecting, one is automatically selected from the results of the paraphrases that are possible.

[0019]

EXAMPLE The operation of the machine translation apparatus according to the embodiment of the present invention will be described below with reference to example sentences. For the sake of explanation, it is assumed that the source language is English and the target language is Japanese, but in another embodiment, the present invention is used for translating from Japanese to English or between different languages. It is not restricted to be used for. Here, the source language sentence is "Could I have you
r name, sir? ".

The processing control unit 10 uses the source language sentence "Could I ha
ve your name, sir? "is input, the source language paraphrase processing unit 11 is input with the source language paraphrase processing unit 11 and the source language paraphrase processing unit 11 is input.
Wait for output from.

The source language paraphrase processing unit 11 uses the source language sentence "Coul.
When "d I have your name, sir?" is input, the source language analysis processing unit 111 morphologically analyzes the source language sentence and transforms it into a format that can be collated with paraphrase knowledge to obtain the source language sentence analysis result.
In this example, the source language sentence analysis result is as follows in terms of notation.

[0022]

[Table 2] ―――――――――――――――――――――――――――――――――――― <Start of sentence> could I have your name <reading mark> sir <end of sentence> ――――――――――――――――――――――――――――――――――――

The paraphrase knowledge is the information stored in the paraphrase knowledge memory 114, which is a pair of the paraphrase operation in the source language and the difference information. An example is shown in the table below.

[0024]

[Table 3] ――――――――――――――――――――――――――――――――――――               Paraphrase operation Difference information ―――――――――――――――――――――――――――――――――――― <Text> could I → <text> can I politeness − sir <end of sentence> → (empty) <end of sentence> politeness − ――――――――――――――――――――――――――――――――――――

The paraphrasing operation in the source language is, for example, information in which a constituent element before a paraphrase of a certain word or phrase and a constituent element after the paraphrase are paired in the source language. Of course, the paraphrasing operation may be described as an operation such as “sir <end of sentence>: delete first element”. Another example of the difference information may include, but is not limited to, emphasis and euphemism, feminine expression strength, masculine expression strength, and the like.
In this embodiment, only the increase / decrease represented by + and − with respect to “politeness” will be considered.

Since the translatable sentence information is not input at the initial stage, the approximate collation processing unit 112 collates the source language sentence analysis result with the paraphrase knowledge, and uses a predetermined score calculation method to calculate the source language sentence. The first score showing the degree of approximation of the pre-paraphrase component of the source language paraphrase operation to the analysis result is calculated, and the maximum score is obtained from the paraphrase knowledge describing the paraphrase operation applicable to the source language sentence analysis result. The paraphrase knowledge having the first score is selected, and the paraphrase knowledge and the source language sentence analysis result are output to the source language paraphrase sentence generation unit 113. In the present embodiment, it is assumed that the above-mentioned source language sentence analysis result and two paraphrase knowledge are output.

The source language paraphrase sentence generating unit 113 paraphrases the source language paraphrase sentence into the source language paraphrase sentence by using the paraphrase operation of the paraphrase knowledge of the source language by means of the conventional technique. In this embodiment, the source language paraphrase is “Can I have your name?”. Further, the source language paraphrase sentence generation unit 113 outputs the paraphrase knowledge difference information indicating the difference before and after the processing of the selected paraphrase operation. In the present embodiment, the difference information is “Politeness-” is 2
It becomes a matter.

The source language paraphrasing processing unit 11 saves the progress of processing in the history information memory 115 to avoid repeating the same processing.

When the source language paraphrase sentence and the difference information are output from the source language paraphrase processing unit 11, the processing control unit 10 sets the difference information to the difference information of the paraphrase operation opposite to the difference information. Convert to and save temporarily. This conversion is
For example, the correspondence table incorporated in advance is referred to, and in the present embodiment, the two "weights-" are set to "weights +". The difference information will be described later in connection with the description of the target language paraphrasing processing unit 13.

The processing control unit 10 inputs the above-mentioned original language paraphrase sentence into the language conversion processing unit 12 as a translation source sentence, and waits for an output from the language conversion processing unit 12.

In the operation of the language conversion processing unit 12, first, the source language collation processing unit 121 determines that the translated original sentence "Can I have.
When "your name?" is input, the translation source sentence is collated with the source language information of the language translation knowledge stored in the language translation knowledge memory 124, and the language translation knowledge for the translation source sentence is calculated using a predetermined score calculation method. A third score indicating the degree of approximation of the source language information of is calculated, and the possibility of being the basis of the language conversion in the language conversion processing, which is represented by the value of the calculated third score, is higher than a certain threshold value. The language conversion knowledge and the translated original sentence are selected and output to the approximate matching determination unit 122.

The language conversion knowledge is stored in the language conversion knowledge memory 12
It is the information stored in No. 4 as a pair of the source language information and the target language information as shown below.

[0033]

[Table 4] ――――――――――――――――――――――――――――――――――――         Source language information Target language information ―――――――――――――――――――――――――――――――――――― May I have your name? Please tell me your name. Can I help you? Yes, please. Can I <VP>? Can I use <VP>? ――――――――――――――――――――――――――――――――――――

Here, <VP> represents a verb phrase. In the present embodiment, it is assumed that the above-mentioned three pieces of language conversion knowledge are output as those that may be higher than the above threshold value.

The approximate collation judging unit 122 sequentially collates the translation source sentence with the source language information of the language conversion knowledge to determine whether or not to perform the language conversion. When performing the language conversion, the language conversion knowledge and the translation source sentence are compared. Is output to the target language sentence generation unit 123, and in the case where the language is not converted, the source language information of the language conversion knowledge that is the closest to the translated original sentence, that is, has the maximum third score is used as the translatable sentence information as the process control unit. Output to 10.
At this time, the approximate matching determination unit 122 determines whether or not to perform the language conversion by gradually relaxing the approximate matching based on, for example, the number of times of the approximate matching for the same translated original sentence and the elapsed time of the translation process.

In other words, the original translation "Can I have your name
"" For source language information May I have your name?
Is relatively approximate information in which only one word has been replaced, but since it is the first approximate matching determination at the present time in the description, it is determined that the language conversion is not performed. However, depending on the setting of the relaxation of the approximate matching, for example, when the second approximate matching determination is made or the elapsed time of the translation processing exceeds 1 second, it is determined that the language conversion is performed with the language conversion knowledge. It will be.

In this embodiment, out of the above three pieces of language conversion knowledge, the source language information May I have yourname? Is the source language information that is the closest to the translated source sentence, and is output as translatable sentence information. The translatable sentence information is, for example, "May", "
It is a list of the words I "," have "," your "," name ","? ".

When the translatable sentence information is output from the language conversion processing unit 12, the processing control unit 10 outputs the source language sentence "Co
uld I have your name, sir? "
y, "I", "have", "your", "name", and "?" are output to the source language paraphrasing processing unit 11 and the source language paraphrasing processing unit 11 waits for output.

In the above steps of this embodiment, since no language conversion knowledge applicable to the source language paraphrase sentence was found, the source language paraphrase processing unit 11 next refers to the translatable sentence information to refer to the source language paraphrase information. Execute the paraphrase of the sentence again. The source language paraphrasing processing unit 11 obtains a source language sentence analysis result of the source language sentence from the history information stored in the history information memory 115 in which the progress of processing is stored, and together with the translatable sentence information,
The approximate collation processing unit 112 performs collation with the paraphrase knowledge of the source language, and by a predetermined score calculation method, the degree of approximation of the components before the paraphrase of the paraphrase operation with respect to the source language sentence and the paraphrase operation with respect to the translatable sentence information. Compute a first score that indicates the degree of approximation of the component after paraphrasing. If the source language sentence analysis result of the source language sentence is not stored, the source language analysis processing unit 111 performs morphological analysis.

The approximate matching processing unit 112 refers to the calculated value of the first score, and in the paraphrase knowledge stored in the paraphrase knowledge memory 114, uses the translatable sentence information as the basis for the language conversion in the language conversion processing. If there is something that exceeds the above threshold, it is added to the paraphrase knowledge to be output.

The following paraphrase knowledge, for example, is assumed not to exist in advance in the paraphrase knowledge memory 114 in this embodiment.

[0042]

[Table 5] ――――――――――――――――――――――――――――――――――――               Paraphrase operation Difference information ―――――――――――――――――――――――――――――――――――― <Sentence> could I → <Sentence> may I Courtesy − ――――――――――――――――――――――――――――――――――――

Further, the source language paraphrasing processing unit 11 morphologically analyzes the source language paraphrase sentence "Can Ihave your name?" For the source language sentence by the source language analysis processing unit 111 to obtain the source language sentence analysis result, Together with the translatable sentence information, the approximate collation processing unit 112 collates the paraphrase knowledge. If a source language sentence analysis result of the source language paraphrase sentence is stored in the history information memory 115 as the processing progress of the source language paraphrase sentence generation unit 113, the morpheme analysis process is omitted by using it. To do. Hereinafter, such supplementary description will be omitted.

[0044]

[Table 6] ――――――――――――――――――――――――――――――――――――               Paraphrase operation Difference information ―――――――――――――――――――――――――――――――――――― <Start> can I → <Start> may I (empty) ――――――――――――――――――――――――――――――――――――

The approximate collation processing unit 112, similarly to the above,
The first result indicating the degree of approximation of the pre-paraphrase component of the paraphrase operation of the source language with respect to the source language sentence analysis result using a predetermined score calculation method by collating the source language sentence analysis result with the paraphrase knowledge. The paraphrase knowledge having the maximum first score is selected from the paraphrase knowledge describing the paraphrase operation applicable to the source language sentence analysis result by calculating the score, and the paraphrase knowledge and the source language sentence analysis result are selected. It is output to the source language paraphrase sentence generation unit 113. In the present embodiment, it is assumed that the source language sentence analysis result and the paraphrase knowledge described above are output.

The source language paraphrase sentence generation unit 113 paraphrases the source language parsing result into the source language paraphrase sentence and outputs it. In this embodiment, the source language paraphrase is "May I have your n
In addition, the source language paraphrase sentence generation unit 113 replaces the paraphrase knowledge difference information indicating the difference before and after the selected paraphrase operation with the paraphrase knowledge difference information of the previous paraphrase of the source language sentence. In addition, since there is no difference information in the paraphrase knowledge in Table 5, the difference information in the present embodiment is "Collecting-" is still two.

The processing control unit 10 outputs the original language paraphrase sentence as a translation source sentence to the language conversion processing unit 12, and waits for the output from the language conversion processing unit 12.

The language conversion processing unit 12 uses the translated original sentence "May I
If "have your name?" is input, the source language collation processing unit 121 collates the source language information with the source language information of the language conversion knowledge in the same manner as above, and the language for the source language is translated using a predetermined score calculation method. A third score, which indicates the degree of approximation of the source language information of the conversion knowledge, is calculated, and the possibility of being the basis of the language conversion in the language conversion processing indicated by the calculated third score value is greater than a certain threshold value. The high language conversion knowledge and the translated original text are output to the approximate matching determination unit 122.

The language conversion knowledge is information in which the source language information and the target language information are paired.

[0050]

[Table 7] ――――――――――――――――――――――――――――――――――――         Source language information Target language information ―――――――――――――――――――――――――――――――――――― May I have your name? Please tell me your name. May I help you? Yes, please. May I <VP>? Can I use <VP>? ――――――――――――――――――――――――――――――――――――

In this embodiment, it is assumed that the above-mentioned three pieces of language conversion knowledge are output as those that may be higher than the above threshold value.

The approximate collation judging unit 122, similarly to the above,
The translation source sentence and the source language information of the language conversion knowledge are sequentially compared to determine whether or not to perform the language conversion. When performing the language conversion, the language conversion knowledge and the translation source sentence are combined into the target language sentence generation unit 1
When the language is not converted, the source language information of the language conversion knowledge that is the closest to the translation source sentence, that is, has the maximum third score is output as the translatable sentence information.

In this embodiment, it is determined that language conversion is to be performed based on the following language conversion knowledge.

[0054]

[Table 8] ――――――――――――――――――――――――――――――――――――         Source language information Target language information ―――――――――――――――――――――――――――――――――――― May I have your name? Please tell me your name. ――――――――――――――――――――――――――――――――――――

The approximate matching determination unit 122 determines that the translated original sentence "May
"I have your name?" And one piece of the above language conversion knowledge are output to the target language sentence generation unit 123.

The target language sentence generator 123 converts the source language sentence into a target language translated sentence by using the language conversion knowledge, and outputs the translated sentence by the conventional technique. In this embodiment, the target language translation is "Please tell me your name."

Next, the paraphrasing process in the target language executed by the target language paraphrasing unit 13 will be described. In the present embodiment, description will be made with the same source language sentence as the description of the paraphrasing process in the source language, that is, the process executed by the source language paraphrasing processing unit 11 and the language conversion processing unit 12,
The description of the part overlapping with the paraphrasing process in the source language will be omitted and will be described as a continuation.

When the source language paraphrase sentence and the difference information are output from the source language paraphrase processing unit 11, the processing control unit 10 sets the input difference information to the difference of the paraphrase operation opposite to the difference information. Convert it to information and save it temporarily. This conversion is performed, for example, by referring to a correspondence table incorporated in advance, and in the present embodiment, "politeness-" is taken as "politeness +".
In this example, the difference information was not added in the second paraphrase.

The processing control unit 10 outputs the above two pieces of difference information "Dignity +" and the target language translation "Tell me your name." To the target language paraphrasing processing unit 13, and the target language paraphrase is output. The output from the processing unit 13 is awaited.

When the target language translation sentence, "Tell me your name," is input to the target language paraphrase processing unit 13, the translation sentence analysis processing unit 131 morphologically analyzes the target language translation sentence and collates it with the paraphrase knowledge. It is transformed into a possible format and used as the translated sentence analysis result. In this embodiment, the translated sentence analysis result is as follows in terms of notation.

[0061]

[Table 9] ―――――――――――――――――――――――――――――――――――― <Beginning> Tell us your name <End> ――――――――――――――――――――――――――――――――――――

The paraphrase knowledge is the information stored in the paraphrase knowledge memory 134, which is a pair of the paraphrase operation and the difference information in the target language. The paraphrase operation is, for example, information in which the pre-paraphrase component and the post-paraphrase component of the target language are paired.

[0063]

[Table 10] ――――――――――――――――――――――――――――――――――――               Paraphrase operation Difference information ―――――――――――――――――――――――――――――――――――― Your → Your politeness + Your → Your politeness − Please tell me <End of sentence> → Can you accept it <End of sentence> Courtesy + Please <end of sentence> → can you please <end of sentence> politeness + Please tell me <End of sentence> → Notify <End of sentence> High pressure + ――――――――――――――――――――――――――――――――――――

The approximate collation processing unit 132 collates the translation sentence analysis result with the paraphrase knowledge and the difference information input from the processing control unit 10, and uses a predetermined score calculation method to calculate the translation sentence analysis result and the translation sentence analysis result. A second score of the paraphrase knowledge of the target language for the difference information is calculated. Approximate matching processing unit 13
2 selects the paraphrase knowledge describing the paraphrase operation applicable to the translated sentence analysis result and the difference information opposite to the difference information input from the processing control unit 10 based on the calculated second score. The paraphrase knowledge and the translated sentence analysis result are output to the target language paraphrase sentence generation unit 133. In the present embodiment, it is assumed that the above translated sentence analysis result and three out of the five paraphrase knowledge illustrated in the table including the difference information including “politeness +” are output.

The target language paraphrase sentence generation unit 133 refers to the grammatical knowledge of the target language stored in advance, determines the difference information to be actually applied, and uses the conventional technique to convert the translated sentence analysis result to the target language. The target language paraphrase text is paraphrased and output using the paraphrase operation of paraphrase knowledge. In this embodiment, the target language paraphrase is "Can you give me your name?"

When the target language paraphrase sentence is output from the target language paraphrase processing unit 13, the processing control unit 10 outputs it as a target language sentence to the speech synthesizer 4 for inputting the next source language sentence. stand by. Here, the information related to one source language sentence stored in the history information memory 115 may be updated for each input of the source language sentence, or may be accumulated without being updated.

In the above embodiment, the source language paraphrasing processing unit 11 passes the difference information between the source language sentence and the paraphrased source language sentence to the target language paraphrasing processing unit 13, and the target language paraphrasing processing unit 13 sends the difference information. It may be configured such that only the most appropriate paraphrased target language sentence is automatically selected and output based on the above.

In the above embodiment, the language conversion processing unit 12 feeds back the translatable sentence information to the source language paraphrasing processing unit 11, and the source language paraphrasing processing unit 11 automatically makes a proper paraphrase based on it. You may comprise so that only the structure which outputs a source language sentence may be provided.

In the above embodiment, the sentence is converted into its constituent elements by morphological analysis, but syntactic analysis may be used instead of morphological analysis.

[0070]

As described above in detail, the machine translation apparatus according to the present invention automatically selects from a plurality of possible paraphrases for a source language sentence or a source language sentence translated into a target language. A paraphrasing mechanism for both languages for selectively selecting one is provided in the source language paraphrasing processing unit and the target language paraphrasing processing unit on both sides of the language conversion processing unit, and the source language paraphrasing processing unit 11 is provided in the target language paraphrasing processing unit 13. The difference information between the source language sentence and the paraphrased source language sentence is delivered, and the target language paraphrase processing unit 13 automatically selects and outputs the most appropriate paraphrased target language sentence based on the difference information, or / And the language conversion processing unit 12 feeds back the translatable sentence information to the source language paraphrasing processing unit 11,
The source language paraphrasing processing unit 11 automatically outputs a source language sentence that is originally paraphrased based on the source language paraphrase processing unit 11.

Therefore, since the difference information in the paraphrase in the source language is used for the paraphrase in the target language, it is possible to improve the accuracy in automatically selecting the vocabulary to be paraphrased in the target language. Therefore, it is not necessary for the operator of the machine translation device to have knowledge of the source language and the target language, and the use of the machine translation device is expanded. For example, as shown in FIG. 1, a speech translation (interpretation) device can be combined with a speech recognition device or a speech synthesis device to allow an ordinary person to take it on an overseas trip. Also, since it is possible to automatically select an appropriate translated word when executing translation and improve paraphrase accuracy, it is possible to reduce the prior knowledge necessary for developing a machine translation device, which is necessary for developing a machine translation device. Resource can be reduced. This facilitates the development of a machine translation device between languages other than Japanese and English, for example. In addition, even with Japanese-English translation devices, currently some technical changes such as replacement of the technical term dictionary are compatible with some fields and applications.
It facilitates the development of machine translation devices specialized for fields and applications.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a speech translation synthesis system including a machine translation device 3 according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a detailed configuration of a machine translation device 3 of FIG.

3 is a block diagram showing a detailed configuration of a source language paraphrasing processing unit 11 in FIG.

4 is a block diagram showing a detailed configuration of a language conversion processing unit 12 in FIG.

5 is a block diagram showing a detailed configuration of a target language paraphrasing processing unit 13 in FIG.

[Explanation of symbols]

1 ... Microphone, 2 ... voice recognition device, 3 ... Machine translation device, 4 ... voice synthesizer, 5 ... speaker, 10 ... Process control unit, 11 ... Source language paraphrasing unit, 12 ... Language conversion processing unit, 13 ... Target language paraphrasing processing unit, 111 ... Source language analysis processing unit, 112, 132 ... Approximate matching processing unit, 113 ... Source language paraphrase sentence generation unit, 114,134 ... paraphrase knowledge memory, 115 ... History information memory, 121 ... Source language collation processing unit, 122 ... Approximate matching determination unit, 123 ... a target language sentence generator, 124 ... Language conversion knowledge memory, 131 ... Translated sentence analysis processing unit, 133 ... Target language paraphrase sentence generation unit.

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor, Tama Kiyoshi, 2-2, Kodai 2-chome, Seika-cho, Soraku-gun, Kyoto Pref., Spoken Language Communication Research Laboratories, Ltd. (56) Bibliography Satoshi Sato, title of paper In other words, Journal of Information Processing Society of Japan, Japan, July 15, 1999, Vol. 40, No. 7, p. 2937-p. 2945 Kazuhide Yamamoto, Satoshi Shirai, Hitoshi Sakamoto, Kiyoshi Zhangtama, SANDGLASS: Speech translation based on bilingual paraphrasing mechanism, Proceedings of the 7th Annual Conference of Linguistic Processing Society, Japan, March 27, 2001, p. 221-p. 224, "1 Motivation and Proposal Outline" and "References" (58) Fields investigated (Int.Cl. ⁷ , DB name) G06F 17/21-17/28 JISST file (JOIS)

Claims (3)

(57) [Claims] 1. A machine translation device for translating a source language sentence into a target language sentence, analyzing the source language sentence to extract constituent elements of the source language sentence, which are words or phrases of the source language sentence, and extract the source language sentence. With reference to the source language paraphrasing operation consisting of a pair of components before and after the paraphrase of the paraphrasing operation, the degree of approximation of the component before the paraphrase of the source language paraphrasing operation with respect to the component of the source language sentence is described. Calculating a first score, selecting the source language paraphrase operation that includes the pre-paraphrase component of the source language paraphrase operation having the largest first score, and selecting the selected source language paraphrase operation A source language paraphrase processing unit that paraphrases the source language sentence into a source language sentence after paraphrase using, and generates difference information that indicates a difference between before and after processing of the selected source language paraphrase operation; In pairs with the components of the corresponding target language A language conversion processing unit that converts the source language sentence after the paraphrase into a target language sentence by referring to the language conversion knowledge, and a target language that is a word or a phrase of the target language sentence by analyzing the target language sentence. Extract the constituent elements of the sentence, based on the difference information, referring to the paraphrasing operation of the target language consisting of a pair of constituent elements before and after paraphrasing the paraphrasing operation of the target language, for the constituent elements of the target language sentence, A second score indicating the degree of approximation of the pre-paraphrase component of the paraphrase operation of the target language that performs the paraphrase operation opposite to the difference information is calculated, and the paraphrase operation of the target language having the maximum second score is calculated. And a target language paraphrasing processing unit that paraphrases the target language sentence into a post-paraphrase target language sentence by using a paraphrasing operation of the target language including the component before paraphrasing. Machine equipped with Translation apparatus. 2. A machine translation device for translating a source language sentence into a target language sentence, analyzing the source language sentence to extract constituent elements of the source language sentence, which are words or phrases of the source language sentence, and extract the source language sentence. With reference to the source language paraphrasing operation consisting of a pair of components before and after the paraphrase of the paraphrasing operation, the degree of approximation of the component before the paraphrase of the source language paraphrasing operation with respect to the component of the source language sentence is described. Calculating a first score, selecting the source language paraphrase operation that includes the pre-paraphrase component of the source language paraphrase operation having the largest first score, and selecting the selected source language paraphrase operation , The source language paraphrase processing unit that paraphrases the source language sentence into the source language sentence after paraphrase, and the language conversion knowledge composed of a pair of the source language constituent element and the corresponding target language constituent element. , Language conversion for converting the source language sentence after the above paraphrase into the target language sentence And a constituent part of the target language sentence that is a word or a phrase of the target language sentence by analyzing the target language sentence, and based on the difference information, the components before and after the paraphrase of the paraphrase operation of the target language. A second score indicating the degree of approximation of the component before the paraphrase of the paraphrase operation of the target language with respect to the component of the target language sentence is calculated with reference to the paraphrase operation of the predetermined target language consisting of a pair of Then, the paraphrase operation of the target language including the component before paraphrase of the paraphrase operation of the target language having the maximum second score is selected, and the target language sentence is selected using the paraphrase operation of the selected target language. And a target language paraphrasing processing unit that paraphrases the target language sentence after paraphrasing, wherein the language conversion processing unit refers to the constituent elements of the source language of the language conversion knowledge, with respect to the source language sentence after paraphrasing, Source language configuration required for the above language conversion knowledge A third score indicating the degree of approximation of the prime is calculated, and when the value of the calculated third score is higher than a predetermined threshold value, the language conversion knowledge of the source language having the maximum third score is calculated. The language conversion knowledge including the constituent elements is selected, and the source language sentence after the paraphrase is language-converted into the target language sentence using the selected language conversion knowledge, while the value of the calculated third score is the above. When it is less than or equal to the threshold value, the constituent element of the source language of the language conversion knowledge having the maximum third score is selected as translatable sentence information and output to the source language paraphrasing processing unit. The part is a degree of approximation of a component before paraphrase of the paraphrase operation of the source language with respect to a component of the source language sentence, and an approximation of a component after paraphrase of the paraphrase operation of the source language with respect to the translatable sentence information. Calculate the first score, which indicates the degree and After paraphrasing the source language sentence using the paraphrasing operation of the source language including components before and after paraphrasing the paraphrasing operation of the source language having the largest first score, and using the paraphrasing operation of the selected source language A machine translation device characterized by paraphrasing the original language sentence of. 3. The degree of approximation of the source language constituent element of the language conversion knowledge with respect to the source language sentence after the paraphrase, by the language conversion processing unit referring to the source language constituent element of the language conversion knowledge. Is calculated, and when the value of the calculated third score is higher than a predetermined threshold value, the source language component of the language conversion knowledge having the maximum third score is included. Language conversion knowledge is selected, and the source language sentence after the paraphrase is language converted into the target language sentence by using the selected language conversion knowledge, while the calculated third score value is equal to or less than the threshold value. In this case, the constituent element of the source language of the language conversion knowledge having the maximum third score is selected as translatable sentence information and output to the source language paraphrasing processing unit. Paraphrase of the above source language for the constituent elements of the language sentence A first score indicating the degree of approximation of the component before the paraphrase of the work and the degree of approximation of the component after the paraphrase of the paraphrase operation of the source language with respect to the translatable sentence information is calculated, and the maximum first score is calculated. Selecting a source language paraphrase operation including components before and after the source language paraphrase operation having a score of, and using the selected source language paraphrase operation to convert the source language sentence to the source language after the paraphrase. The phrase information is paraphrased, and difference information representing a difference before and after processing of the selected source language paraphrase operation is generated.
The machine translation device described.