JP4940606B2 - Translation system, translation apparatus, translation method, and program - Google Patents

Description

  The present invention relates to a technique for translating between different languages.

  Words that humans use for everyday communication, such as Japanese and English, are called “natural languages”. Natural languages have a natural origin and have evolved with the history of mankind, and there are now a wide variety of natural languages. Natural languages originally have abstract and ambiguous properties, but various computer processes can be performed by mathematically handling sentences (natural language sentences) composed of these natural languages. By such computer processing, various applications and services relating to natural language such as machine translation, dialogue system, and search system are realized. Among these, “machine translation” is an application or service that supports communication performed by persons having different languages using computer processing.

Currently available machine translation systems include a method called “direct method” and a method called “transfer method”. In the direct method, words in the translation source language are simply replaced with words in the translation destination language based on a word dictionary prepared in advance. This is an effective method only when the grammar of the translation source language and the translation destination language are almost similar, such as translation between Japanese and Korean. On the other hand, the transfer method includes processing for replacing a syntax structure as well as replacement of a word. Therefore, according to this transfer method, it is possible to cope with translation of different languages having different grammars. Furthermore, for example, Patent Document 1 discloses a technique in which the machine translation system is improved.
JP 10-32627 A

  However, no matter how high-precision machine translation system is used, the resulting translation is unnatural when viewed from the person (native speaker) who uses the target language as the native language. There are many things. In other words, as long as translation is based on mathematical processing by a computer, there is no room for human and ambiguous elements in the translation result, so it must be unnatural.

  The present invention has been made in view of such problems, and its purpose is to produce a more natural sentence even when viewed from a person (native speaker) who uses the translated language as a native language. To realize translation.

To achieve the above object, the present invention is connected to a plurality of computers via a network, and when a search keyword is designated, the search keyword hits information stored in the plurality of computers. A search device that outputs the number of hits or hit probabilities, a translation unit that translates a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language, and a computer that has a plurality of designated search keywords For the search device that outputs the number of hits to the information stored in the information or the probability of the hit, the phrase included in the plurality of translation sentence candidates is designated as a search keyword, and the number of hits output by the search device Search means for obtaining hit probabilities;
Among the plurality of translation sentence candidates, the translation sentence candidate including a phrase having a large hit probability or hit probability is output as a translation sentence, or, among the plurality of translation sentence candidates, the hit number or hit probability is large. Output means for generating and outputting a translation sentence using translation sentence candidates including phrases, and the search means performs a syntactic and semantic analysis on the plurality of translation sentence candidates and is included in each translation sentence candidate Means for identifying the attribute of the phrase to be extracted, and means for extracting, from each of the plurality of translation sentence candidates, a basic structure sentence consisting of a phrase corresponding to a predetermined first attribute based on the result of the syntax-semantic analysis And means for designating each extracted basic structure sentence as the search keyword for the search device, and obtaining the number of hits or the hit probability obtained as a result of the search by the search device; Among words constituting a basic structure sentence having a large hit probability or hit probability obtained as a result of the search by the search device, a word corresponding to a predetermined second attribute is extracted from the basic structure sentence, and the word or phrase is extracted. Means for extracting a word / phrase to be modified from each of the plurality of translation sentence candidates, specifying a plurality of child structures comprising the extracted word / phrase and a word / phrase to be modified, and the search keyword for the search device for the specified child structure; And a means for obtaining the hit count or hit probability obtained as a result of the search by the search device, and the output means has a high hit count or hit probability among the plurality of basic structure sentences. A basic structure sentence, a child structure having a high hit probability or a hit probability among the plurality of child structures, and a word not included in the basic structure sentence among the translation sentence candidates Providing a translation system characterized in that it comprises a translation device, characterized in that to create a translation using and.

In the present invention, translation means for translating a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language, and designated search keywords are stored in a plurality of computers. For a search device that outputs the number of hits to information or the probability of hitting, a phrase included in the plurality of translated sentence candidates is specified as a search keyword, and the number of hits or hit probability output by the search device is acquired. A search means and, among the plurality of translation sentence candidates, output a translation sentence candidate including a phrase having a high hit probability or a hit probability as a translation sentence, or among the plurality of translation sentence candidates, the number of hits using a translation candidate including the word hits probability is large and an output means for outputting to create a translation, said searching means, to said plurality of translation candidates A basic structure sentence consisting of a phrase corresponding to a first attribute determined in advance based on the result of the syntactic and semantic analysis, and means for performing syntactic and semantic analysis and specifying the attribute of the phrase included in each translation sentence candidate Is extracted from each of the plurality of translation sentence candidates, each of the extracted basic structure sentences is designated as the search keyword for the search device, and the number of hits or hits obtained as a result of the search by the search device Of the words constituting the basic structure sentence having a high probability of hits or the number of hits acquired as a result of the search by the search device, the word corresponding to the predetermined second attribute is selected as the basic structure sentence. The phrase that modifies the phrase is extracted from each of the plurality of translation sentence candidates, and a plurality of child structures including the extracted phrase and the phrase that modifies the phrase are specified. It has a stage, specifying the identified said terminal structure as the search keywords for the search device, and means for obtaining the search results obtained number of hits to hit probability by the search device,
The output means includes a basic structure sentence having a high hit probability or a hit probability among the plurality of basic structure sentences, a child structure having a high hit probability or a hit probability among the plurality of child structures, and the translation sentence candidates. A translation apparatus is provided that creates a translation sentence using a phrase that is not included in the basic structure sentence .

Further, the present invention is a computer, a natural language sentence expressed in a language, a translation step of translating the plurality of translation candidates expressed in a different language, the computer has been designated search keyword of the plurality For the search device that outputs the number of hits to the information stored in the computer or the probability of the hit, the phrase included in the plurality of translated sentence candidates is specified as a search keyword, and the number of hits output by the search device Or a search step for obtaining hit probabilities , wherein the plurality of translation sentence candidates are subjected to syntax semantic analysis, attributes of phrases included in the respective translation sentence candidates are specified, and based on the result of the syntax semantic analysis A basic structure sentence consisting of a phrase corresponding to a predetermined first attribute is extracted from each of the plurality of translation sentence candidates, and each of the extracted base sentences is A structure sentence is designated as the search keyword for the search device, the number of hits or the hit probability obtained as a result of the search by the search device is obtained, and the number of hits or the hit probability obtained as a result of the search by the search device is large. A word corresponding to a predetermined second attribute is extracted from the basic structure sentence among words constituting the structure sentence, and a word that modifies the word is extracted from each of the plurality of translation sentence candidates, and extracted. A plurality of child structures consisting of the phrase and the phrase that modifies the phrase, the identified child structure as the search keyword for the search device, and the number of hits or hits obtained as a result of the search by the search device a searching step of obtaining a probability, the computer of the plurality of translation candidates, the number of hits to the phrase hit probability is greater Output step whether to output a translation candidate as a translation, or, among the plurality of translation candidates, and outputs the created a translation using a translation candidate including the number of hits to the phrase hit probability is greater comprising Among the plurality of basic structure sentences, the basic structure sentence having a large hit probability or hit probability, the child structure having the large hit number or hit probability among the plurality of child structures, and the translation sentence candidates There is provided a translation method comprising: an output step of creating a translation sentence using a phrase not included in the basic structure sentence .

Further, the present invention is the computer, the natural language sentence expressed in a language, a translation function for translating a plurality of translation candidates expressed in a different language, the designated search keyword are multiple computers For the search device that outputs the number of hits to the stored information or the hit probability, the phrase included in the plurality of translation sentence candidates is designated as a search keyword, and the number of hits or hits output by the search device A search function for acquiring a probability, performing a syntax semantic analysis on the plurality of translation sentence candidates, specifying an attribute of a phrase included in each translation sentence candidate, and based on a result of the syntax semantic analysis, A basic structure sentence composed of words corresponding to a predetermined first attribute is extracted from each of the plurality of translation sentence candidates, and each of the extracted basic structure sentences is stored in the search device. Phrases that constitute a basic structure sentence that is designated as the search keyword, obtains the hit number or hit probability obtained as a result of the search by the search device, and has a large hit probability or hit probability obtained as a result of the search by the search device Among them, a phrase corresponding to a predetermined second attribute is extracted from the basic structure sentence, a phrase that modifies the phrase is extracted from each of the plurality of translation sentence candidates, and the extracted phrase and its modification A search function for specifying a plurality of child structures composed of words and phrases, specifying the specified child structure as the search keyword for the search device, and acquiring the number of hits or the hit probability obtained as a result of the search by the search device A translation sentence candidate including a phrase having a high hit probability or a hit probability among the plurality of translation sentence candidates is output as a translation sentence, or Among the plurality of translation candidates, the number of hits to hit probability with the translation candidate including phrase large an output function of outputting to create a translation, the plurality of the basic structure statement said Using a basic structure sentence with a high hit probability or hit probability, a child structure with a high hit number or hit probability among the plurality of child structures, and a phrase that is not included in the basic structure sentence among the translation sentence candidates A program for realizing an output function for creating a translation is provided.

  According to the present invention, a phrase included in a plurality of translation sentence candidates is designated as a search keyword for the search device, and the number of hits or the hit probability as a search result is acquired. Since the search device searches various information stored in a plurality of computers using a specified search keyword, a phrase with a high hit count or a high hit probability is a phrase used by many people. . Therefore, a translation sentence candidate including a phrase with a high hit probability or a hit probability among a plurality of translation sentence candidates can be said to be a sentence having a more natural combination and order of phrases. Such a translation sentence candidate is output as a translation sentence, or if necessary, a translation sentence is created and output using such a translation sentence, so that a translation sentence with a natural phrase for a native speaker can be obtained. Can do.

Next, the best mode for carrying out the invention will be described. In the following description, “word / phrase” is literally used to mean not only “word” (word) but also a phrase (phrase).
In the present embodiment, a natural language sentence expressed in one language is translated into a plurality of translation sentence candidates expressed in another language by conventional machine translation. As a result, a search is executed by specifying a search phrase on the Internet for a phrase included in the obtained plurality of translation sentence candidates as a search keyword. The search engine compares the specified search keyword with various texts published on the WWW (World Wide Web) sites established by individuals and organizations, and outputs the number of hits and the hit probability. is there. That is, an expression with a high number of cases or a probability of hitting in this search engine is an expression used by many people. Therefore, a translation sentence candidate having a large number of hits or a hit probability among a plurality of translation sentence candidates can be said to be a general sentence with a more natural combination and order of phrases. Therefore, if a translated sentence is created using candidate translation sentences including words or phrases having a high probability or number of hits in the search engine, it is possible to obtain a translated sentence that is natural for native speakers.

  FIG. 1 is a diagram illustrating a configuration of a translation apparatus 10 according to the present embodiment. This translation apparatus 10 is, for example, a general-purpose personal computer in which a translation program describing a procedure for translation processing and a WWW browser program describing a procedure for performing communication via the Internet are installed. These translation programs and WWW browser programs are stored in the storage unit 11 such as a hard disk device. The translation program includes a translation dictionary used for machine translation. In this translation dictionary, a phrase in the translation source language and a phrase in the translation destination language are described in association with each other and used for so-called machine translation.

  The control unit 12 is a CPU, for example, and executes various programs stored in the storage unit 11. The communication unit 13 is connected to the Internet 20 via, for example, a public telephone line, an ISDN (Integrated Services Digital Network) line, or an ADSL (Asymmetric Digital Subscriber Line) line. The Internet 20 is a huge computer network that connects networks all over the world using a standardized communication protocol (TCP / IP). Many computers are connected to the Internet 20 and can communicate with each other via the Internet 20. One of the computers connected to the Internet 20 is a search engine server device 30. The search engine server device 30 is a kind of WWW server device, and is based on keywords specified from the computer via the Internet 20, and is made public on the Internet 20 (by a computer connected to the Internet 20). A search service is provided for searching stored information) and returning the search result to a computer. The search result includes the number of hits or the hit probability that the search keyword hits information stored in a plurality of computers. Typical search services include “Google”, “goo”, and “Yahoo!” (all are registered trademarks).

  The data input / output unit 14 inputs / outputs data to / from the translation device 10, and for example, via a drive device that reads / writes data from / to various storage media or a network such as a LAN (Local Area Network). A communication device that exchanges data with other devices. The display unit 15 is, for example, a liquid crystal display or a CRT (Cathode Ray Tube) display. The operation unit 16 is configured by, for example, a keyboard and a mouse.

In this embodiment, since a language analysis theory called LFG (Lexical Functional Grammar) is used, the contents thereof will be briefly described.
In this language analysis based on LFG, syntactic analysis and semantic analysis (case structure analysis) are performed to output structures called constant structure (c-structure) and functional structure (f-structure) as analysis results. The f-structure is characterized in that the syntactic and semantic information of a sentence is expressed by a nested structure of pairs of attributes and attribute values. Each word constituting the sentence is described in the f-structure as an attribute value corresponding to an attribute called PRED (predicate: predicate). In this f-structure, the part that changes depending on each language is only the attribute value (phrase) corresponding to the above PRED, and other parts are standardized (standardized) throughout all languages. In other words, sentences that represent the same meaning and content, even in different languages, will be f-structures with exactly the same structure, except for the words themselves.

This f-structure will be described with an example.
FIG. 2 is a diagram illustrating an example of an f-structure obtained by performing LFG analysis on a Japanese sentence “Taro gave a present to Hanako”, for example. In FIG. 2, the correspondence between the attribute and the attribute value is expressed by arranging them at horizontal positions. For example, the attribute “PRED” and the attribute value “pass” are in a correspondence relationship. The part underlined in the figure is a word (attribute value corresponding to the PRED attribute), and all the contents except for the word / phrase are concepts common to all languages. However, parts common to all languages are expressed in English for notation. In the figure, attribute “PRED” is a predicate, attribute “SUBJ” is the main case, attribute “OBJ” is the purpose case, attribute “OBL” is the oblique case, attribute “TENSE” is the tense, and “PAST” is the past. Represents meaning. FIG. 3 is a more detailed list of attributes. Details of the f-structure are disclosed in the document “Miriam Butt, et. al.,“ A Grammar Writer's Cookbook ”, CSLI Publication (1999)”.

Next, the operation of this embodiment will be described.
4 to 7 are flowcharts showing the execution procedure of the control unit 12 described in the translation program. The translation procedure represented by this flowchart can be roughly divided into the following contents.
Extraction and selection of basic structure sentence (FIG. 4: step S1 to step S5)
Extraction and selection of child structure of basic structure sentence (FIG. 4: step S6 to FIG. 5: step S17)
Extraction and selection of words not included in the basic structure sentence (FIG. 5: Step S23 to Step S26, FIG. 6: Step S18 to FIG. 7: Step S31)

First, the user operates the operation unit 16 of the translation device 10 to input a sentence expressed in the translation source language (hereinafter referred to as an original sentence) from the data input / output unit 14 to the translation device 10 and instruct to translate it. To do. Here, the source language is English, and the original English text is as follows. Also, the target language will be Japanese.
“Japanese children enter the first grade of elementary school in the April after their sixth birthday.”
The control unit 12 executes the translation program in response to a translation instruction from the user, and starts the process shown in FIG. First, the control unit 12 machine-translates the original text into Japanese using a translation dictionary (step S1). It is assumed that the control unit 12 obtains the following three translation results by this translation process.
(A) “Japanese children enter the first grade of elementary school after their sixth birthday in April.”
(B) “Japanese children enter the first grade elementary school in April after their 6th birthday.”
(C) “Japanese children enter the first grade of elementary school in April after their sixth birthday.”
Hereinafter, the above three sentences (a), (b), and (c) are referred to as “translation sentence candidates”.

  Next, the control unit 12 performs these translation sentence candidates. Syntactic and semantic analysis using LFG is performed (step S2). Accordingly, the control unit 12 performs f-structure (see FIG. 8) for the translation sentence candidate (a), f-structure (see FIG. 9) for the translation sentence candidate (b), and f-structure for the translation sentence candidate (c). (See FIG. 10).

  Next, the control unit 12 determines the attribute values (phrases) of the attributes “PRED”, “SUBJ”, “OBJ”, and “OBL” from the f-structures of the translation sentence candidates (a), (b), and (c). ) And a sentence composed of the extracted attribute values (words) (hereinafter referred to as a basic structure sentence) is generated (step S3). Since these attributes "PRED", "SUBJ", "OBJ", and "OBL" can be said to be the attributes of the main words in the sentence, the basic structure sentence consisting of attribute values (phrases) of these attributes is generally "case" It is called a sentence that represents the basic structure of the original text. Note that the case particle (“Ni”, “O”, etc.) is described as an attribute value in “PRED” of the attribute “OBL”, but the control unit 12 generates this attribute when generating the basic structure sentence. Associate self-supporting words with “PRED” in “OBL”.

Thereby, the control part 12 obtains the following three basic structure sentences. The translation sentence candidate (a) and the basic structure sentence (a1) correspond, the translation sentence candidate (b) and the basic structure sentence (b1) correspond, and the translation sentence candidate (c) and the basic structure sentence. (C1) corresponds.
(A1) "Child enters the first grade"
(B1) “Child enters elementary school”
(C1) “Children enter grades”

Next, the control unit 12 activates the WWW browser program to access the search engine server device 30, and uses the basic structure sentences (a1), (b1), and (c1) as search keywords in the search engine server device 30. Transmit (step S4). In response to this, the search engine server device 30 returns the search result. The search result when “google” (registered trademark) is used as the search engine server device 30 is shown below.
(A1) “Children enter the first grade” Number of hits: 0 Same phrase rate: 0/0 = 0.000
(B1) “Children enter elementary school” Number of hits: 11 Same phrase rate: 10/11 = 0.909
(C1) “Children enter grade” Number of hits: 0 Same phrase rate: 0/0 = 0.000
“Google” has a service that searches for information that does not exactly match the search keyword, such as “children enter a nearby elementary school” when the search keyword “kid enters elementary school” is specified. . “Number of hits” is the number of such hits including those that are not completely matched. On the other hand, the “same phrase rate” is one type of hit probability in the search, and is a value obtained by dividing the number of hits that completely match the search keyword by the total number of hits including the number that does not completely match.

  Next, the control unit 12 determines that the sentence is the most natural sentence based on the search result by the search engine server device 30 among the above three basic structure sentences (a1), (b1), and (c1). A possible basic structure sentence is selected (step S5). Specifically, the control unit 12 selects a basic structure sentence with the highest phrase ratio from the above three basic structure sentences (a1), (b1), and (c1). This is because the same phrase rate in the basic structure sentence means that the basic structure sentence is used by more people. However, when the same phrase ratios of a plurality of basic structure sentences have the same value, the control unit 12 selects a basic structure sentence having a large number of hits from among the basic structure sentences. This is because even if the number of hits in the basic structure sentence is large, sentences similar to the basic structure sentence are used by more people. In the example of the present embodiment, since the same phrase ratio of “children enter elementary school”, which is the basic structure sentence (b1), is the highest “0.909”, the control unit 12 selects the basic structure sentence (b1). Will do.

Next, the control unit 12 extracts a child structure based on the content of the basic structure sentence (b2) selected in step S5 (step S6). The child structure is attribute values (phrases) of the attributes “SUBJ”, “OBJ”, “OBL” that constitute the basic structure sentence, and the attribute values in the translation sentence candidates (a), (b), (c) Refers to a phrase that is composed of the word or phrase being modified. More specifically, the control unit 12 first extracts the attribute value of the attribute “SUBJ” from the selected basic structure sentence (b2). In this example, the attribute value “child” of the attribute “SUBJ” in the basic structure sentence (b1) is extracted. Then, the control unit 12 extracts a phrase that modifies the phrase “child” in the translation sentence candidates (a), (b), and (c). In other words, the phrases “Japanese” and “Japan”. Then, the control unit 12 extracts a child structure composed of the phrase “child” extracted from the basic structure sentence and the phrase “Japanese” or “Japan” that modifies it. As a result, the control unit 12 obtains the following two child structures. The child structure (a2) corresponds to the translation sentence candidate (a) and the basic structure sentence (a1), and the child structure (b2) corresponds to the translation sentence candidate (b) and the basic structure sentence (b1). is doing.
(A2) “Japanese children”
(B2) “Japanese children”

Next, the control part 12 accesses the search engine server apparatus 30 by a WWW browser program, and transmits said child structure (a2), (b2) to the search engine server apparatus 30 as a search keyword (step S7). The search results are shown below.
(A2) “Japanese children” hits: 4480 (b2) “Japanese children” hits: 18,800

  After confirming that the number of hits is not zero (step S8; No), the control unit 12 compares the number of hits with a predetermined threshold (here, 1000), and compares the magnitudes of the hits with each other. The relationship is determined (step S9). Here, if the number of hits does not exceed the threshold value (step S9; No), the control unit 12 selects the child structure with the larger number of hits (step S10). After selecting the child structure, the control unit 12 proceeds to the process of step S17 in FIG.

On the other hand, if the number of hits exceeds the threshold value (step S9: Yes), it means that the number of hits of any child structure is relatively large, so the control unit 12 has a natural expression. Cannot judge whether there is. In this case, the control unit 12 inserts the above-described child structures (a2) and (b2) into the basic structure sentence (b1) “child enters elementary school” selected in step S5 (hereinafter, detailed structure). Sentence) is generated (step S11). In this example, the detailed structure sentence is as follows. The detailed structure sentence (a3) corresponds to the child structure (a2), and the detailed structure sentence (b3) corresponds to the child structure (b2).
(A3) “Japanese children enter elementary school”
(B3) “Japanese children enter elementary school”

  And the control part 12 designates this detailed structure sentence as a search keyword, and performs the search by the search engine server apparatus 30 (step S12). Then, after confirming that the number of hits in the search result is not 0 (step S13: No), the control unit 12 selects the child structure with the larger number of hits (step S10).

On the other hand, in step S13, if the number of hits in the search result is zero (step S13; Yes), the process of the control unit 12 proceeds to step S14 in FIG.
(A3) “Japanese children enter elementary school” Number of hits: 0 (b3) “Japanese children enter elementary school” Number of hits: 0

In step S <b> 14, the control unit 12 separates the detailed structure sentence into a plurality of words by dividing the detailed structure sentence with a case particle and an auxiliary particle. And the control part 12 designates these as a search keyword by AND condition, and performs the search by the search engine server apparatus 30 (step S15). An example of this search result is shown below.
(A4) “Japanese children” & “To elementary school” & “Enter” Hits: 18
(B4) “Japanese children” & “To elementary school” & “Enter” Hits: 47
The control unit 12 selects the child structure inserted in the detailed structure sentence with the larger number of hits, here (b2) “Japanese children” (step S16).

Through the processing described above, the control unit 12 can determine the child structure constituted by the attribute value “child” of the attribute “SUBJ” in the selected basic structure sentence (b2) and the phrase that modifies it. It was. Next, the control unit 12 determines whether or not a child structure can be further extracted based on the content of the basic structure sentence (b2) selected in step S5 (step S17). In this example, in the basic structure sentence (b2), the child structure “first grade elementary school” composed of the attribute value “elementary school” of the attribute “OBL” and the phrase “first grade” that modifies it. Is present (step S17; Yes), the control unit 12 may repeat the steps S6 to S16 in FIGS. 4 to 5 to confirm whether or not the contents of the child structure are acceptable. However, in this example, the search result of step S7 is
(B4) “First grade elementary school” Number of hits: 0
Thus, it is assumed that the number of hits is zero.

Thus, when the number of hits is zero (step S8; Yes), the control unit 12 performs a syntactic and semantic analysis on the expression “first grade elementary school” and, as follows, The dependency relationship is specified (step S18 in FIG. 6).
(B5) “First” → “Grade” → “Primary school”
Then, the control unit 12 designates the phrase “grade elementary school”, which combines the attribute value “child” of the attribute “OBL” and the phrase “grade” directly modifying the attribute value, as a search keyword. A search is performed by the engine server device 30 (step S19). The result of this search is as follows.
(B5) “Grade elementary school” Number of hits: 0 In this way, when the number of hits is zero (step S20; Yes), the control unit 12 does not qualify the attribute value of the attribute “OBL” at all. Two words (phrases) of “elementary school” or “first grade elementary school” which is the child structure itself are stored as translation candidates in the storage unit 11 (step S21). On the other hand, when the number of hits is not zero (step S20; No), the control unit 12 stores the child structure “first grade elementary school” in the storage unit 11 as a translation candidate (step S22).

  Referring back to FIG. 5 again, when the process related to the extraction / selection of the child structure of the basic structure sentence is completed (Step S17; No), the control unit 12 performs the translation sentence candidates (a), (b), and (c), respectively. Words that are not included in the basic structure sentences (a1), (b1), and (c1) are extracted (step S23). Next, the control unit 12 determines whether or not there is a word that is included in all of the translated sentence candidates (a), (b), and (c) among the extracted words (step S24). ). In this example, the words “their”, “birth”, and “day” correspond respectively. When there is such a phrase (step S24; Yes), the control unit 12 extracts a phrase (phrase) obtained by connecting these phrases from each translation sentence candidate (a), (b), (c), A search by the search engine server device 30 is performed using these words (phrases) as search keywords (step S25). In this case, the phrase (phrase) “6th birthday” is extracted from the translation sentence candidate (a), and the phrase (phrase) “6th birthday” is extracted from the translation sentence candidate (b). The phrase (sixth birthday) is extracted from the translation sentence candidate (c) and searched. The search results are as follows.

(A6) “6th birthday” Number of hits: 16,500 Same phrase rate: 209 / 16,500 = 0.013
(B6) “6th birthday” Hits: 9,180 Same phrase rate: 262 / 9,180 = 0.029
(C6) “6th birthday” Number of hits: 158,000 Same phrase rate: 1 / 158,000 = 0.000
The control unit 12 selects a word / phrase having the highest phrase ratio from the search result (step S26). In this example, the phrase “6th birthday” will be selected.

Then, the process of the control unit 12 proceeds to FIG. The control part 12 is a part in which the order of the phrases differs among the phrases that are not included in the basic structure sentences (a1), (b1), and (c1) in the translation sentence candidates (a), (b), and (c). It is determined whether or not there is (step S27). In this example, in translation candidate (a), “in April” is before “after their 6th birthday”, but in translation candidates (b) and (c) The expression equivalent to “after the sixth birthday of” is after “in April”. When the word order is different as described above (step S27; Yes), the control unit 12 performs the syntactic and semantic analysis on the original sentence, and specifies the dependency relationship of these words based on the result (step S28). In this example, the result of the syntactic and semantic analysis is as shown in FIG. In other words, the phrase "in April" is modified by "after their 6th birthday", "after their 6th birthday", or "after their 6th birthday" It can be seen that it is. And the control part 12 produces the phrase (phrase) which reflected this dependency relationship in translation sentence candidate (a), (b), (c), and searches by the search engine server apparatus 30 by using this phrase as a search keyword. Is performed (step S29). The results are as follows.
(A7) “April after April” Number of hits: 0 Same phrase rate 0.000
(B8) “April after birthday” Number of hits: 52 Same phrase rate 0.000
(C8) “April after birthday” Number of hits: 30 Same phrase rate 0.100

  The control unit 12 refers to the search result and selects a word (phrase) having the highest same phrase rate (step S30). In this example, the phrase “April after birthday” is selected. Thereby, the control part 12 employ | adopts the phrase "in April after a 6th birthday".

Then, the control unit 12 obtains a final translation result by appropriately combining words other than the basic structure sentence with the contents of the basic structure sentence selected by the above processing (step S31). That is, the control unit 12
“Japanese children enter elementary school in April after their 6th birthday.”
Or "Japanese children enter elementary school (first grade) in April after their 6th birthday."
Is obtained and displayed on the display unit 15. In step S27, the control unit 12 determines whether the translation sentence candidates (a), (b), and (c) are not included in the basic structure sentences (a1), (b1), and (c1). If it is determined that there is no part in which the order of the phrases is different (step S27; No), the process may proceed to step S31 described above.

According to the above embodiment, a translation is created using words / phrases with a high probability or number of hits in the search engine, so a translation with a natural phrase for a native speaker can be obtained.
In step S10 of FIG. 4 described above, the control unit 12 does not select the child structure with the larger number of hits, but calculates the distance between the search keywords in the search result, and the distance is short. One child structure may be selected. The distance between the search keywords is the number of morphemes between the search keywords in a piece of text information including the search keywords among the information stored by the computer, and the distance increases as the number of morphemes increases. The farther away, the smaller the number of morphemes, the shorter the distance. Specifically, the control unit 12 performs morpheme analysis processing on the text information included in the search results, counts the morphemes between the search keywords, obtains an average thereof, and has a smaller average value. Select the child structure.
In the embodiment, the number of hits or the hit probability (same phrase rate) is properly used as a criterion for selecting a basic structure sentence, a child structure, and other phrases from the search results. The designer can determine which of the probabilities (or both) to use. In short, the translation apparatus 10 may select a search keyword based on the number of hits or hit probability that hits information stored in a plurality of computers.
The translation program executed by the control unit 12 can be recorded on a recording medium such as a magnetic recording medium, an optical recording medium, or a ROM that can be read by a computer and provided to the translation apparatus 10. It is also possible to download the translation apparatus 10 via a network such as the Internet 20.

It is a block diagram which shows the structure of the translation apparatus which concerns on embodiment of this invention. It is a figure which shows an example of f-structure. It is a figure which shows the detailed list | wrist of an attribute. It is a flowchart which shows the execution procedure of the control part described in the translation program It is a flowchart which shows the execution procedure of the control part described in the translation program It is a flowchart which shows the execution procedure of the control part described in the translation program It is a flowchart which shows the execution procedure of the control part described in the translation program F-structure for translation candidates. F-structure for translation candidates. F-structure for translation candidates. It is a figure which shows the result of having performed the syntax semantic analysis to the original sentence.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Translation apparatus, 11 ... Memory | storage part, 12 ... Control part, 13 ... Communication part, 14 ... Data input / output part, 15 ... Display part, 16 ... Operation part , 20... Internet, 30... Search engine server device.

Claims (10)

A search device that is connected to a plurality of computers via a network and outputs a hit probability or hit probability that the search keyword hits information stored in the plurality of computers when a search keyword is specified;
A translation means for translating a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language;
For the search device that outputs the number of hits to the information stored in a plurality of computers or the probability of hitting the specified search keyword, the phrase included in the plurality of translation sentence candidates is specified as a search keyword, Search means for obtaining the hit count or hit probability output by the search device;
Among the plurality of translation sentence candidates, the translation sentence candidate including a phrase having a large hit probability or hit probability is output as a translation sentence, or, among the plurality of translation sentence candidates, the hit number or hit probability is large. An output means for generating and outputting a translation using candidate translations including words ,
The search means includes
Means for performing syntactic and semantic analysis on the plurality of translation sentences candidates, and specifying attributes of phrases included in each translation sentence candidate;
Means for extracting, from each of the plurality of translation sentence candidates, a basic structure sentence consisting of a phrase corresponding to a predetermined first attribute based on the result of the syntax and semantic analysis;
Means for specifying each of the extracted basic structure sentences as the search keyword for the search device, and obtaining the number of hits or the hit probability obtained as a result of the search by the search device;
Among words constituting a basic structure sentence having a large hit probability or hit probability obtained as a result of the search by the search device, a word corresponding to a predetermined second attribute is extracted from the basic structure sentence, and the word or phrase is extracted. Means for extracting a phrase to be modified from each of the plurality of translation sentence candidates, and identifying a plurality of child structures composed of the extracted phrase and the phrase to be modified;
Means for obtaining the number of hits or the hit probability obtained as a result of the search by the search device by designating the identified child structure as the search keyword for the search device;
The output means includes a basic structure sentence having a high hit probability or a hit probability among the plurality of basic structure sentences, a child structure having a high hit probability or a hit probability among the plurality of child structures, and the translation sentence candidates. A translation system comprising: a translation device that creates a translated sentence using a phrase that is not included in the basic structure sentence . A translation means for translating a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language;
For the search device that outputs the number of hits to the information stored in a plurality of computers or the probability of hitting the specified search keyword, the phrase included in the plurality of translation sentence candidates is specified as a search keyword, Search means for obtaining the hit count or hit probability output by the search device;
Among the plurality of translation sentence candidates, the translation sentence candidate including a phrase having a large hit probability or hit probability is output as a translation sentence, or, among the plurality of translation sentence candidates, the hit number or hit probability is large. Output means for creating and outputting a translation using candidate translations including words ,
The search means includes
Means for performing syntactic and semantic analysis on the plurality of translation sentences candidates, and specifying attributes of phrases included in each translation sentence candidate;
Means for extracting, from each of the plurality of translation sentence candidates, a basic structure sentence consisting of a phrase corresponding to a predetermined first attribute based on the result of the syntax and semantic analysis;
Means for specifying each of the extracted basic structure sentences as the search keyword for the search device, and obtaining the number of hits or the hit probability obtained as a result of the search by the search device;
Among words constituting a basic structure sentence having a large hit probability or hit probability obtained as a result of the search by the search device, a word corresponding to a predetermined second attribute is extracted from the basic structure sentence, and the word or phrase is extracted. Means for extracting a phrase to be modified from each of the plurality of translation sentence candidates, and identifying a plurality of child structures composed of the extracted phrase and the phrase to be modified;
Means for obtaining the number of hits or the hit probability obtained as a result of the search by the search device by designating the identified child structure as the search keyword for the search device;
The output means includes a basic structure sentence having a high hit probability or a hit probability among the plurality of basic structure sentences, a child structure having a high hit probability or a hit probability among the plurality of child structures, and the translation sentence candidates. A translation apparatus that creates a translation sentence using a phrase that is not included in the basic structure sentence . Attribute "PRED" Prior Symbol first attribute in LFG functional obtained in language analysis based on (Lexical Functional Grammar) structure (f -structure), "SUBJ", Ri "OBJ" and "OBL" Der, the first The translation apparatus according to claim 2, wherein the two attributes are attributes "SUBJ", "OBJ", and "OBL" in a functional structure obtained by language analysis based on LFG . Before Symbol retrieval means,
When the number of hits or the hit probability acquired by specifying the child structure as the search keyword for the search device exceeds a threshold, each of the plurality of child structures is a basic structure having a large hit number or hit probability. It means for generating a detailed structure statement is inserted into the sentence,
Specifying the generated detailed structure sentence as the search keyword for the search device, and obtaining the number of hits or the hit probability obtained as a result of the search by the search device,
The output means includes, among the plurality of detailed structure sentences, a child structure inserted in the detailed structure sentence having a high hit probability or hit probability, a basic structure sentence having a high hit number or hit probability, and the translated sentence The translation apparatus according to claim 2, wherein a translation sentence is created using a phrase that is not included in the basic structure sentence among candidates . Before Symbol retrieval means,
When the number of hits or the hit probability obtained by designating the plurality of detailed structure sentences as the search keyword for the search device are all zero, the plurality of detailed structure sentences are separated by a particle, respectively. Means for separating words,
A unit for obtaining the number of hits or the hit probability obtained as a result of the search by the search device by designating the plurality of phrases as the search keyword for the search device under an AND condition;
The output means includes a child structure inserted in a detailed structure sentence including a phrase having a high hit probability or a hit probability among the plurality of detailed structure sentences, a basic structure sentence having a high hit probability or a hit probability, The translation apparatus according to claim 4, wherein a translation sentence is created using a phrase that is not included in the basic structure sentence among translation sentence candidates . The search means replaces the search device with the specified child structure as the search keyword for the search device, and obtains the number of hits or the hit probability obtained as a result of the search by the search device. As a search result by, comprising means for acquiring text information including the search keyword,
Before Symbol output means,
Among the plurality of basic structure sentences, the basic structure sentence having a large hit probability or hit probability, the child structure having the large number of hit cases or hit probability among the plurality of child structures, and the basic structure sentence among the candidate translation sentences Instead of creating translations using words not included in the
The obtained sentence information is subjected to morphological analysis to count the morphemes between the respective search keywords, the average of the counted values is obtained, and the child structure corresponding to the smallest average value, the number of hits and the hit probability The translation apparatus according to claim 2 , wherein a translation sentence is created by using a basic structure sentence having a large and a phrase that is not included in the basic structure sentence among the translation sentence candidates . Before Symbol retrieval means,
Means for extracting a phrase that is not included in the basic structure sentence and that is included in all of the plurality of translation sentence candidates, and generating a plurality of phrases by connecting the phrases extracted from each of the translation sentence candidates When,
A plurality of generated phrases are specified as the search keyword for the search device, and the number of hits or the hit probability obtained as a result of the search by the search device is obtained,
The output means translates the phrase having a high hit probability or hit probability among the plurality of phrases, a basic structure sentence having a high hit probability or hit probability, and a child structure having a high hit probability or hit probability. The translation device according to claim 2 , wherein a sentence is created. Before Symbol retrieval means,
Means for determining whether or not there is a part in which the order of words / phrases not included in the basic structure sentence is different in the plurality of translation sentence candidates;
When it is determined that there is a part in which the order of words is different, syntactic / semantic analysis is performed on the natural language sentence expressed in the certain language to identify dependency relations of words with different orders, and the dependency Means for creating a plurality of phrases reflecting the relationship; means for designating the plurality of created phrases as the search keyword for the search device; and acquiring the number of hits or the hit probability obtained as a result of the search by the search device; Have
The output means translates the phrase having a high hit probability or hit probability among the plurality of phrases, a basic structure sentence having a high hit probability or hit probability, and a child structure having a high hit probability or hit probability. The translation device according to claim 2 , wherein a sentence is created. A translation step in which a computer translates a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language;
A computer specifies a word or phrase included in a plurality of translation sentence candidates as a search keyword for a search device that outputs the number of hits to the information stored in a plurality of computers or the probability of the hit. A search step for obtaining the number of hits or the hit probability output by the search device , wherein the plurality of translation sentence candidates are subjected to syntactic and semantic analysis, and the attribute of the phrase included in each translation sentence candidate is determined. Based on the result of the syntactic and semantic analysis, a basic structure sentence consisting of words corresponding to a predetermined first attribute is extracted from each of the plurality of translation sentence candidates, and each of the extracted basic sentences A structure sentence is designated as the search keyword for the search device, and the number of hits or hit probability obtained as a result of the search by the search device is acquired. Of the words constituting the basic structure sentence having a high hit probability or the hit probability obtained as a result of the search by the search device, the word corresponding to the predetermined second attribute is extracted from the basic structure sentence, and the phrase Is extracted from each of the plurality of translation sentence candidates, a plurality of child structures composed of the extracted words and phrases and the phrases to be modified are specified, and the specified child structure is used as the search keyword for the search device A search step for specifying and obtaining the hit number or hit probability obtained as a result of the search by the search device ;
A computer outputs a translation sentence candidate including a phrase having a high hit probability or a hit probability among the plurality of translation sentence candidates as a translation sentence, or the hit number or hit among the plurality of translation sentence candidates An output step of creating and outputting a translation sentence using a translation sentence candidate including a word with a high probability , wherein the basic structure sentence having a high hit probability or the hit probability among the plurality of basic structure sentences; An output step of creating a translated sentence using a child structure having a high hit probability or a hit probability among the child structures and a phrase that is not included in the basic structure sentence among the candidate translation sentences. Translation method. On your computer,
A translation function for translating a natural language sentence expressed in one language into a plurality of translation sentence candidates expressed in another language;
For the search device that outputs the number of hits to the information stored in a plurality of computers or the probability of hitting the specified search keyword, the phrase included in the plurality of translation sentence candidates is specified as a search keyword, A search function for obtaining the number of hits or hit probability output by the search device, performing a syntactic and semantic analysis on the plurality of translation sentence candidates, and identifying an attribute of a phrase included in each translation sentence candidate, Based on the result of the syntactic and semantic analysis, a basic structure sentence composed of words corresponding to a predetermined first attribute is extracted from each of the plurality of translation sentence candidates, and each of the extracted basic structure sentences is The search keyword is specified as the search keyword for the search device, the number of hits or the hit probability obtained as a result of the search by the search device is acquired, and the search device Of words constituting a basic structure sentence having a large hit probability or hit probability obtained as a result of search, a word corresponding to a predetermined second attribute is extracted from the basic structure sentence, and a word or phrase that modifies the word is obtained. Extracting from each of the plurality of translated sentence candidates, specifying a plurality of child structures consisting of the extracted words and phrases that modify them, specifying the specified child structures as the search keywords for the search device, A search function for acquiring the number of hits or the hit probability obtained as a result of the search by the search device ;
Among the plurality of translation sentence candidates, the translation sentence candidate including a phrase having a large hit probability or hit probability is output as a translation sentence, or, among the plurality of translation sentence candidates, the hit number or hit probability is large. An output function for creating and outputting a translation sentence using translation sentence candidates including a phrase , wherein the basic structure sentence having a large hit probability or hit probability among the plurality of basic structure sentences, and the plurality of child structures A program for realizing a child structure having a large hit probability or a hit probability, and an output function for creating a translated sentence using a phrase that is not included in the basic structure sentence among the translated sentence candidates .

Images