JP2008176536A - Device, method and program for mechanically translating input original language sentence to target language - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a mechanical translation device improved in translation accuracy of voice translation. <P>SOLUTION: The device comprises: an example storage part 123 storing an example of a first language in association with an example of a second language; an example translation part 104 acquiring an example of the second language corresponding to an input first language sentence from the example storage part 123; a rule translation part 103 translating the first language sentence to the second language based on a predetermined rule; a selection part 102 selecting the most probable translation result from the translation result of the example translation part 104 and the translation result of the rule translation part 103; a cache memory 122 storing a correspondence between a first word contained in the first language sentence and a second word contained in the selected translation result; and an output part 105 outputting the selected translation result. The rule translation part 103 further acquires a second word corresponding to the first word from the cache memory 122 and substitutes a second word obtained by converting the first word by the acquired second word. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an apparatus, a method, and a program for machine-translating an input source language sentence into a target language for translation by a plurality of translation methods.

  Conventionally, the application area of machine translation is limited to technical documents with limited sentence patterns. However, with recent advances in machine translation technology, it has become possible to handle documents in a wide range of fields such as Web documents. Furthermore, by integrating with speech recognition technology, a speech translator that translates human speech in real world and real time ( (Interpreters) are expected to be put into practical use. However, with regard to the practical use of speech translators, due to the nature of handling spoken language, words such as omission of words, omission of particles, inversion of phrases and sentences, rephrasing, insertion of unnecessary words, etc. have not so much appeared so far. It is necessary to cope with language phenomena, and there are many problems to be overcome.

  A useful solution to these problems is an example-driven translation method that improves the translation quality by collating the input sentence with a prepared bilingual example database and outputting the translation corresponding to the whole or part of the collation. (EBMT: Example Based Machine Translation) has been proposed. However, the cost of developing a huge parallel translation example database that can be collated for any input sentence is enormous, and EBMT alone cannot translate any input.

  For this reason, using the rule-based translation method (RBMT: Ruled Based Machine Translation) that translates in accordance with the translation rules, which has been used in the past, the coverage is enhanced and the advantages of both methods are further utilized for the final translation. A hybrid translation system (Hybrid MT) that improves performance has been proposed.

  In Hybrid MT, consistent tone in a series of conversations, such as when the translation of the same word of the same speaker is different due to the difference in the translation method of EBMT and RBMT, or when the difference in politeness appears in the treatment expression, etc. May not be available for translation.

  As an example, (1) A Japanese sentence (I want to rent a board and shoes) that indicates that I want to rent a ski board and ski boots is entered, and EBMT “I'd like to rent skis and "boots." is output, and (2) a Japanese sentence indicating that you want to see another ski shoe (please show another shoe) is input, and "Please show me different shoes." The case where is output will be described.

  In this example, for the input (1), since the example sentence is registered in the parallel translation example database, the translation is performed by EBMT. As a result, “boots” that is optimal for the utterance context is selected as the translation of “shoes”. On the other hand, in the continuously spoken input (2), the similar example sentence does not exist in the parallel translation example database, and as a result, the translation is performed by the RBMT.

  In RBMT, a translation that is usually considered optimal in many contexts is selected. As a result, “shoes” is selected as the translation of “shoes”. As described above, there may be a case where the translated words for the same word are different due to the difference in the translation method used internally, despite the series of conversations. In the first sentence, the listener who heard the output thought of ski boots, but in the second sentence it could be paraphrased as “shoes”, so he felt embarrassed to draw the intention of the speaker. Become.

  As a technique related to the above problem, in Patent Document 1, a professional translator or the like uses a machine translation device (RBMT) and a translation memory device (EBMT) in an interactive manner when translating a manual sentence or the like. Systems that support the process of creating translations have been proposed.

  In Patent Document 1, the RBMT result is manually corrected and added to an EBMT example sentence by a mechanism that is a data compatibility processing unit, or a bilingual pattern or a bilingual dictionary obtained by analyzing the EBMT example sentence is an RBMT dictionary. To add to. As a result, the translation result of RBMT and EBMT can be handled with the same data structure.

  As a result, if many translations of “shoes” appear as “boots” in the translation memory, the translation “boots” is added to the machine translation dictionary. As a result, RBMT can also obtain “boots” as a translation of “shoes”.

JP 2002-278964 A

  However, the method disclosed in Patent Document 1 has a problem in that the parallel translation information becomes non-volatile because the translation information potentially possessed by the EBMT is directly registered on the RBMT side. That is, there is a tendency that the translation of a certain input word is fixed to a specific one regardless of the input context. This is suitable for the purpose of improving translation efficiency by effectively using a past translation result of a professional translator or the like in a translation support environment assumed in Patent Document 1.

  However, the speech translation device used in the conversation support environment is assumed to be used in various scenes, and the optimal translation differs from scene to scene. Arise.

  For example, in a conversation sentence at a rental ski store, “shoes” should be translated as “boots”, but if used in a shoe shop, it should normally be translated as “shoes”. However, when “boots” is registered as a translated word of “shoes” by the method of Patent Document 1, “boots” is also translated into a conversation sentence in a shoe shop, and translation accuracy decreases.

  The present invention has been made in view of the above, and an object of the present invention is to provide an apparatus, a method, and a program that can improve translation accuracy in speech translation in which many contexts are assumed.

  In order to solve the above-described problems and achieve the object, the present invention includes an example storage unit that stores an example of the first language and an example of the second language in which the example of the first language is translated in association with each other. A dictionary storage unit that stores a word dictionary in which a first word in a first language and a second word in a second language are associated with each other, an input receiving unit that receives an input of a first language sentence in a first language, An example translation unit for example-translating the first language sentence into a second language by acquiring an example of the second language corresponding to the first language sentence from the example storage unit, and included in the first language sentence The second word corresponding to the first word is acquired from the dictionary storage unit, and the first word included in the first language sentence is acquired as the acquired second word based on a predetermined rule. Translating the first language sentence into a second language by conversion A rule translation unit; a selection unit that selects a most likely translation result among a translation result of the example translation unit and a translation result of the rule translation unit; and the first word included in the first language sentence; A correspondence storage unit that stores a correspondence relationship between the second word included in the selected translation result, and an output unit that outputs the selected translation result, and the rule translation unit further includes: The first language sentence of the first rule is translated into a second language, and for each of the first words included in the other first language sentence, the second word corresponding to the first word is stored in the correspondence storage unit. And the second word obtained by regular translation of the first word is replaced with the acquired second word.

  Further, the present invention is a method and program capable of executing the above-described apparatus.

  According to the present invention, even if different translation methods are used internally, the consistency of the translation result can be obtained, so that the translation accuracy can be improved, and an unnatural impression is not given to the listener. Play.

  Exemplary embodiments of an apparatus, a method, and a program according to the present invention will be described below in detail with reference to the accompanying drawings.

  The machine translation apparatus according to the present embodiment stores additional information obtained from the translation results of rule translation and example translation constituting the hybrid translation system in a cache memory, and refers to the additional information in the subsequent translation processing. It improves the accuracy of translation.

  In the following, description will be made on the premise of machine translation between Japanese and English, but the combination of the source language and the target language of translation is not limited to this, and any combination of languages can be applied.

  FIG. 1 is a block diagram showing a configuration of a machine translation apparatus 100 according to the present embodiment. As shown in the figure, the machine translation apparatus 100 includes a dictionary storage unit 121, a cache memory 122, an example storage unit 123, an input reception unit 101, a translation control unit 102, a rule translation unit 103, and an example translation. Unit 104 and output unit 105.

  The dictionary storage unit 121 stores a translation dictionary in which words in the source language are associated with words in the target language. The dictionary storage unit 121 is referred to when the rule translation unit 103 to be described later translates. FIG. 2 is an explanatory diagram showing an example of the data structure of the translation dictionary stored in the dictionary storage unit 121. In the example of the figure, the translation dictionary stores Japanese words, Japanese word politeness, English words, English word politeness, and screens in association with each other. .

  The politeness indicates the degree of politeness of the meaning content of each word. For example, as shown in the figure, there are two Japanese translations for the English word “visit” (“go” and “listen”), but each has a different level of politeness. The degree values (“3”, “4”) are different. The politeness is referred to select a translated word that matches the politeness in the cache memory 122 described later.

  The scene is information representing a scene where the translation is used, and is used to select an appropriate translation by referring to the scene when there are a plurality of corresponding translations.

  The cache memory 122 includes additional information relating to translation such as correspondence between words between a source sentence that is a sentence in the input source language and a translation result (translation) in the target language of the source sentence, scene information about a scene, politeness, etc. Are temporarily stored, and are configured by a RAM (Random Access Memory). Note that the cache memory 122 may be configured by other commonly used storage media such as an HDD (Hard Disk Drive), an optical disk, and a memory card. Further, additional information and the like are added to the cache memory 122 in chronological order, and the stored information is initialized when the translation processing related to a series of dialogs is completed.

  FIG. 3 is an explanatory diagram illustrating an example of information stored in the cache memory 122. The figure shows the processing result when a Japanese sentence (I want to rent a board and shoes) is input, as shown in the above example. The registered state is shown.

  As shown in the figure, identification information and additional information are stored in the cache memory 122. The identification information includes an ID for uniquely identifying the input sentence, a translation direction, and a translation method.

  The ID is expressed in a format to which “speaker identification information” such as a user ID and a user name, a symbol “#”, and a “order number” indicating the utterance order for each speaker are added. In the figure, for example, “S # 1” and “S # 2” that are IDs corresponding to the first and second utterances of the Japanese speaker “S” are registered.

  In this embodiment, “JE” is set for Japanese-English translation and “EJ” is set for English-Japanese translation. In the method, the method of translation that generated the translation result selected by the selection unit 102a described later is set, and “EB” is used for the example-driven translation method, and “RB” is used for the rule-driven translation method. Is set.

  As shown in the first record of the figure, for the above example in which “I'd like to rent skis and boots.” Is output by EBMT, ID = “S # 1” and the translation direction is Japanese-English. (JE), the example translation type (EB) is set as the translation method used.

  The additional information is information that the rule translation unit 103 and the example translation unit 104 commonly output together with the translation result, and includes scene information, alignment information, politeness, and certainty.

  The scene information is information that is output when the example-driven translation method is used for translation, and is information that indicates a scene in which the example searched during example translation is used. For the above example, “ski” is set as the scene information. As will be described later, the scene information is stored in the example storage unit 123 together with each example.

  Note that when no scene information is stored in the example storage unit 123, an empty value (null) is set. The rule-driven translation method may also be configured to specify a use scene of an utterance input from a scene of a word included in a sentence and output scene information.

  The alignment information is information indicating the correspondence between the constituent elements of the original sentence and the translated sentence. The component represents a morpheme such as a word or a morpheme string including a plurality of morphemes. In the figure, it is shown that the constituent elements on both sides of the arrow are associated by designating the Japanese constituent element to the left of the arrow and the English constituent element to the right of the arrow.

  For example, for the above-described example, it is shown that a morpheme (board) that means a ski in the original corresponds to a morpheme “ski” in the translation. In addition, it is shown that the morpheme string in the original sentence (Ta + +) is corresponding to the morpheme string “wool + like + to” in the translated sentence.

  Moreover, you may comprise so that the attribute information of a translation word may be added to alignment information. For example, in the figure, an example is shown in which attribute information indicating that the translated word (ski) is plural (pl) is specified in parentheses, such as “ski (pl)”.

  The politeness indicates the degree of politeness of the meaning content of the translation. The politeness in the cache memory 122 is set by acquiring the politeness corresponding to the example from the example storage unit 123 when the translation result of the example translation is selected.

  The certainty factor is a numerical value of 0 to 1 representing the certainty of the translation result. The larger the value, the more reliable the translation result. Details of the certainty factor calculation method will be described later.

  The example storage unit 123 stores an example of a source language sentence and an example of a target language sentence in association with each other. FIG. 4 is an explanatory diagram showing an example of the data structure of the example stored in the example storage unit 123. As shown in FIG.

  As shown in the figure, the example storage unit 123 includes a Japanese example, an English example, scene information, and alignment information that is information indicating associations between components in the Japanese example and the English example. The politeness is stored in association with each other.

  As the alignment information, information obtained by analyzing the association of the components in advance when the example is registered in the example storage unit 123 is set. The alignment information may not be stored in the example storage unit 123 but may be configured to generate alignment information when the example translation unit 104 translates. However, from the viewpoint of the accuracy of the alignment information generation process, it is desirable that the information subjected to manual verification is given in advance.

  As described above, the politeness indicates the degree of politeness of the meaning content of the translation. The figure shows an example in which seven translations with different politeness are set for Japanese meaning "Please do it". The politeness is referred to select an example or translation so that each translation has an approximately equivalent value. This makes it possible to maintain the consistency of tone.

  The dictionary storage unit 121 and the example storage unit 123 can be configured by any commonly used storage medium such as an HDD (Hard Disk Drive), an optical disk, a memory card, and a RAM.

  The input receiving unit 101 receives input of the original text by voice recognition, handwritten character recognition, key input, or the like.

  The translation control unit 102 drives the rule translation unit 103 and the example translation unit 104, and controls the process of selecting the final translation result from the translation results output by both, and includes a selection unit 102a. Yes.

  The selection unit 102a selects, as a final translation result, a translation result having a high degree of certainty output together with the translation result from the translation results output by the rule translation unit 103 and the example translation unit 104, respectively.

  The translation control unit 102 generates additional information and corresponding identification information for the selected translation result, and registers the generated additional information and identification information in the cache memory 122. A method of generating additional information and identification information will be described later.

  The rule translation unit 103 is a rule-driven translation method translation unit that translates an original text into a target language with reference to dictionary information stored in the dictionary storage unit 121 and a predetermined translation rule. Specifically, the rule translation unit 103 first performs normal sentence analysis processing such as morphological analysis, syntactic analysis, and dependency structure analysis on the input original sentence, and generates a dependency structure tree reflecting the semantic structure of the original sentence. Generate. Then, the rule translation unit 103 performs a transfer process for converting the dependency structure tree into the target language dependency structure tree while referring to the dictionary storage unit 121 and a predetermined translation rule. Finally, the rule translation unit 103 generates a translation in the target language from the dependency structure tree of the target language.

  Further, the rule translation unit 103 generates alignment information between each component of the original sentence and the translated sentence from the correspondence between the elements of each dependency structure tree before and after the conversion.

  The example translation unit 104 is an example-driven translation unit that performs translation by comparing an original sentence with an example in the example storage unit 123 and selecting an example of a target language corresponding to a matching or similar example.

  Note that the rule translation unit 103 and the example translation unit 104 output the translation result, and also output the above-described additional information including the certainty of the translation result and the correspondence (alignment information) between each component of the original sentence and the translated sentence. To do.

  The output unit 105 synthesizes the translation result selected by the translation control unit 102 and outputs it to the user. Note that the output unit 105 may be configured to present the translation result to the user via a display or the like.

  Next, a machine translation process performed by the machine translation apparatus 100 according to the present embodiment configured as described above will be described. FIG. 5 is a flowchart showing the overall flow of the machine translation process in the present embodiment.

  First, the input receiving unit 101 receives an input of the original text (step S501). Next, rule translation processing (step S502) by the rule translation unit 103 and example translation processing (step S503) by the example translation unit 104 are performed. Note that the processing order of the rule translation process and the example translation process is not limited to this, and the example translation process may be executed first or may be configured to process in parallel. Details of the rule translation process and the example translation process will be described later.

  Next, the selection unit 102a compares the certainty factors of the translation result by the rule translation process and the translation result by the example translation process, and selects a translation result having a high certainty factor (step S504).

  Next, the translation control unit 102 generates additional information and identification information of the selected translation result, adds the identification information to the generated additional information, and registers it in the cache memory 122 (step S505). Specifically, the translation control unit 102 generates additional information and identification information as follows.

  First, the translation control unit 102 generates an ID of the identification information from a user ID specified or input by the user and an utterance order for each user managed by a storage unit (not shown) such as a RAM. Also, the translation control unit 102 sets the translation direction specified by the user as the direction of the identification information. Alternatively, in the case of a speech translation system, a user ID that is automatically generated using a speaker recognition technique for identifying a speaker or a translation direction obtained by a language identification technique for determining the language type of a utterance language may be set. Good. The translation control unit 102 sets the translation method selected by the translation control unit 102 itself as the identification information method.

  Of the additional information, the scene information acquired from the example storage unit 123 is set as the scene information when the example-driven type is selected as the method. As the alignment information, the alignment information acquired from the example storage unit 123 is set when the example-driven type is selected, and the alignment information generated by the rule translation unit 103 is set when the rule-driven type is selected.

  As a rule, the politeness obtained from the example storage unit 123 is set when the example-driven type is selected. In addition, referring to the politeness of the cache memory 122 and selecting a translated word having the same politeness during rule translation, the referenced politeness may be inherited and set. As the certainty factor, the certainty factor output by the rule translation unit 103 and the example translation unit 104 is set.

  Finally, the output unit 105 synthesizes the translation result selected by the translation control unit 102 and outputs it to a speaker or the like (not shown) (step S506), and the machine translation process ends.

  Next, details of the rule translation process in step S502 will be described. FIG. 6 is a flowchart showing the overall flow of rule translation processing in the present embodiment.

  First, the rule translation unit 103 analyzes the input original text and generates a dependency structure tree (step S601). As described above, the rule translation unit 103 generates a dependency structure tree by executing sentence analysis processes such as morphological analysis, syntactic analysis, and dependency structure analysis that have been conventionally used.

  Next, the rule translation unit 103 transfers the generated dependency structure tree to the target language dependency structure tree (step S602). This process is the same as the process in the machine translation transfer system used conventionally.

  Thereafter, the rule translation unit 103 executes a cache application process for changing the words in the dependency structure tree by applying the contents of the cache memory 122 (steps S603 to S607). In the rule translation cache application process, records in the cache memory 122 that are basically arranged in chronological order are retrospectively searched, and when applicable additional information is searched, using the searched additional information, Modify the dependency tree. The scope and application method of retrospective search differ depending on the type of additional information. Hereinafter, the cache application process in steps S603 to S607 will be described.

  First, the rule translation unit 103 acquires alignment information from the cache memory 122 (step S603), and determines whether or not the acquired alignment information includes a word in the dependency structure tree (step S604).

  If it is included (step S604: YES), the rule translation unit 103 changes the word in the dependency structure tree with the word in the alignment information (step S605). Specifically, alignment information is applied as follows.

  First, the rule translation unit 103 searches the acquired alignment information for a range in which the speaker is the same, or the translation direction is the reverse direction and the translation method is the example initiative type (EB). The search key is a morpheme or morpheme string in the source language when the speakers are the same. When the translation direction is the reverse direction, the translated word side morpheme or morpheme string is used as a search key.

  For example, information as shown in ID = “S # 1” and “H # 1” in FIG. 3 is stored in the cache memory 122, and the speaker (S) is in a different skiing manner as in the above example. Assume that a Japanese sentence indicating that you want to see shoes (please show me another shoe) is entered (ID = “S # 2”). In this case, since there is no record in which the translation direction is the reverse direction and the translation method is EB, the search range is only the record of ID = “S # 1” with the same speaker.

  For the original text corresponding to ID = “S # 2”, alignment information (shoes → boot (pl)) which is one of the words in the original text and includes Japanese (shoes) meaning ski boots. It exists in the cache memory 122. Therefore, for example, when “shoe (pl)” is selected as the translation of the word (shoes) in the rule translation, the translation is changed to “boot (pl)” according to the alignment information.

  At this time, the translation attribute information (such as pl) is also inherited. For example, the Japanese word meaning “I lost one shoe” was input in the example-driven type, the translated word “I've lost a boot.” Was output by the example translation, and then the rule-driven type “Shoes” When translating Japanese meaning, the singular "boot (sg)" is selected instead of the normal translation "boot (pl)".

  If it is determined in step S604 that the alignment information does not include a word in the dependency structure tree (step S604: NO), or after the alignment information is applied (step S605), the rule translation unit 103 Apply politeness. That is, the rule translation unit 103 acquires the scene information and the politeness from the cache memory 122 (Step S606), and changes the word in the dependency structure tree to a word that matches the scene information and the politeness (Step S607). Scene information and politeness are applied as follows.

  First, application of scene information will be described. The scene information is searched in a range where the translation method is an example-driven type (EB), and the first searched scene information, that is, the latest scene information is assumed to be the scene of the current utterance. If there is a translation that matches the scene information in the translation candidate, it is changed to the first priority translation (first translation).

  For example, the normal first translation for Japanese meaning “clothes” is “dress”, but “ski wear” is changed to the first translation in the usage scene “ski”. Therefore, if the information as shown in FIG. 3 is stored in the cache memory 122 and a word meaning “clothes” appears in the input Japanese text after that, the translation of the word Will be changed to “ski wear”.

  Next, application of politeness will be described with reference to FIG. FIG. 7 is an explanatory diagram showing another example of information stored in the cache memory 122.

  In the politeness search, the search is performed in a range where the speakers are the same. This is because it is meaningless to adjust the politeness to the conversation partner. Here, for example, “My name is Brown.” Is input as the preceding utterance, and the corresponding Japanese (I am Brown) is output by example translation, and the information as shown in FIG. Is stored.

  In this state, when the same speaker speaks “I'll visit tomorrow.”, The search range is ID = “S # 1”, and the politeness 4 is acquired. A politeness level of 4 represents the use of a humble word in English-Japanese translation.

  Here, for example, referring to the dictionary storage unit 121 as shown in FIG. 2, if Japanese (going) with politeness 3 is selected as the translated word of “visit”, the Japanese is the acquired politeness. Will be changed to another Japanese (asking) corresponding to the politeness (4) that matches.

  Returning to FIG. 6, when the above-described cache application processing (step S603 to step S607) ends, the rule translation unit 103 generates a translation from the dependency structure tree of the target language (step S608). That is, the rule translation unit 103 executes a syntax generation / morpheme generation process to generate a target language translation in the same manner as in a conventionally used machine translation transfer system.

  Next, the rule translation unit 103 generates alignment information from the correspondence between the generated translated text and the original text (step S609). Specifically, the rule translation unit 103 reflects the new correspondence obtained by the cache application process in the correspondence between the dependency structure tree of the source language and the dependency structure tree of the target language converted at the time of transfer. Is generated as alignment information.

  Next, the rule translation unit 103 calculates a certainty factor of the translation result obtained by the rule translation (step S610). The certainty factor is calculated by setting a specific value smaller than 1 as the maximum value (maximum value 1 in the case of example-driven translation) and subtracting the number of points according to the degree of ambiguity of the dependency and the ambiguity of the translation word. If the ambiguity of the translation is reduced by applying the cache, the penalty will be reduced accordingly.

Specifically, rule translation unit 103 calculates the certainty factor C _r in rule-driven translation by the following equation (1). Here, α and β are constants satisfying 0 <α and β <1, N is the total number of constituent words, N _s is the number of syntactic ambiguity, and N _tw is the number of words having ambiguity in the translated word. .

  Α is a constant that defines an upper limit value in rule-driven translation, and β is a constant that defines an upper limit value of a penalty related to the ambiguity of the translated word.

  Next, the rule translation unit 103 outputs a translation and additional information (step S611). Of the additional information, the alignment information is output from the alignment information generated in step S609. Of the additional information, the scene information and the politeness are not output by rule translation in principle. However, when the cache memory 122 is referred to and the translated information matches the scene information or the politeness, the same scene as the record of the reference source Information or politeness may be set. Further, a scene and politeness corresponding to a word included in the original sentence may be acquired from the dictionary storage unit 121, and the acquired scene and politeness may be set.

  Next, details of the example translation process in step S503 will be described. FIG. 8 is a flowchart showing the overall flow of the example translation process in the present embodiment.

  First, the example translation unit 104 searches the example storage unit 123 for an example of a target language corresponding to a sentence similar to the original sentence (step S801). As described above, the example search is performed by the similar sentence search method, and similar examples are obtained in order with a certain degree of certainty even if they do not completely match. In the case of a perfect match, 1 is set as the certainty factor.

  Although not shown in the figure, if the search fails or if only a certain degree of certainty is obtained, a blank translation result and blank additional information are output and the process ends.

  When the search is successful, a cache application process for changing the example is executed with reference to the cache memory 122 (steps S802 to S803). In the example application cache application process, basically no change process other than the politeness is performed. This is because the bilingual example is the result of the optimal interpretation already taking into account the usage situation. On the other hand, the politeness is dynamically determined based on the relationship between the speaker and the listener, and needs to be acquired from the context. Although there is a strategy that always outputs with maximum politeness, there may be a problem that it may be rude.

  In many cases, the politeness can also be obtained from the original text that is currently being translated. However, depending on the target language pair, the politeness does not always correspond uniquely. For example, if there is no preceding context for a Japanese text meaning "Are you sure you want to ask?", "May I ask a question?" And "I wonder if I might" Two English translations of "ask a question."

  Actually, there are many cases where a plurality of translated sentences differing only in politeness for the same original sentence are registered in the example storage unit 123, particularly as in an idiomatic expression. In such a case, by referring to the preceding context by the cache application process, that is, referring to the cache memory 122, one of a plurality of translated sentences having the same politeness is selected as a translation result. .

  Next, the example translation unit 104 detects a difference between the selected example and the original text, and performs word replacement (step S803). For example, if the original text is "Please give me a glass" and the searched example is "Please give me two dishes", the alignment information "Plate → dish (pl)" and "2 “Sheet → two” is changed to “cup → glass (sg)” and “one → a”, respectively, and the corresponding part of the translation is changed.

  Next, the example translation unit 104 calculates the final certainty of the translation result (step S805). The certainty factor is determined by setting the value obtained during the example search as a maximum value and subtracting the number of points according to the number of word replacements and the semantic distance between replacement words.

Specifically, the example-based translation unit 104 calculates the certainty factor C _e in the example driven translated by the following equation (2). Here, S _i represents the similarity between the input word and the example word in the correspondence between the word in the input sentence (input word) and the word in the example (example word), and 0 <= S _i ≦ = 1 Shall be satisfied.

  Note that the similarity between the input word and the example word is determined by referring to a thesaurus (not shown) that stores the hierarchical relationship between words, and a method of calculating the similarity higher as the number of hierarchies between the two is smaller. Any similarity calculation method used can be applied.

  Next, the example translation part 104 outputs a translation and additional information (step S806). The scene information, the alignment information, and the politeness included in the additional information are obtained by outputting information corresponding to the selected example from the example storage unit 123, respectively.

  Next, a specific example of machine translation processing according to this embodiment will be described. In the following, a case where information as shown in FIG. 3 is stored in the cache memory 122 and Japanese meaning “Please show me another shoe” is input will be described as an example.

  The rule translation unit 103 and the example translation unit 104 perform rule translation processing (step S502) and example translation processing (step S503), respectively, on the original text. FIG. 9 is an explanatory diagram showing an example of the dependency structure tree generated by the rule translation process at this time. FIG. 10 is an explanatory diagram showing an example of the correspondence between the example selected by the example translation processing at this time and the input sentence.

At the time when the rule translation unit 103 generates the dependency structure tree, among all the constituent words (the number of constituent words N = 3), there are a plurality of translation word candidates for all the constituent words (N _tw = 3). On the other hand, there is no syntactic ambiguity (N _s = 1). Therefore, when α = 0.9 and β = 0.3, and the certainty factor C _r at this time is calculated by the above-described equation (1), C _r = 0.9 * (1−0.3 * 3/3) ) = 0.63.

Thereafter, when the alignment information is applied in step S605, the ambiguity of the translated word of the node 901 is resolved and N _tw = 2, so the certainty factor is C _r = 0.9 * (1-0.3 *). 2/3) = 0.72. In step S610, this value is calculated as the certainty factor of the translation result of the rule translation. Thus, according to the present embodiment, the accuracy of rule translation can be improved by referring to the cache memory 122.

  On the other hand, in the example translation process, as shown in FIG. 10, an example that exactly matches the input sentence 1010 is not searched, and a similar source language example 1011 and a target language example 1012 corresponding to the source language example 1011 are obtained. Suppose that it was searched.

  In this case, the example translation unit 104 does not replace the element 1001 and the element 1002 that match in the input sentence 1010 and the source language example 1011, but the element 1007 corresponding to the element 1004 that is a different part corresponds to the element 1003. Replacement with English translation (step S804). Similarly, the example translation unit 104 replaces the element 1008 corresponding to the element 1006 which is a different part with an English translation corresponding to the element 1005 (step S804).

Example translation unit 104 calculates the certainty factor C _e of example translation of the translation result by the above formula (2) (step S805). In the example of FIG. 10, the element 1001 (other) and the element 1002 (other) completely coincide with each other, so the similarity is 1. Also, if the similarity between the element 1003 (shoe) and the element 1004 (person) is 0.2, and the similarity between the element 1005 (show me) and the element 1006 (please call) is 0.1 From Equation (2), C _e = (1 + 0.2 + 0.1) /3=0.43 is obtained as the certainty factor.

  In the example translation, if there is no matching information in the cache application process, the final certainty factor remains 0.43. By comparing the certainty factor 0.72 of the rule translation thus obtained with the certainty factor 0.43 of the example translation, the selection unit 102a finally selects the translation result of the rule translation having a high certainty factor. (Step S504). In addition, the translation control unit 102 registers additional information as a translation result of the selected rule translation in the cache memory 122 (step S505).

  As described above, in the machine translation device according to the present embodiment, additional information such as word correspondence, scene information, and politeness obtained from the translation results of rule translation and example translation is stored in the cache memory. The translation accuracy can be improved by referring to the additional information in the translation process.

  FIG. 11 is an explanatory diagram of a hardware configuration of the machine translation apparatus according to the present embodiment.

  The machine translation apparatus according to this embodiment includes a communication I / O that communicates with a control device such as a CPU (Central Processing Unit) 51 and a storage device such as a ROM (Read Only Memory) 52 and a RAM 53 by connecting to a network. F54 and a bus 61 for connecting each part are provided.

  The machine translation program executed by the machine translation apparatus according to the present embodiment is provided by being incorporated in advance in the ROM 52 or the like.

  The machine translation program executed by the machine translation apparatus according to the present embodiment is a file in an installable format or an executable format, and is a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R. (Compact Disk Recordable), DVD (Digital Versatile Disk) or the like may be provided by being recorded on a computer-readable recording medium.

  Furthermore, the machine translation program executed by the machine translation apparatus according to the present embodiment may be provided by being stored on a computer connected to a network such as the Internet and downloaded via the network. . The machine translation program executed by the machine translation apparatus according to the present embodiment may be provided or distributed via a network such as the Internet.

  The machine translation program executed by the machine translation apparatus according to the present embodiment has a module configuration including the above-described units (input reception unit, translation control unit, rule translation unit, example translation unit, output unit), As actual hardware, the CPU 51 reads out the machine translation program from the ROM 52 and executes the machine translation program, so that each unit is loaded on the main storage device, and each unit is generated on the main storage device.

  As described above, the machine translation apparatus, method, and program according to the present invention are suitable for a machine translation apparatus, method, and program that translate a source language sentence into a target language by combining a plurality of translation methods.

It is a block diagram which shows the structure of the machine translation apparatus concerning this Embodiment. It is explanatory drawing which shows an example of the data structure of the translation dictionary stored in the dictionary memory | storage part. It is explanatory drawing which shows an example of the information memorize | stored in a cache memory. It is explanatory drawing which shows an example of the data structure of the example memorize | stored in the example storage part. It is a flowchart which shows the whole flow of the machine translation process in this Embodiment. It is a flowchart which shows the whole flow of the rule translation process in this Embodiment. It is explanatory drawing which shows another example of the information memorize | stored in a cache memory. It is a flowchart which shows the whole flow of the example translation process in this Embodiment. It is explanatory drawing which shows an example of the dependence structure tree produced | generated by the rule translation process. It is explanatory drawing which shows an example of a response | compatibility with the example selected by the example translation process, and an input sentence. It is explanatory drawing which shows the hardware constitutions of the machine translation apparatus concerning this Embodiment.

Explanation of symbols

51 CPU
52 ROM
53 RAM
54 Communication I / F
61 Bus 100 Machine Translation Device 101 Input Accepting Unit 102 Translation Control Unit 102a Selection Unit 103 Rule Translation Unit 104 Example Translation Unit 105 Output Unit 121 Dictionary Storage Unit 122 Cache Memory 123 Example Storage Unit 901 Nodes 1001, 1002, 1003, 1004, 1005 , 1006, 1007, 1008 Element 1010 Input sentence 1011 Example for source language 1012 Example for target language

Claims (12)

An example storage unit for storing an example of the first language and an example of the second language obtained by translating the example of the first language in association with each other;
A dictionary storage unit for storing a word dictionary in which the first word in the first language is associated with the second word in the second language;
An input receiving unit for receiving an input of a first language sentence in a first language;
An example translation unit for example-translating the first language sentence into a second language by obtaining an example of the second language corresponding to the first language sentence from the example storage unit;
The second word corresponding to the first word included in the first language sentence is acquired from the dictionary storage unit, and the first word included in the first language sentence is determined based on a predetermined rule. A rule translation unit that translates the first language sentence into a second language by converting the acquired second word;
Among the translation results of the example translation unit and the translation results of the rule translation unit, a selection unit that selects the most probable translation result;
A correspondence storage unit that stores a correspondence relationship between the first word included in the first language sentence and the second word included in the selected translation result;
An output unit for outputting the selected translation result,
The rule translation unit further performs rule translation of another first language sentence into a second language and corresponds to the first word for each of the first words included in the other first language sentence. Obtaining the second word from the correspondence storage unit and replacing the second word obtained by regular translation of the first word with the obtained second word;
Machine translation device characterized by the above. The dictionary storage unit stores a word dictionary that associates the first word, the second word, and first scene information representing a scene using the second word,
The correspondence storage unit further stores second scene information representing a scene using the first language sentence,
The rule translation unit acquires the second scene information from the correspondence storage unit, the first word included in the first language sentence, and the first scene information that matches the acquired second scene information; By acquiring the second word associated with the first word included in the first language sentence based on the rules, and converting the first word into the acquired second word, Regularly translating the first language sentence into a second language;
The machine translation apparatus according to claim 1. The correspondence storage unit stores the second scene information in association with the correspondence relationship,
The rule translation unit acquires the second scene information associated with the correspondence stored last, from the correspondence storage unit, and the first word included in the first language sentence and the acquired The second word associated with the first scene information that matches the second scene information is acquired from the dictionary storage unit, and the first word included in the first language sentence is obtained based on the rules. Translating the first language sentence into a second language by converting it into the acquired second word;
The machine translation apparatus according to claim 2. The dictionary storage unit stores a word dictionary that associates the first word, the second word, and a first politeness level indicating the degree of politeness of the semantic content of the second word;
The correspondence storage unit further stores a second politeness level indicating the degree of politeness of the semantic content of the first language sentence;
The rule translation unit acquires the second politeness from the correspondence storage unit, the first word included in the first language sentence, and the first politeness that matches the acquired second politeness By acquiring the second word associated with the first word included in the first language sentence based on the rules, and converting the first word into the acquired second word, Regularly translating the first language sentence into a second language;
The machine translation apparatus according to claim 1. The correspondence storage unit stores the second politeness and identification information for identifying a speaker of the first language sentence in association with each other,
The input unit receives the input of the identification information of a speaker of the first language sentence together with the first language sentence,
The rule translation unit acquires the second politeness level associated with the received identification information from the correspondence storage unit, and acquires the first word included in the first language sentence and the acquired second politeness. The second word associated with the first politeness that matches the degree is acquired from the dictionary storage unit, and the first word included in the first language sentence is acquired based on the rules Translating the first language sentence into a second language by converting to the second word;
The machine translation apparatus according to claim 4. The input receiving unit further receives an input of a second language sentence in a second language,
The rule translation unit further acquires the first word corresponding to the second word included in the second language sentence from the dictionary storage unit, and is included in the second language sentence based on the rule. By converting the second word into the acquired first word, the second language sentence is regularly translated into the first language, and each of the second words included in the second language sentence is converted into the first word. Obtaining the first word corresponding to two words from the correspondence storage unit, replacing the first word obtained by regular translation of the second word with the obtained first word;
The machine translation apparatus according to claim 1. The example translation unit obtains an example of the first language corresponding to the second language sentence from the example storage unit, thereby translating the second language sentence to the first language.
The selection unit further selects the most probable translation result from the translation result of the example translation unit for the second language sentence and the translation result of the rule translation unit for the second language sentence,
The correspondence storage unit further stores a correspondence relationship between the second word included in the second language sentence and the first word included in the selected translation result;
The machine translation apparatus according to claim 6. The example storage unit stores the first language example, the second language example, and the first politeness indicating the degree of politeness of the semantic content of the second language example in association with each other,
The correspondence storage unit further stores a second politeness level indicating the degree of politeness of the semantic content of the first language sentence;
The example translation unit acquires the second politeness from the correspondence storage unit and associates the first politeness that matches the acquired second politeness with the first language sentence. Obtaining an example of the second language from the example storage unit;
The machine translation apparatus according to claim 1. The correspondence storage unit stores the second politeness and identification information for identifying a speaker of the first language sentence in association with each other,
The input unit receives the input of the identification information of a speaker of the first language sentence together with the first language sentence,
The example translation unit acquires the second politeness associated with the received identification information from the correspondence storage unit, the first politeness that matches the acquired second politeness, and the first Obtaining an example of the second language associated with a language sentence from the example storage unit;
The machine translation apparatus according to claim 8. The example translation unit further calculates a first certainty factor representing the certainty of the translation result,
The rule translation unit further calculates a second certainty factor representing the certainty of the translation result,
The selection unit selects a translation result of the example translation unit when the first certainty factor is larger than the second certainty factor, and the rule translation unit when the first certainty factor is smaller than the second certainty factor. Selecting translation results for
The machine translation apparatus according to claim 1. A first input receiving step of receiving an input of a first language sentence in a first language by an input receiving unit;
The second language corresponding to the first language sentence from the example storage unit that stores the example of the first language and the example of the second language translated from the example of the first language by the example translation unit An example translation step of translating the first language sentence into a second language by obtaining an example of
Corresponding to the first word included in the first language sentence from the dictionary storage unit that stores the word dictionary in which the first word in the first language and the second word in the second language are associated by the rule translation unit The second word is acquired, and the first language sentence is converted into the acquired second word by converting the first word included in the first language sentence based on a predetermined rule. A first rule translation step for rule translation into a second language;
The selection unit selects the most likely translation result from the translation result of the example translation step and the translation result of the rule translation step, the first word included in the first language sentence, and the selected translation result A selection step of storing a correspondence relationship between the second word included in the correspondence storage unit;
An output step for outputting the selected translation result by the output unit; and
A second input receiving step of further receiving an input of the first language sentence by an input receiving unit;
The example translation unit obtains the second language example corresponding to the first language sentence received by the second input receiving step from the example storage unit, thereby converting the first language sentence into the second language. A second example translation step for translating the example into
The rule translation unit obtains the second word corresponding to the first word contained in the first language sentence accepted by the second input acceptance step from the dictionary storage unit, and based on the rule, A second rule translation step of translating the first language sentence into a second language by converting the first word contained in the first language sentence into the acquired second word;
The rule translation unit obtains the second word corresponding to the first word from the correspondence storage unit for each of the first words included in the first language sentence received in the second input reception step. Replacing the second word obtained by regular translation of the first word with the acquired second word;
A machine translation method comprising: A first input acceptance procedure for accepting input of a first language sentence in a first language;
An example of the second language corresponding to the first language sentence is acquired from an example storage unit that stores an example of the first language and an example of the second language translated from the example of the first language. An example translation procedure for example-translating the first language sentence into a second language;
The second word corresponding to the first word included in the first language sentence from a dictionary storage unit that stores a word dictionary in which the first word in the first language is associated with the second word in the second language And converting the first word contained in the first language sentence into the acquired second word based on a predetermined rule, thereby converting the first language sentence into a second language. A first rule translation procedure to translate;
The most probable translation result is selected from the translation result of the example translation procedure and the translation result of the rule translation procedure, and the first word included in the first language sentence and the translation result included in the selected translation result A selection procedure for storing the correspondence relationship with the second word in the correspondence storage unit;
An output procedure for outputting the selected translation result;
A second input receiving procedure for further receiving input of the first language sentence;
A second example of translating the first language sentence into a second language is obtained by obtaining an example of the second language corresponding to the first language sentence accepted by the second input acceptance procedure from the example storage unit. 2 example translation procedures;
The second word corresponding to the first word included in the first language sentence accepted by the second input acceptance procedure is acquired from the dictionary storage unit, and the first language sentence is obtained based on the rules. A second rule translation procedure for rule-translating the first language sentence into a second language by converting the first word contained in the acquired second word;
For each of the first words included in the first language sentence accepted by the second input acceptance procedure, the second word corresponding to the first word is obtained from the correspondence storage unit, and the obtained second word A replacement procedure for replacing the second word obtained by regular translation of the first word with two words;
Machine translation program that causes a computer to execute

Images