JP2015170094A - Translation device and translation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To properly translate a sentence even in a case where the single sentence is not enough to determine a dictionary to be used to translate the sentence.SOLUTION: The translation device 10 comprises: a translation engine group 11 which stores a plurality of dictionaries to be used for translation by associating the dictionaries with categories to perform translation processing; an input unit 12 to which the sentence to be translated is input; a determination unit 13 which determines a category of the sentence; an information storage unit 14 which stores category information for specifying the category of the sentence; an update unit 15 which updates the category information when the category is determined and does not update the category information when the category is not determined; a translation engine distribution unit 16 which outputs the sentence to a translation engine 110 that is associated with the category specified according to the stored category information; an automatic input correction unit 17 which corrects the sentence to be translated; and an output unit 18 which outputs the translated sentence.

Description

  The present invention relates to a translation apparatus and a translation method for translating a sentence.

  The machine translation service generates a translation by obtaining a translation result having the highest degree of appropriateness based on the certainty of the language and the certainty of the translation. In Patent Document 1, a translation target feature vector indicating the characteristics of the translation target sentence is created, and a specialized field feature vector indicating the characteristics of the specialized field sentence is created. A technique for estimating and selecting specialized dictionary data suitable for translation based on the estimation has been proposed.

JP 2007-80221 A

  However, the category (field) to which the sentence belongs cannot be determined only from each sentence to be translated, and the translation may not be performed adequately. For example, consider a case where a sentence to be translated is "How is your body?" The main body chapter is translated as “How is state of body?”, For example, in a general machine translation engine. However, if the above sentence relates to a conversation in the medical field, for example, the above translation is not necessarily appropriate. For example, the above sentence is translated as “How is condition of body?” Or the like in a translation engine having a medical dictionary database. Thus, depending on the background of conversation, it is desirable to translate using a translation engine having a dictionary database in a specific field rather than a general machine translation engine. However, it is difficult to sufficiently determine whether or not it is only necessary to perform translation specialized for the medical field by using the above sentence alone.

  The present invention has been made in view of the above problems, and even when it is not possible to sufficiently determine what dictionary should be used for translation from a single sentence, the sentence is appropriately It is an object of the present invention to provide a translation apparatus and a translation method that can translate.

  In order to achieve the above object, a translation apparatus according to the present invention includes dictionary holding means for holding a plurality of dictionaries used for translation in association with a category, and category information for specifying a category of a sentence to be translated. A category information holding means for holding; an input means for inputting a sentence in a first language to be translated; a determining means for determining a category of the sentence based on a sentence input by the input means; and a determining means When the category is determined, the category information held by the category information holding unit is updated with the category information for specifying the determined category. When the category is not determined by the determining unit, the category information holding unit The update means that keeps the category information that is held, and the category information holding means Translation means for translating a sentence input by the input means from the first language to the second language using a dictionary associated with the category specified by the category information and held in the dictionary holding means. Output means for outputting a sentence translated by the translation means.

  In the translation apparatus according to the present invention, the category is determined from the input sentence to be translated. When the category is determined, translation is performed using a dictionary associated with the determined category. On the other hand, if the category is not determined, translation is performed using a dictionary associated with the category determined based on the sentence input before that. That is, in the translation device according to the present invention, even if it is not possible to sufficiently determine what dictionary should be used for translation from a single sentence, it is determined based on a series of sentences such as conversations. Translation is performed using a dictionary associated with the category. Therefore, according to the translation apparatus according to the present invention, it is possible to appropriately translate a sentence even in the above case.

  The determination means may estimate the degree to which the text input by the input means belongs to a category by machine learning, and determine the category of the text based on the degree. According to this configuration, it is possible to appropriately determine the category including whether or not the category can be determined. Thereby, a sentence can be translated more appropriately.

  The input means inputs sentences in the second language to be translated, and the translation means is associated with the category specified by the category information held by the category information holding means and held in the dictionary holding means. The sentence input by the input means may be translated from the second language into the first language using the existing dictionary. According to this configuration, for example, a conversation in a different language can be performed after appropriate translation is performed.

  The translation unit may convert the sentence to be translated based on a rule set in advance according to the category specified by the category information held by the category information holding unit, and may translate the converted sentence. . According to this configuration, the translated sentence can be made more appropriate.

  By the way, the present invention can be described as an invention of a translation apparatus as described below, in addition to the invention of a translation apparatus as described above. This is substantially the same invention only in different categories, and has the same operations and effects.

  That is, the translation method according to the present invention includes a dictionary holding unit that holds a plurality of dictionaries used for translation in association with a category, and a category information holding unit that holds category information for specifying a category of a sentence to be translated. A translation method that is an operation method of a translation apparatus comprising: an input step for inputting a sentence in a first language to be translated; and a category of the sentence based on the sentence input in the input step And when the category is determined in the determination step, the category information held by the category information holding means is updated with the category information for specifying the determined category, and the category is not determined in the determination step Keeps the category information held by the category information holding means. Using the dictionary stored in the dictionary holding unit in association with the category specified by the update step and the category information held by the category information holding unit, the sentence input in the input step is converted into the first language. A translation step for translating the text into the second language, and an output step for outputting the text translated in the translation step.

  In the present invention, even if it is not possible to sufficiently determine what dictionary should be used for translation from a single sentence, it is associated with a category determined based on a series of sentences such as conversation. A dictionary is used for translation. Therefore, according to the present invention, even in the above case, the sentence can be appropriately translated.

It is a figure which shows the structure of the translation apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the text translated by the translation apparatus. It is a figure which shows the example of the display screen of a user terminal. It is a certainty factor determination pattern table | surface which shows the threshold value used in order to judge whether the presumed label is judged. It is a translation engine distribution pattern table | surface with which the presumed label and the translation engine number were matched. It is a figure which shows the example of an input sentence, and the example of a machine translation sentence. It is a figure which shows the example of the E estimation dictionary containing the word used as the candidate of E category for every P category. It is a figure which shows the outline | summary of estimation of E category. It is a figure which shows the example of the standard phrase memorize | stored for every category and P category. It is a figure which shows the outline | summary of determination of a standard phrase. It is a figure which shows the hardware constitutions of the translation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the whole process (translation method) performed with the translation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the estimation label determination process among the processes (translation method) performed with the translation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the automatic correction process of an input sentence among the processes (translation method) performed with the translation apparatus which concerns on embodiment of this invention.

  Hereinafter, embodiments of a translation apparatus and a translation method according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 shows a translation apparatus 10 according to this embodiment. The translation apparatus 10 is an apparatus that inputs a sentence to be translated, translates the sentence (machine translation), and outputs the translated sentence. That is, the translation device 10 is a device that provides a translation service. For example, the translation apparatus 10 is connected to the user terminal 20 via a communication network such as a mobile communication network or the Internet. The translation device 10 receives a sentence to be translated from the user terminal 20 and transmits the translated sentence to the user terminal 20.

  The translation device 10 sequentially translates a plurality of series of sentences. The translation device 10 may translate the sentence in the first language into the second language and translate the sentence in the opposite second language into the first language. For example, the translation apparatus 10 may translate Japanese sentences into English and translate English sentences into Japanese. Specifically, the translation apparatus 10 translates a series of conversations between the users 30a and 30b as shown in FIG.

  The user terminal 20 is a device that can communicate with the translation device 10 via a communication network, and corresponds to, for example, a smartphone, a mobile phone, or a PC (Personal Computer). The user terminal 20 establishes a session with the translation apparatus 10, inputs a translation target sentence, and transmits it to the translation apparatus 10. In addition, the user terminal 20 inputs translation direction information (for example, information indicating that the translation is from Japanese to English) indicating which language is translated from which language, and matches the transmission of the sentence. The information may be transmitted to the translation apparatus 10. The session between the user terminal 20 and the translation device 10 is uniquely identified by a session ID or the like, and the translation device 10 can grasp from which user terminal 20 the sentence to be translated is transmitted.

  The user terminal 20 inputs a sentence to be translated, for example, by inputting a voice from the user and recognizing the voice as a sentence. Speech recognition is performed using a conventionally used speech recognition technique. Alternatively, the user terminal 20 may transmit the input speech to the translation device 10 and speech recognition may be performed on the translation device 10 side. Alternatively, the speech recognition may be performed by a speech recognition server different from the translation device 10. In that case, the user terminal 20 transmits voice to the voice recognition server, and receives a sentence that is a voice recognition result from the voice recognition server. The sentence to be translated may be input by the user operating a button on the user terminal 20 or the like other than the input by voice.

  Moreover, the user terminal 20 inputs translation direction information based on a user's operation. This input is performed by an interface as shown in FIG. 3, for example. FIG. 3 shows a display screen 21 of a touch panel provided in the user terminal 20. The display screen 21 is provided with buttons 21a and 21b for inputting translation direction information. For example, when the user speaks after touching the button 21a described as “speak”, translation direction information indicating that Japanese is translated into English is input together with a sentence based on the utterance. Further, when the user utters after touching the button 21b described as “Speak”, translation direction information indicating that English is translated into Japanese is input together with a sentence based on the utterance.

  Further, the user terminal 20 receives the translation result text from the translation device 10. The user terminal 20 displays the received translation result text on the display screen 21 of the touch panel. Thereby, the user can recognize the translation result. At this time, as shown in FIG. 3, a sentence (translation target sentence) 21 c input from the user and a translated sentence 21 d of the sentence may be displayed together. Alternatively, a sentence related to translation from English to Japanese may be displayed in a balloon 21e from the right side, and a sentence related to translation from Japanese to English may be displayed in a balloon 21f from the left side.

  Subsequently, the functional configuration of the translation apparatus 10 according to the present embodiment will be described. As shown in FIG. 1, the translation apparatus 10 includes a translation engine group 11, an input unit 12, a determination unit 13, an update unit 15, a category information holding unit 14, a translation engine distribution unit 16, and an automatic input correction unit. 17 and an output unit 18. Among these components, the determination unit 13, the update unit 15, the category information holding unit 14, the translation engine distribution unit 16, and the automatic input correction unit 17 can be realized by a category estimation / distribution program. it can.

  The translation engine group 11 is a functional means group that performs a translation process on the text input to the translation apparatus 10. The translation engine group 11 includes a plurality of translation engines 110a to 110d (translation engines A to D). As will be described later, the text input to the translation device 10 is input to any of a plurality of translation engines 110 (generic name of the translation engines 110a to 110d) included in the translation engine group 11. Each translation engine 110 translates the input sentence and outputs it to the output unit 18.

  The translation engine 110 includes a processing unit 111 (generic name for the processing units 111a to 111d) and a translation dictionary 112 (generic name for the translation dictionaries 112a to 112d). The translation dictionary 112 is a dictionary holding unit that holds dictionary information used for translation. The dictionary information held by the translation dictionary 112 is, for example, information indicating correspondence between words in the translation source language and words in the translated language, and is similar to the dictionary used in conventional machine translation. Also good. The dictionary held by the translation dictionary 112 is associated with a category corresponding to the translation engine 110. The category indicates a field (specialized field) to which the sentence to be translated belongs, and includes, for example, “medical”, “finance”, “science (technology)”, and the like. The translation dictionary 112 is prepared in advance by an administrator of the translation apparatus 10 or the like.

  For example, the translation dictionary 112a included in the translation engine 110a (translation engine A) holds a basic translation dictionary for translating general sentences, not a category related to a special field. When translation is performed using the basic translation dictionary, a translation that is the translation result with the highest probability of probability is generated based on the certainty of the language and the certainty of the translation. For this reason, the translation accuracy of input sentences with expertise such as medical care, finance, and science and technology is not necessarily appropriate, and there is a possibility that the translation result may be strange.

  The translation dictionary 112b included in the translation engine 110b (translation engine B) holds a medical translation dictionary related to the “medical” category. The translation dictionary 112c included in the translation engine 110c (translation engine C) holds a financial translation dictionary related to the category of “finance”. The translation dictionary 112d included in the translation engine 110d (translation engine D) holds a scientific translation dictionary related to the category of “science”.

  The processing unit 111 is a translation unit that translates the text input to the translation engine 110 from the first language to the second language using the translation dictionary 112. The processing unit 111 may perform translation between a plurality of languages (any one or both of the first language and the second language described above). In that case, translation direction information is also input to the translation engine 110 together with the text to be translated, and translation is performed in the direction between languages indicated by the translation direction information. For example, the processing unit 111 is realized by a technique similar to that of conventional machine translation. Note that the processing unit 111 may be different for each translation engine 110 or may be shared among the translation engines 110.

  Note that each translation engine 110 is not necessarily realized by a device similar to the constituent elements of the translation apparatus 10 other than each translation engine 110. Each translation engine 110 and the components other than each translation engine 110 of the translation apparatus 10 may be configured as separate devices. That is, the translation apparatus 10 may be configured by a computer system including a plurality of computers.

  The input unit 12 is input means for inputting a sentence to be translated. Specifically, the input unit 12 establishes a session between the user terminal 20 and the translation device 10 and inputs a sentence by receiving a sentence to be translated from the user terminal 20 via the session. . In the translation apparatus 10, the session can be grasped by the session ID, and thereby it can be grasped from which user terminal 20 the information is inputted. Further, as described above, it is also possible to receive a voice from the user terminal 20 and recognize the voice to obtain a sentence. Further, when the translation apparatus 10 can execute a plurality of translations, the input unit 12 may input translation direction information. The translation direction information includes information indicating the language (input language) of the sentence to be translated and information indicating the language of the translated sentence (translation language). The following processing in this embodiment is performed based on the input language and the translation language. That is, for each sentence before translation, processing for each input language is performed, and translation from the input language to the translation language is performed. The input unit 12 receives the information from the user terminal 20 together with the text to be translated. The input unit 12 outputs the input information to the determination unit 13.

  The determination unit 13 is a determination unit that determines the category of the sentence based on the sentence input from the input unit 12. For example, the determination unit 13 estimates the degree to which a sentence input from the input unit 12 belongs to a category by machine learning, and determines the category of the sentence based on the degree. In the present embodiment, the category of the sentence to be determined is referred to as an estimated label. In addition, the degree to which a sentence belongs to a category is called confidence. For example, if the input sentence that is the input sentence is “Which is internal medicine?” By the determination by the determination unit 13, an output such that the estimated label is “medical” and the certainty factor is 78% is output. .

  In order to implement the determination unit 13 according to the present embodiment, sentences associated with categories are collected in advance. Supervised machine learning is performed as correct answer data (sample data) associated with a category to obtain a learning database (learning model). For example, a general-purpose SVM (Support Vector Machine) can be used as the logic for machine learning. The determination unit 13 inputs the text input from the input unit 12 to a discriminator based on the learning database, and determines the category by the discriminator.

  Specifically, the determination unit 13 performs morphological analysis of the input sentence input from the input unit 12. Morphological analysis can be performed using a general-purpose morphological analysis engine. For example, if the input sentence is “Tell me the disease name”, the result of morphological analysis of “hospital (noun) / to (particle) / teaching (verb) / te (particle)” is obtained. “/” In the result of morpheme analysis is a morpheme break, and parentheses indicate the part of speech of each morpheme. Subsequently, the determination unit 13 generates a feature ID that is SVM input data for each morpheme from the morpheme analysis result. Specifically, the determination unit 13 converts each morpheme into a feature ID. The determination unit 13 stores a correspondence relationship between the morpheme and the feature ID in advance, and performs the above modification based on the correspondence relationship. For example, the morpheme “hospital” converts the feature ID to “1”, and the morpheme “teach” converts the feature ID to “2”.

  The determination unit 13 outputs (generates) an estimated label and a certainty factor by SVM using the obtained input data of SVM. The above can be performed by utilizing general-purpose SVM engine logic. In order to estimate a plurality of estimated labels, multi-class SVM estimation may be performed, or a learning database for each estimated label may be used. The determination unit 13 outputs an estimation label indicating the determination result and information indicating the certainty factor to the update unit 15. Further, the determination unit 13 outputs the information input from the input unit 12 to the translation engine distribution unit 16.

  Note that the determination unit 13 only needs to be able to use a discriminator based on a learning database obtained by machine learning, and machine translation need not necessarily be performed in the translation apparatus 10. In that case, the translation apparatus 10 acquires in advance information indicating a learning database from the apparatus that performed the machine learning. The determination unit 13 performs the above determination by SVM, but other methods such as a pattern recognition method may be used. Moreover, the determination part 13 may determine a category from a sentence by methods other than said machine learning.

  The category information holding unit 14 is a category information holding unit that holds category information that specifies an estimated label of a sentence to be translated. Specifically, the category information holding unit 14 holds a session ID and category information in association with each other. The category information holding unit 14 receives the update instruction from the update unit 15 and updates the above information. The category information holding unit 14 receives an inquiry about the information held from the translation engine sorting unit 16 from the translation engine sorting unit 16 and returns the category information. The category information holding unit 14 releases (deletes) information held for each session when there is no information update or inquiry for a predetermined time for the same session ID. Also, when a session is disconnected, it may be released in the same manner.

  When the determination unit 13 determines the estimated label, the update unit 15 updates the category information held by the category information holding unit 14 with the category information that identifies the determined estimated label, and the determination unit 13 When the estimated label is not determined, the updating unit keeps the category information held by the category information holding unit 14.

  Specifically, the update unit 15 determines whether or not the estimated label is determined by the determination unit 13 based on the information indicating the estimated label and the certainty factor input from the determination unit 13 (the category is determined from the sentence). Whether it was possible). The update unit 15 makes the above determination by comparing the certainty factor with a preset threshold value. For example, if the certainty factor exceeds the threshold value, the update unit 15 determines that the estimated label is determined by the determination unit 13. If the certainty factor is below the threshold value, the update unit 15 determines that the estimation label is not determined by the determination unit 13 (the sentence does not belong to a specific category). The threshold for making this determination may be set for each estimated label as shown in the certainty determination pattern table of FIG.

  When the update unit 15 determines that the estimated label is determined, the update unit 15 inputs the category information specifying the estimated label to the category information holding unit 14 and updates the category information held in the category information holding unit 14. This update is performed on the category information held in association with the session ID related to the input sentence. When there is no category information held in the category information holding unit 14, the category information input from the updating unit 15 is newly held. If the update unit 15 determines that the estimated label has not been determined, the update unit 15 does not update the category information (in this case, no special processing is performed).

  The translation engine sorting unit 16 is a function of a translation unit that sorts the translation engine 110 that translates the text input from the determination unit 13. The translation engine sorting unit 16 performs the sorting described above based on the category information held by the category information holding unit 14. This distribution process is performed after the above process by the updating unit 15 is completed.

  Specifically, the translation engine sorting unit 16 stores in advance information in which an estimated label and a translation engine number that identifies the translation engine 110 are associated with each other, as shown in the translation engine sorting pattern table of FIG. . Further, as shown in FIG. 5, this information may also be associated with a combination of the input language and the translation language. When a sentence is input from the determination unit 13, the translation engine distribution unit 16 refers to the category information held in the category information holding unit 14 in association with the session ID related to the input sentence. Specifically, the translation engine sorting unit 16 may request the category information holding unit 14 to make a category inquiry request together with the session ID related to the input sentence, and acquire the category information from the category information holding unit 14. Good. The translation engine distribution unit 16 uses the translation engine 110 identified by the translation engine number associated with the estimated label indicated by the referenced category information in the translation engine distribution pattern table of FIG. decide. If there is no category information associated with the session ID associated with the input sentence and retained in the category information retaining unit 14, the translation engine allocating unit 16 selects a preset translation engine 110 as a distribution destination. The translation engine 110 is determined. For example, the translation engine 110 identified by the translation engine number associated with the estimated label “others” in FIG. 5 (for example, the translation engine 110a including the translation dictionary 112a for translating general sentences (translation engine A)) ) Is determined as the translation engine 110 of the distribution destination.

  Further, when information indicating the input language and the translation language is input as the translation direction information from the determination unit 13, the translation engine 110 specified by the translation engine number associated with the combination is further distributed. It is determined as the previous translation engine 110.

  The translation engine sorting unit 16 outputs the sentence to be translated to the translation engine 110 that is the sorting destination. When the translation engine 110 is configured by another device, the translation engine number is associated with the IP address of the server that constitutes the translation engine 110, and the sentence to be translated is transmitted to the IP address. The translation engine number itself may be an IP address. Further, the translation engine distribution unit 16 may output information indicating the input language and the translation language to the translation engine 110 as translation direction information.

  The translation engine distribution unit 16 may cause the automatic input correction unit 17 to correct the sentence before outputting the sentence to be translated to the translation engine 110. In that case, the translation engine distribution unit 16 outputs the sentence corrected by the automatic input correction unit 17 to the translation engine 110.

  The automatic input correction unit 17 converts the sentence to be translated input from the translation engine distribution unit 16 based on a rule set in advance according to the estimated label specified by the category information held by the category information holding unit 14. This is a function of translation means for conversion.

  When the category of the input sentence is estimated, an appropriate translation engine 110 can be selected. However, by further correcting the input sentence for each category into a standard wording and inputting it to the translation engine 110, it is possible to derive a more accurate translation result for each category and improve the translation quality. For example, FIG. 6 shows an example of an input sentence when it is claimed that the head is very painful in medical treatment. However, depending on the input format, it may be mistranslated, such as an input sentence such as “I'll be angry with a headache”. The automatic input correction unit 17 corrects the input sentence to a standard phrase according to the category.

  The automatic input correction unit 17 stores in advance information in which an estimated label and a rule for converting (correcting) a sentence are associated with each other. The automatic input correction unit 17 refers to the category information held in the category information holding unit 14 in association with the session ID related to the input sentence. The automatic input correction unit 17 converts (corrects) a sentence according to a rule stored in association with the referenced category information.

  Hereinafter, an example of conversion in the “medical” category will be described. First, the automatic input correction unit 17 estimates a P category indicating what the input sentence means in the category from the input sentence. For this estimation, for example, a learning database created by machine learning in advance to estimate the P category is used. The correct answer data (sample data) used for machine learning is, for example, a sentence associated with the P category. For example, if the input sentence includes phrases such as “~ hurts like dull”, “~ hurts like rippling”, “~ hurts like being tightened”, “~ hurts like cracking”, the P category Is estimated to be “(disease state) pain: high”.

  Subsequently, the automatic input correction unit 17 estimates the E category indicating the target of the P category from the input sentence. The automatic input correction unit 17 previously stores words that are candidates for the E category for each P category as a dictionary (E estimation dictionary). FIG. 7 shows an example of the E estimation dictionary. In the example illustrated in FIG. 7, words that are candidates for the E category with respect to the P category of “pain: high” are associated with each other. The automatic input correction unit 17 determines whether or not the input sentence includes a word that is a candidate for the E category associated with the estimated P category. If a word that is a candidate for the E category is included, the automatic input correction unit 17 extracts the word as a word indicating the E category. For example, as shown in FIG. 8, when the input sentence is “slow and painful head” and the P category is “pain: high”, the E category is “head”.

  Subsequently, the automatic input correction unit 17 generates a standard phrase from the category (estimated label), the P category, and the E category. Based on the input sentence, the E category indicating the target of the P category is estimated. The automatic input correction unit 17 stores a standard phrase for each category and P category in advance. FIG. 9 shows an example of standard phrases stored for each category and P category. The stored standard phrase includes a portion where a word is input. In the example of FIG. 9, this part is indicated by “˜”.

  The automatic input correction unit 17 extracts a standard phrase stored in association with the combination of the category and P category determined (estimated) as described above. The automatic input correction unit 17 inserts the E category in the portion where the extracted standard phrase word is input, and completes the standard phrase. For example, as shown in FIG. 10, when the category is “medical”, the P category is “pain: high”, and the E category is “head”, the extracted standard phrase is “~ is sharp and painful”. , The standard phrase generated is “Heart is sharp and painful.”

  The automatic input correction unit 17 outputs the generated standard phrase to the translation engine distribution unit 16 as corrected text. When the input sentence is automatically corrected to a standard phrase in this way, the (machine) translation engine 110 can output an optimal translation result according to the input meaning. In addition, when automatic correction is performed in this way, a sentence automatically corrected together with the sentence of the translation result may be transmitted to the user terminal 20 to notify the user of the result of automatic correction.

  The output unit 18 is an output unit that outputs sentences translated by the translation engine 110. Specifically, the output unit 18 outputs the sentence by transmitting the translated sentence to the user terminal 20 via a session between the user terminal 20 and the translation device 10. The text received by the user terminal 20 is displayed and output on the user terminal 20 as described above. The functional configuration of the translation apparatus 10 according to the present embodiment has been described above.

  FIG. 11 shows a hardware configuration of the translation apparatus 10 according to the present embodiment. As shown in FIG. 11, the translation apparatus 10 includes a central processing unit (CPU) 101, a random access memory (RAM) 102 and a read only memory (ROM) 103 that are main storage devices, a communication module 104 for communication, and a hard disk. The computer is configured to include a computer including hardware such as the auxiliary storage device 105. The functions of the translation apparatus 10 described above are exhibited when these components are operated by a program or the like. The above is the configuration of the translation apparatus 10 according to the present embodiment.

  Subsequently, a translation method, which is an operation method of the translation apparatus 10 according to the present embodiment (a process executed by the translation apparatus 10), will be described using the flowcharts of FIGS. The entire translation method will be described with reference to the flowchart of FIG. In this process, first, an input sentence that is a sentence to be translated is transmitted from the user terminal 20 to the translation apparatus 10. In the translation apparatus 10, the input sentence is received by the input unit 12 (S01, input step). The input sentence is output from the input unit 12 to the determination unit 13.

  Subsequently, the determination unit 13 determines the estimated label based on the input sentence (S02, determination step). This determination process will be described with reference to the flowchart of FIG. First, morphological analysis is performed on the input sentence (S21). Subsequently, from the morpheme analysis result, a feature ID that is SVM input data is generated for each morpheme (S22). Subsequently, multi-class SVM estimation is performed using the obtained SVM input data (S23). The estimated label and the certainty factor obtained by the estimation are output (generated) (S24). Information indicating the estimated label and the certainty factor is output from the determination unit 13 to the update unit 15. Information input from the input unit 12 is output from the determination unit 13 to the translation engine distribution unit 16.

  Returning to FIG. 12, subsequently, the updating unit 15 determines whether or not the estimated label is determined by the determining unit 13 (S03, update step). Specifically, this determination is made by comparing the certainty factor with a preset threshold value. If the certainty level exceeds the threshold value, it is determined that the estimated label is determined (YES in S03). In this case, the update unit 15 inputs the category information specifying the estimated label to the category information holding unit 14, and the category information held in the category information holding unit 14 is updated (S04, update step). On the other hand, if the certainty factor is below the threshold value, it is determined that the estimated label has not been determined (NO in S03). In this case, the above update process is not performed.

  Subsequently (after NO in S04 or S03), the input sentence is output from the translation engine distribution unit 16 to the automatic input correction unit 17. Subsequently, the automatic input correction unit 17 corrects (converts) the input sentence based on a rule set in advance according to the estimated label specified by the category information held by the category information holding unit 14 (S05). , Translation step). This correction process will be described with reference to the flowchart of FIG. First, the P category is estimated from the input sentence based on a learning database created by machine learning (S51). Subsequently, the E estimation dictionary is used to estimate the E category from the input sentence (S52). Subsequently, a standard phrase is generated as a corrected sentence from the category (estimated label), the P category, and the E category based on the previously stored category and the standard phrase for each P category (S53). The corrected sentence is output from the automatic input correction unit 17 to the translation engine distribution unit 16.

  Returning to FIG. 12, subsequently, the translation engine sorting unit 16 sorts the translation engine 110 that translates the sentence to be translated (corrected sentence) based on the category information held by the category information holding unit 14. (S06, translation step). This distribution is performed based on information stored in advance in which the estimated label and the translation engine number are associated with each other. The translation target text (corrected text) is output from the translation engine sorting section 16 to the translation engine 110 that is the sorting destination.

  Subsequently, the distribution destination translation engine 110 inputs a sentence, and the translation dictionary 112 provided in the translation engine 110 is used to translate the sentence (S07, translation step). The translated text is output from the translation engine 110 to the output unit 18. Subsequently, the output unit 18 transmits the translated sentence to the user terminal 20 (S08, output step). The user terminal 20 receives the translated text and outputs it in a format that the user can recognize, such as a display. The above is the translation method that is the process executed by the translation apparatus 10 according to the present embodiment.

  The above process is performed every time an input sentence is transmitted from the user terminal 20. Therefore, translation is performed on a series of input sentences. As described above, in this embodiment, the category is determined from the input sentence. When the category is determined, translation is performed using a dictionary associated with the determined category. On the other hand, if the category is not determined (the input sentence is determined not to belong to a specific category), the dictionary associated with the category determined based on the previous input sentence is Used to translate. That is, in this embodiment, even if it is not possible to sufficiently determine what dictionary should be used for translation from a single sentence, it corresponds to a category determined based on a series of sentences such as conversation. The attached dictionary is used for translation. Therefore, according to this embodiment, even in the above case, the sentence can be appropriately translated. It is possible to select an optimal translation engine that is suitable for translation of specialized input sentences such as medical care, finance, and science and technology, and improve the quality of translation results. In addition, since the dictionary of a specific category can be expanded independently of the basic dictionary, the translation accuracy for consumers is not affected.

  Specifically, as in the example illustrated in FIG. 2, for example, after the category is determined as “medical” based on the sentence “the head hurts like undulating”, until the category is newly determined, Translation is performed using a medical dictionary. Accordingly, even if it is a sentence that cannot be determined as to which category, “How is the state?” After that, the “medical” translation dictionary is used for appropriate translation.

  Further, as in the present embodiment, the category determination may be performed by machine learning. According to this configuration, it is possible to appropriately determine the category including whether or not the category can be determined. Thereby, a sentence can be translated more appropriately.

  Further, as in the present embodiment, the sentence in the first language may be translated into the second language, and the sentence in the opposite second language may be translated into the first language. For example, as shown in FIG. 2, a series of sentences to be translated may include a sentence translated from Japanese to English and a sentence translated from Japanese to English. In this case, when the category (for example, “medical”) is determined in one direction, the category (for example, “medical”) can be automatically applied to the conversation translation in the reverse direction. Thereby, it is possible to perform two-way conversation in different languages after appropriate translation. Note that it is not always necessary to perform translation in the reverse direction, and translation in only one direction may be performed.

  Moreover, it is good also as performing an automatic correction | amendment of an input sentence like this embodiment. According to this configuration, the sentence translated as described above can be made more appropriate. However, when it is not necessary to automatically correct the input sentence, it is not always necessary to automatically correct the input sentence.

  In the present embodiment, the translation apparatus 10 is configured to communicate with the user terminal 20 to transmit / receive an input sentence and a translated sentence. However, the terminal used by the user such as the user terminal 20 may be provided with the function of the translation device 10. In that case, the terminal is a translation apparatus according to an embodiment of the present invention. In addition, the terminal may have some functions of the translation device 10.

  Further, the category information may be updated as in the present embodiment only in certain cases such as conversation translation. In this case, except for the fixed case, the category is determined for each sentence, and translation according to the determined category is performed (in the case of the present embodiment, the previous category determination result is not used). In the case of conversational translation, it may be input to the translation apparatus 10 that it is conversational translation before the translation process. Or, if there is a reverse translation request with the same session ID (for example, if a translation from Japanese to English was performed before that and there was a translation request from English to Japanese), conversational translation It may be set to be.

  DESCRIPTION OF SYMBOLS 10 ... Translation apparatus, 11 ... Translation engine group, 110 (110a-110d) ... Translation engine, 111 (111a-111d) ... Processing part, 112 (112a-112d) ... Translation dictionary, 12 ... Input part, 13 ... Determination part 14 ... Category information holding unit, 15 ... Update unit, 16 ... Translation engine sorting unit, 17 ... Automatic input correction unit, 18 ... Output unit, 101 ... CPU, 102 ... RAM, 103 ... ROM, 104 ... Communication module, 105 ... Auxiliary storage device, 20 ... User terminal.

Claims (5)

Dictionary holding means for holding a plurality of dictionaries used for translation in association with categories;
Category information holding means for holding category information for specifying a category of a sentence to be translated;
An input means for inputting a sentence in a first language to be translated;
Determination means for determining a category of the sentence based on the sentence input by the input means;
When the category is determined by the determination unit, the category information held by the category information holding unit is updated with the category information that identifies the determined category, and the category is not determined by the determination unit. Updating means for keeping the category information held by the category information holding means;
Using the dictionary held in the dictionary holding means in association with the category specified by the category information held by the category information holding means, the sentence input by the input means is read from the first language. Translation means for translating into a second language;
Output means for outputting sentences translated by the translation means;
A translation device comprising:   The translation device according to claim 1, wherein the determination unit estimates a degree to which the sentence input by the input unit belongs to the category by machine learning, and determines the category of the sentence based on the degree. The input means inputs a sentence in a second language to be translated,
The translation unit associates a sentence input by the input unit with a dictionary that is associated with the category specified by the category information held by the category information holding unit and is held by the dictionary holding unit. The translation device according to claim 1, wherein the translation device translates the second language into the first language.   The translation unit converts a sentence to be translated based on a rule set in advance according to a category specified by category information held by the category information holding unit, and translates the converted sentence. The translation apparatus as described in any one of 1-3. An operation method of a translation apparatus comprising: a dictionary holding unit that holds a plurality of dictionaries used in translation in association with a category; and a category information holding unit that holds category information that specifies a category of a sentence to be translated. A translation method,
An input step for inputting a sentence in a first language to be translated;
A determination step of determining a category of the sentence based on the sentence input in the input step;
When the category is determined in the determination step, the category information held by the category information holding unit is updated with the category information specifying the determined category, and when the category is not determined in the determination step An update step for keeping the category information held by the category information holding means;
Using the dictionary held in the dictionary holding unit in association with the category specified by the category information held by the category information holding unit, the sentence input in the input step is read from the first language. A translation step for translating into a second language;
An output step of outputting the sentence translated in the translation step;
Translation method including

Images