JP2017199363A - Machine translation device and computer program for machine translation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation device which appropriately translates a text differently in accordance with information of a range exceeding the text.SOLUTION: A machine translation device 230 includes: a grammar type determination unit 282 which specifies a grammar type of a text; a grammar type specific tag assigning unit 286 which respectively inserts a first tag and a second tag corresponding to the grammar type at a head position and an end position of the text; and a phrase based machine translation unit 288 which receives as an input the text to which the first tag and the second tag are assigned. Plural kinds of predetermined grammar types are defined. The grammar type determination unit 282 selects a first tag and a second tag according to a kind of the grammar type.SELECTED DRAWING: Figure 6

Description

  The present invention relates to a machine translation apparatus, and more particularly, to a machine translation apparatus capable of accurately reflecting a difference in a translation original sentence in a translated sentence and performing a highly accurate translation, and a computer program for machine translation.

  There are various types of statistical machine translation, and a phrase-based (phrase-based) statistical machine translation (PBSMT) is one of the most promising machine translation systems. PBSMT divides the original text into a chain of several words called phrases, translates each chain into a phrase in the partner language, and rearranges the translated phrases (Non-Patent Document 1). The phrase used here is different from the phrase used in linguistics and simply refers to a chain of words. Phrase-based translation can be learned automatically from parallel translation data. For example, in the English-Japanese translation, "Hello!" Is can be dealt with automatically "Hello!" Or "Hello." And the like. In the following description, the original translation is described in Japanese and the translation is in English, but the same applies to other languages.

  PBSMT is fast and can perform machine translation with high accuracy, especially between languages with similar structures. Furthermore, as a recent development, by applying the PBSMT after converting the original phrases so that the phrases in the original language are close to the word order of the counterpart language, the word order becomes larger, such as English and Japanese, and Chinese and Japanese. Translation between different languages is now possible with high accuracy. This technique of translating after changing the word order is referred to as “pre-ordering”. The prior rearrangement method is described in Patent Document 1 described later.

  In the PBSMT learning, a phrase table is created. The phrase table contains a large number of phrase pairs. A phrase pair is a combination of phrases that are mutually translated in two languages.

  In learning the phrase table, a bilingual corpus including a large number of bilingual translations is used. The bilingual translation is, for example, the sentence pair 30 shown in FIG. 1 and is a combination of sentences in two languages that are translated by each other. The main part of the learning of PBSMT is to extract a correspondence relationship of phrases composed of a chain of words constituting the pair and create a phrase pair.

  In PBSMT learning, the translation accuracy can be improved by inserting a symbol representing the beginning or end of the sentence into both the original and translated sentences of each bilingual corpus and inserting the same symbol into the original sentence at the time of translation. Are known. For example, referring to FIG. 1, in bilingual sentence 32, tags <s> 40 and 44 representing the beginning of the sentence are attached to the beginning of the original sentence and the beginning of the translated sentence, respectively, and tags representing the end of the sentence at the end of each sentence </ s> 42 and 46 are assigned. In this way, the translation performance of PBSMT is improved by using bilingual sentences with tags <s> and </ s> at the beginning and end of the sentence, respectively, and adding the same tags to the original sentence during translation. . This is due to the following reasons.

  In learning of PBSMT, tags attached to the beginning and end of sentences are also processed as one word. As a result, the correspondence between phrases can be performed more accurately. In the above example, when “this” appears at the beginning of the sentence in the original sentence (Japanese), “<s> this” in the original sentence is associated with “<s> This” in the translated sentence (English). When “this” appears in the original sentence (a place other than the beginning of the sentence), the translated sentence “this” is associated. That is, even if phrases (phrase pairs) are paired in the original sentence and the translated sentence, cases where the positions where they appear in the parallel sentence are different can be appropriately distinguished and handled. That is, by attaching a tag as auxiliary information for indicating the position of a word, an effect that phrases can be appropriately associated as a result is obtained.

JP 2013-250605 JP

Taro Watanabe, Kenji Imamura, Hideto Kazawa, GrahamNeubig, Toshiaki Nakazawa, Machine Translation (Natural Language Processing Series 4), Corona, ISBN: 978-4-339-02754-9. Andrew Finch, EiichiroSumita, Dynamic Model Interpolation for Statistical Machine Translation, Proceedingsof the Third Workshop on Statistical Machine Translation, pages 208-215, 2008. Sutskever, I., Vinyals, O., and Le, Q. V. Sequence to sequence learning with neural networks.In Advances in Neural Information Processing Systems (2014), pp. 3104-3112.

  As described above, PBSMT can perform machine translation at high speed and with high accuracy. However, there is still room for improvement in PBSMT. One of the problems of PBSMT is that it is difficult to introduce information beyond the range of phrases into the translation even if tags indicating the beginning and end of the sentence are added. Specific problems are listed below.

  (1) Difficult to translate the original text by grammatical type.

  In the conventional PBSMT, when the grammatical type of the original text is different, it is difficult to accurately translate it. The reasons for this are as follows.

  Consider the case where the noun phrase 60 “result of audit” is translated into English as shown in the upper part of FIG. 2 in Japanese-English translation. In the PBSMT, first, the original text is converted into a word order 62 in the order of words close to English to obtain a word string 64 of “result audit”. A tag addition process 66 is performed on the word string 64 to add a start tag <s> at the beginning of the sentence and an end tag </ s> at the end of the sentence. As a result, a word string 68 is obtained. When the translation 70 by PBSMT is applied to the word string 68, the adverb phrase 72 “As a result of the audit” is obtained as a translation of the noun phrase 60. That is, there is a problem that the result of translating the noun phrase 60 may be an adverb phrase 72 instead of a noun phrase.

  As a similar example, the example shown in the lower part of FIG. 2 can be considered. This example is an example of translating the question sentence 80 “Is the Web server service in operation?” Into English. This question sentence 80 is converted into a word order 82, and a word string 84 “is the Web server service in operation?” Is obtained. By performing the tagging process 86 on this word string, a word string 88 “<s> Web server service is operating? </ S>” is obtained. By applying the translation 90 by PBSMT to the word string 88, a translated sentence 92 that does not have a question sentence and a plain sentence is obtained, "the web server service running?".

  Such a problem occurs for the following reason.

Consider a case where the tagged translation used for learning is a question sentence such as the following in Japanese-English translation.

<s> Is the Web server service running? </ s>
<s> Is the Web server service is running? </ s>

On the other hand, the following translated text may be used as a parallel translation used for learning.

<s> Web server service is running. </ s>
<s> The Web server service is running. </ s>

The difference in notation between the two is negligible.

When the word order is converted, these pairs are as follows.

Is the <s> web server service running? </ s>
<s> Is the Web server service is running? </ s>

<s> web server service is running. </ s>
<s> The Web server service is running. </ s>
The difference in notation between the two is negligible. Therefore, when such parallel translation data is used, appropriate learning cannot be performed with respect to the phrase table. Specifically, the phrase “<s> Web server”, which is the same Japanese phrase, corresponds to “<s> The Web server service is” on one side and “ <s> Is the Web server service ”. For this reason, it is impossible to determine which one should be selected as the translation of “<s> Web server” within the range of this phrase. As a result, the plain text with higher frequency is always used, and the translation of the question text fails.

  A proposal for solving these problems is disclosed in Non-Patent Document 2. Non-Patent Document 2 proposes a model obtained by linearly interpolating a model obtained from a bilingual sentence that is a question sentence and a model created from a bilingual sentence other than the question sentence as a model used in PBSMT.

  On the other hand, as another measure, in order to translate question sentences as question sentences and noun phrases as noun phrases, separate translation engine for question sentences and translation engine for noun phrases Can be considered. A typical example of such a translation apparatus is shown in FIG.

  Referring to FIG. 3, a parallel corpus 110 is prepared. The model learning unit 114 uses the bilingual corpus 110 to learn a plurality of models 112 for translation by grammar type. The model learning unit 114 divides each bilingual sentence of this bilingual corpus into a plurality of partial corpora 130 according to their grammatical types (question sentence, plain text, noun phrase, etc.). The model learning unit 114 further uses the partial corpus 130 to perform learning 132 for PBSMT by a method similar to the conventional method, and constructs a plurality of models 112 for translation. Each of these models 112 includes a phrase table and has a configuration suitable for translation of a specific grammar type. By incorporating these models as models in different machine translation apparatuses, a translation engine suitable for each grammar type can be obtained. For example, by incorporating a model for noun phrases into the machine translation apparatus 120, the machine translation apparatus 120 becomes a dedicated translation engine for translation of noun phrases.

  At the time of translation, an appropriate translation engine is used among a plurality of translation engines for each grammar type according to the grammar type of the input sentence 118. For example, the machine translation device 120 for a noun phrase performs a syntax analysis on a morpheme sequence output from the morpheme analysis unit 140 and a morpheme analysis unit 140 that performs a morphological analysis on the input sentence 118 and performs a pre-ordering on the original sentence. A syntactic analysis unit 142 to perform, a pre-ordering unit 144 that rearranges words in the input sentence 118 so that the word order is close to the English word order using the syntax analysis result by the syntax analysis unit 142, and the input in which the word order is rearranged A tag assigning unit 146 that assigns start tags <s> and end </ s> to the beginning and end of the sentence 118, respectively, and PBSMT is performed on the input sentence 118 with the word order converted and tagged, PBSMT device 148 that outputs 122.

  FIG. 4 is a flowchart showing the control structure of the computer program that implements the tag assignment unit 146 shown in FIG. Referring to FIG. 4, this program is called from the parent routine with an input sentence (word string) as an argument. This program declares a variable STR that stores a character string, concatenates a start tag <s>, an input sentence (word string), and an end tag </ s> in this order and stores them in the variable STR. And step 162 for returning control to the parent routine using STR as a return value.

  By using such a method, for example, a translation engine for a noun phrase can be constructed. For learning, for example, a bilingual corpus consisting only of titles of scientific and technical papers and patent documents can be used, thereby making it possible to create a translation engine dedicated to the titles.

  However, in the above-described method, it is necessary to divide the bilingual corpus 110 for each question sentence type and to build a translation engine. As a result, the amount of parallel translation data used for translation engine learning is reduced. It is already known that the amount of parallel translation data used for learning greatly affects the accuracy of the translation engine, and the translation performance of the translation engine constructed for each grammar type is degraded. Furthermore, since a plurality of translation engines are used, there is a problem that the operation cost becomes high.

  In the conventional technology, there is a problem that not only proper translation for each grammar type but also translation by scene cannot be performed. For example, in English, "Hello" is, at the time of face-to-face may be being interpreted as "Hello," but, at the time of the conversation on the phone, it is necessary to translate as "Hello". In the conventional PBSMT, such a translation cannot be made unless the measures shown in FIG. 3 are taken.

  The conventional technology also has a problem that it cannot be divided by a speaker. For example, in medical translation, it may be necessary for a patient and a nurse to properly translate even the same sentence. For example, when translating “take a drug” into English, if the patient is a speaker, the subject must be “I”, but if the nurse is a speaker, "Must be the subject. In the conventional PBSMT, such a translation by a speaker cannot be performed unless a measure as shown in FIG. 3 is taken.

  Furthermore, the conventional PBSMT has a problem that it cannot be divided according to context. For example, consider the case of translating Japanese “yes” into English. In the context of “Do you like apples?” Or “Yes”, “Yes” translates to “Yes”, and in the context of “Do you like apples?” It is necessary to translate “No”. In conventional PBSTM, it was impossible to make a translation according to this context. Even if the measure as shown in FIG. 3 is taken, it is almost impossible to realize it because there are countless contexts.

  The above-mentioned problems of translation by grammar type, translation by scene, translation by speaker, and translation by context are, after all, a problem that is necessary for accurate translation. This means that there is a shortage of information that exceeds. Although it may be possible to input such information to the machine translation device, it is not preferable to increase the cost of translation by performing complicated processing.

  Further, the above-described problem exists even when a translation method other than PBSMT is adopted. For example, a similar problem exists in machine translation using LSTM (Long Short-Term Memory) (see Non-Patent Document 3).

  Therefore, a machine translation device that appropriately translates and separates the original text according to information in a range exceeding the original text is desired.

  The machine translation device according to the first aspect of the present invention corresponds to meta information specifying means for specifying meta information related to translation, and meta information specified by the meta information specifying means at a predetermined position of the original text of translation. It includes a meta information corresponding tag insertion means for inserting a tag, and a machine translation device that receives the original text with the tag as an input. As the meta information, a plurality of predetermined types are defined. The meta information corresponding tag insertion means selects a tag according to the type of meta information.

  Preferably, the meta information corresponding tag insertion means specifies a first tag corresponding to the meta information and a start position and an end position of the range in order to specify a range in which translation using meta information is performed in the original text. Range specifying tag insertion means for inserting each second tag is included.

  Preferably, the meta information specifying unit includes a morpheme analyzing unit for performing morphological analysis of the original sentence, a syntax analyzing unit for performing syntax analysis of the original sentence analyzed by the morpheme analyzing unit, and a syntax analysis of the original sentence by the syntax analyzing unit. Grammar type output means for outputting information indicating the grammatical type of the original sentence obtained as a result as meta information of the original sentence.

  More preferably, meta information relating to the original text is attached to the original text. The meta information specifying means includes meta information separating means for separating the meta information attached to the original text from the original text and giving it to the meta information corresponding tag inserting means.

  More preferably, the meta information is a grammatical type of the original sentence, scene information about a scene in which the original sentence is uttered, speaker information about a speaker who utters the original sentence, and a sentence preceded by the machine translation means. It is selected from the group consisting of the grammatical type of the original text and the linguistic information specifying the language of the translation destination.

  More preferably, the machine translation means is a phrase-based machine translation means.

  More preferably, the meta information specifying unit includes a unit for specifying a translation destination language of the original text of the translation as meta information, and the meta information corresponding tag insertion unit includes a tag indicating the translation language specified by the meta information in the original text. Means for insertion at a predetermined position.

  A computer program according to the second aspect of the present invention causes a computer to function as any of the machine translation devices described above.

It is a schematic diagram which shows tag provision in the conventional PBSMT. It is a schematic diagram which shows the production | generation process of the mistranslation by the conventional PBSMT. It is a block diagram which shows the method of constructing | translating a translation engine according to grammar type by the conventional PBSMT. It is a flowchart which shows the general | schematic control structure of the program which assign | provides a start tag and an end tag to the sentence head and the sentence end of the input sentence at the time of translation of the conventional PBSMT. It is a figure which shows the translation process in PBSMT which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the functional structure of the PBSMT system which concerns on the 1st Embodiment of this invention, and the learning process part of the said PBSMT system. It is a flowchart which shows the outline of the control structure of the program which provides a tag to the head of the input sentence at the time of translation, and the sentence end in the PBSMT system which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the functional structure of the PBSMT system which concerns on the 2nd Embodiment of this invention, and the learning process part of the said PBSMT system. It is a block diagram which shows the functional structure of the PBSMT system which concerns on the 3rd Embodiment of this invention, and the learning process part of the said PBSMT system. It is a block diagram which shows the functional structure of the translation system which concerns on the 4th Embodiment of this invention, and the learning process part of the said translation system. It is a flowchart which shows the control structure of the program which implement | achieves the tag provision part of the learning process part shown in FIG. It is a schematic diagram for demonstrating the learning process of NN by the neural network (NN) learning part shown in FIG. It is a figure which shows the external appearance of the computer which implement | achieves the translation system which concerns on each embodiment of this invention, and the learning process part of the said translation system. FIG. 14 is a block diagram showing a hardware configuration of a computer whose appearance is shown in FIG. 13.

  In the following description and drawings, the same parts are denoted by the same reference numerals. Therefore, detailed description thereof will not be repeated.

[basic way of thinking]
In each embodiment described below, meta information that exceeds the range of a phrase is given to the original text, so that the meta information is appropriately translated by referring to the meta information at the time of translation. As meta information, tags attached to the original text are used in the following embodiments. Prepare multiple types of tags, according to the grammatical type of the original text (first embodiment), according to a scene or speaker (second embodiment), according to the previous context (third embodiment), or translated Appropriate translation can be done by attaching different tags to the original text according to the previous language (fourth embodiment). In learning, it is necessary to perform the same tagging and learn a model for translation including a phrase table.

  In the following first to third embodiments, after performing pre-ordering on the input sentence, tags representing meta information are attached to the beginning and end of the input sentence. Prior rearrangement means that prior to translation, the word order of the original text is converted into a word order that is close to the word order of the language of the translation destination. It is known that the translation accuracy is improved in the statistical translation device by the prior rearrangement (pages 155 to 159 of Non-Patent Document 1). However, the present invention is not limited to such an embodiment. For example, the above-described meta information can be used in PBSMT that does not perform pre-ordering. Moreover, the greatest effect can be obtained by applying meta information to PBSMT. However, since meta information is used to construct a language model, it can be applied to general statistical translation devices other than PBSMT. It is considered effective. The fourth embodiment is a system that performs Sequence-to-Sequence type translation using LSTM, which is a kind of so-called deep neural network (DNN), instead of using PBSMT. In the following embodiments, the grammatical type of the input sentence, the information about the speaker or the other party, the information about the scene, the information about the context, the information specifying the language of the translation destination, etc. are used as the meta information. However, the meta information is not limited to these, and any information that is useful for translation may be used.

  In addition, as a method for prior rearrangement, Non-Patent Document 1 automatically includes a method for manually creating a rearrangement rule, a method for learning a rearrangement model from a corpus, and a parser for rearrangement itself. The method of learning is introduced. In each of the first to third embodiments described below, any of these methods may be used. Also, in the following first to third embodiments, pre-ordering is performed in all cases, but it is expected that the translation accuracy will be improved compared to translation that does not use meta-information without performing pre-ordering. it can.

[First Embodiment]
The PBSMT system according to the first embodiment is a device that performs PBMST, and uses a plurality of types of tags to represent the grammatical type of input as meta information. If the original text of the translation is a noun phrase during learning, PBSMT learns by adding a start tag <NP> to the beginning of the word string after pre-sorting and an end tag </ NP> at the end of the sentence. I do. If the original text of the parallel translation is a question sentence, learning is performed with the start tag <SQ> at the beginning of the word string after the pre-sorting and the end tag </ SQ> at the end of the sentence. At the time of translation, PBSMT is applied to the input sentence that has been rearranged in advance, with tags according to the grammatical type obtained as a result of the syntax analysis being added in the same manner as in the learning.

  For example, referring to the upper part of FIG. 5, consider a case where the input to the translation is a noun phrase 60 “result of audit”. By performing the word order conversion 62, a word string 64 of “result audit” is obtained. The grammatical type tagging process 180 described above is performed on the word string 64. If the tag for the noun phrase is <NP> here, the word string 182 "<NP> Audit result </ NP>" is obtained. By applying the translation 184 by PBSMT to this word string 182, a word string 186 of “results of the audit” is obtained as a translation result.

  A similar example is shown in the lower part of FIG. Consider a case where the input is a question sentence 80 "A Web server service is in operation?" The word string 84 is obtained by applying the word order conversion 82 to the question sentence 80. A tag assignment process 190 for each grammar type is performed on this word string. Here, <SQ> is used as a tag corresponding to the question sentence. As a result, a word string 192 with the tag <SQ> at the beginning of the sentence and the tag </ SQ> at the end of the sentence is obtained. A translation 196 “Is the web server service running?” Is obtained by performing translation 194 using PBSMT on the word string 192.

  In the case of performing pre-ordering, it is necessary to select nodes whose positions should be exchanged in the original sentence parsing tree and exchange their positions. Therefore, when using pre-ordering, the original text is analyzed in advance. The textual grammar type is obtained as a by-product of parsing. In the following embodiment, this grammar type is used to determine the tag type.

<Constitution>
Referring to FIG. 6, the PBSMT system 210 according to this embodiment uses the bilingual data included in the bilingual corpus 220 as learning data and performs tag assignment for each grammar type described above, thereby performing translation including a phrase table. When a learning processing unit 224 for learning a statistical model for output and an input sentence 226 is given to the model storage unit 222, a model for translation stored in the model storage unit 222 is used. A machine translation device 230 for performing PBSMT and outputting a translated sentence 228.

  The learning processing unit 224 using the grammar type reads the parallel translation included in the bilingual corpus 220, separates each parallel translation into an original sentence and a translated sentence, and outputs the bilingual sentence reading section 250. For the original sentence of the parallel translation sentence, the grammatical type of the original sentence is specified, the original sentence processing unit 252 that performs tagging according to the grammar type, and the translation sentence of each parallel translation sentence output by the parallel translation sentence reading unit 250, A translation processing unit 254 that outputs a tag by adding a tag in the same manner as the conventional method, a source text that is output by the source processing unit 252 and that is tagged according to the grammar, and a translation processing unit 254 outputs A learning data storage unit 256 that stores bilingual data paired with translations tagged in the same manner as the conventional learning data as model learning data, and learning data stored in the learning data storage unit 256 are used. Learns the statistical model for translation in a conventional manner and a model learning unit 258 for storing in the model storage unit 222. Although the function itself of the model learning unit 258 is the same as the conventional one, the model stored in the model storage unit 222 is different from the conventional one because the original text is tagged according to the grammatical type.

  The source sentence processing unit 252 performs morpheme analysis on the source sentence given from the parallel translation reading unit 250, performs parse analysis on the morpheme analysis unit 260 that outputs a morpheme sequence, and the morpheme sequence output by the morpheme analysis unit 260. At the same time, the grammar type is determined, and the grammar type determination unit 262 that outputs the syntax analysis result and the grammar type separately, and the syntax analysis result output by the grammar type determination unit 262 is used, and is included in the input original sentence. A pre-ordering unit 264 that outputs the word sequence of the word string by rearranging the words in the order close to the word order of the translation destination language prior to the translation, and the head of the word sequence converted by the pre-ordering unit 264 and converted. And a grammatical type tag adding unit 266 that outputs a word string with a start tag and an end tag corresponding to the grammar type received from the grammar type determination unit 262 to the learning data storage unit 256 at the end of the sentence. Including the.

  An example of a control structure of a computer program that realizes the grammatical type tag assignment unit 266 shown in FIG. 6 is shown in a flowchart form in FIG. Referring to FIG. 7, this program is called from a parent routine that receives an input word string and a grammar type as arguments, and a start tag corresponding to the grammar type is used at the beginning of the sentence, and an end tag corresponding to the grammar type is used at the end of the sentence. The attached word string is returned as a return value. This program includes a step 160 for declaring a character string variable STR used for character string operation, a step 300 for selecting a start tag and an end tag corresponding to the grammar type received as an argument, and a variable STR as a variable STR in step 300. Step 302 includes concatenating the selected start tag, the input word string received as an argument, and the end tag selected in Step 300, and ending this routine with STR as a return value.

  In order to select the start tag and the end tag according to the grammar type in step 300, the grammar type and the tag may be described in association with each other in this routine, or a separate lookup table is stored in the memory. The start tag and the end tag may be read from the lookup table using the grammar type as a key.

  The translated sentence processing unit 254 is the same as the translated sentence processing unit used in the conventional PBSMT. The translated sentence processing unit 254 receives the translated sentence from the parallel translated sentence reading unit 250 and uses the translated sentence processing unit 254 at the beginning and end of the word string constituting the translated sentence. The tag addition part 274 which attaches | subjects the same start tag and end tag, and outputs it to the learning data storage part 256 is included.

  The machine translation device 230 performs parse analysis on the morpheme analysis unit 280 that performs morpheme analysis on the input sentence 226 and the morpheme sequence output by the morpheme analysis unit 280, and the grammar of the input sentence 226 is obtained from the result of the parse analysis result. A grammar type determination unit 282 that determines the type and outputs the parse result and the grammar type of the input sentence 226. The grammar type determination unit 282 receives the parse result from the grammar type determination unit 282, and the word order close to the word order of the translation language in advance of translation. A pre-ordering unit 284 that outputs a word string obtained by rearranging the words to the word, a word string output from the pre-sorting unit 284, and a grammar type output from the grammar type determining unit 282, A grammatical-type tag adding unit 286 that outputs a start tag corresponding to the grammar type at the beginning of a word string output by the H.264 and an end tag corresponding to the same grammar type at the end of the sentence. , And a PBSMT device 288 grammar Type tag attaching portion 286 outputs the tag to output a translation 228 running PBSMT as input word string attached.

  The grammatical type tag assignment unit 286 has the same function as the grammatical type tag assignment unit 266, and has the same configuration as the grammar type tag assignment unit 266 in this embodiment.

<Operation>
The PBSMT system 210 having the configuration shown in FIGS. 6 and 7 operates as follows. There are roughly two operation phases of the PBSMT system 210. The first is a learning phase of the model storage unit 222, and the second is a test or translation phase by the machine translation device 230. In model learning, there is a method of directly learning a model from learning data, and a method of optimizing feature weights given to the model after learning the model from learning data. This embodiment is effective for any method.

  A large number of parallel translation sentences are stored in the parallel corpus 220 in advance. It is assumed that the phrase translations prepared here have already been phrase-aligned.

  The parallel translation reading unit 250 sequentially reads each parallel translation from the parallel corpus 220, gives the original sentence to the morpheme analysis unit 260 of the original sentence processing unit 252, and gives the translated sentence to the tag adding unit 274 of the translation sentence processing unit 254.

  The morpheme analysis unit 260 performs a morpheme analysis on the original sentence given from the parallel translation reading unit 250 and outputs a morpheme string. The grammar type determination unit 262 performs syntax analysis on the morpheme sequence output by the morpheme analysis unit 260, simultaneously determines the grammar type, and outputs the syntax analysis result and the grammar type separately. The prior rearrangement unit 264 uses the parsing result output by the grammar type determination unit 262 to rearrange the word order of the word string included in the input original sentence in the order of words closer to the word order of the translation destination language prior to translation. Change the output. The grammatical type tag assigning unit 266 outputs a start tag and an end tag corresponding to the grammatical type received from the grammar type determining unit 262 at the beginning and end of the word string converted from the word order output from the pre-ordering unit 264. The attached word string is output to the learning data storage unit 256.

  The tag assignment unit 274 of the translated sentence processing unit 254 attaches the same start tag and end tag to the beginning and end of the word string constituting the translated sentence of the parallel translation sentence, and outputs the result to the learning data storage unit 256.

  The learning data storage unit 256 outputs the original sentence to which the tag for each grammar type is output, which is output from the tag assignment unit 266 for each grammar type, and the translated sentence to which the same tag is output, which is output from the tag addition unit 274. Remember in pairs. The model learning unit 258 performs model learning using the learning data stored in the learning data storage unit 256 and stores the parameters in the model storage unit 222.

  At the time of translation, the machine translation device 230 operates as follows.

  The model storage unit 222 that stores the learned model is connected to the machine translation device 230 so that the machine translation device 230 can refer to it. This connection may be realized by storing a model in a hard disk of a computer that implements the machine translation device 230 and then loading the model into a memory so that the model can be read from the CPU of the computer, or by a network A computer may be connected to the model storage unit 222, and the model may be stored in an internal storage device of the computer via a network.

  In response to the input sentence 226 being given, the morpheme analysis unit 280 performs a morpheme analysis on the input sentence 226 and gives the obtained morpheme sequence to the grammar type determination unit 282. The morpheme analysis process of the morpheme analysis unit 280 may be triggered by the input of a specific code after the input sentence 226 is input, or the start of translation is instructed independently of the input of the input sentence 226 May be started in response to receiving an input from the user.

  The grammatical type determination unit 282 performs syntax analysis similar to the grammatical type determination unit 262 on the morpheme sequence output by the morpheme analysis unit 280, determines the grammatical type of the input sentence 226 using the result, and performs the syntax analysis result. Is provided to the pre-sorting unit 284, and the grammatical type is provided to the grammatical type tag assigning unit 286.

  The prior rearrangement unit 284 translates the word order of the words constituting the input sentence 226 so that the parsing result of the input sentence 226 given by the grammar type determination unit 282 is closer to the word order of the translation target language. Is converted and provided to the grammatical type tag assigning unit 286.

  The grammatical type tag assigning unit 286 attaches a start tag corresponding to the grammatical type received from the grammar type determining unit 282 to the beginning of the word string after the word order conversion received from the pre-ordering unit 284, and ends the sentence. Similarly, an end tag corresponding to the grammar type received from the grammar type determination unit 282 is attached. The grammatical type tag assigning unit 286 gives the word string tagged with the grammatical type tag to the PBSMT device 288 as a translation original.

  The PBSMT device 288 performs PBSMT while referring to the model stored in the model storage unit 222 using the word string given from the grammar type tag assigning unit 286 as the original text of translation, and outputs the translated text 228.

<Effects of the present embodiment>
According to the PBSMT system 210 according to the first embodiment, tags that differ depending on the grammar type are attached to the beginning and end of a sentence. In PBSMT, these tags are also considered as words constituting a phrase. Therefore, even when the phrase is the same, it can be distinguished from the case of being at the beginning of the sentence and the case of being within the sentence. Further, since the affirmative sentence and the question sentence can be distinguished from each other by the tag, the phrase pair obtained from the affirmative sentence and the phrase pair obtained from the question sentence include different tags. Therefore, the phrase table can be accurately learned. As a result, translation accuracy is improved. In addition, in this case, there is no need to change the configuration of the PBSMT device itself. Therefore, the accuracy of machine translation can be improved with a simple configuration.

  In the first embodiment described above, syntax analysis is necessary to perform pre-ordering, and a grammar type obtained as a result of syntax analysis is used for tag determination. However, the present invention is not limited to such an embodiment. If pre-sorting is not performed, a classifier that can determine the grammatical type of the original text may be separately constructed by machine learning, and the classifier may be used.

[Second Embodiment]
In the first embodiment, different tags depending on the grammar type are given to the original text as meta information. Therefore, in the first embodiment, a grammar type obtained from the result of syntax analysis performed on the original text at the time of learning and translation is used. However, the present invention is not limited to such an embodiment. For example, a tag representing meta information may be added to the original text in advance. The second embodiment relates to such a translation system. Also in this embodiment, PBSMT is used as a machine translation system.

<Constitution>
FIG. 8 shows a functional configuration of the PBSMT system 320 according to the second embodiment. Referring to FIG. 8, the PBSMT system 320 learns a model for PBSMT using a bilingual corpus 240 with meta information that includes bilingual sentences with meta information, and stores model parameters in the model storage unit 342. A learning processing unit 340 using meta information to be stored, a machine translation device 348 that performs machine translation on the input sentence 344 with meta information using the model parameters stored in the model storage unit 342, and outputs a translation sentence 346; including.

  The learning processing unit 340 receives the translated text reading unit 250, receives the translated text from the translated text reading unit 250, attaches a tag according to the meta information, and outputs it, and the translated text reading unit 250. 6 that receives the translation of the bilingual sentence, attaches and outputs the same tag as in the past, and the original sentence with the tag corresponding to the meta information output from the original sentence processing unit 360 and the same translation sentence processing unit 254 as in FIG. A learning data storage unit 362 that stores the same word sequence and a translation sentence that is output from the translated sentence processing unit 254 and that has the same tag as the conventional one, and learning stored in the learning data storage unit 362 A model learning unit 364 for performing model learning for the PBSMT using the data and storing the model parameters in the model storage unit 342;

  The source text processing unit 360 includes a morphological analysis unit 260 similar to that of the first embodiment, a syntax analysis unit 372 that performs parsing processing similar to the grammar type determination unit 262 of the first embodiment, and a pre-ordering 264, meta information separation unit 370 that separates the meta information from the original text received from the parallel translation reading unit 250, and meta information separation at the beginning and end of the word string after the pre-ordering received from the pre-ordering unit 264 A tag adding unit 374 that adds a tag corresponding to the meta information given from the unit 370 and outputs the tag to the learning data storage unit 362.

  Meta information may include information about the speaker, the gender of the speaker, information indicating the age or occupation of the speaker, the other party, the gender of the other party, information indicating the age or occupation of the other party, and information indicating the relationship between the speaker and the other party. It is done. As information indicating a scene, for example, face-to-face / phone / TV conference can be considered. By assigning meta information to each bilingual sentence stored in the bilingual corpus 240 with meta information in advance, a statistical model for a word string including meta information can be learned.

  The machine translation device 348 includes a morpheme analysis unit 280 that receives a word string and a syntax analysis unit that performs syntax analysis on a morpheme string output by the morpheme analysis unit 280 in the input sentence 344 with meta information. 382, a pre-ordering unit 284 for rearranging the word string constituting the input sentence using the syntax analysis result by the syntax analysis unit 382 close to the order of the word string in the translation target language, and the input sentence with meta information 344 Meta information separating unit 380 that separates meta information, and a tag of a type corresponding to the meta information output from the meta information separating unit 380 at the beginning and end of a pre-ordered word string given from the pre-ordering unit 284 And a tag string for each meta information output by the meta information tag adding unit 384, and a word string to which the meta information tag is output and stored in the model storage unit 342. And a PBSMT apparatus 288 to output a translation 346 performs PBSMT using the model for a machine translation based on the Dell parameters.

<Operation>
In the first embodiment shown in FIG. 6, at the time of learning, a tag for each grammar type is assigned to a word string using the grammar type determined by the grammar type determination unit 262. In the second embodiment, unlike the first embodiment, during learning, the meta information separation unit 370 separates the meta information from the bilingual sentence to which the meta information is added in advance, and the tag addition unit 374 Different tags are assigned to word strings depending on information. Model learning for machine translation using meta information can be efficiently performed by determining in advance what to use as meta information and assigning the meta information to a parallel translation for learning.

  The same applies to translation, and meta information is given to the input sentence 344. The meta information separation unit 380 separates the meta information and gives it to the meta information tag adding unit 384. The meta information tag assigning unit 384 assigns different tags to the word string of the original text according to the meta information and inputs them to the PBSMT device 288. By giving the input sentence 344 the type of meta information used at the time of learning, an appropriate translation sentence 346 corresponding to the meta information can be obtained.

  Similar to the grammatical type obtained as a result of parsing, when using meta information obtained by analyzing the original sentence, meta information is attached to the bilingual sentence in the parallel corpus 220 at the time of learning and the input sentence 344 at the time of translation. There is no need to keep it.

[Third Embodiment]
In the first embodiment, a tag is selected based on grammar type information determined from the result of parsing the original sentence. In the second embodiment, a tag is selected based on meta information given in advance to the original text or meta information obtained by analyzing the original text. In the third embodiment described below, the grammar type of the previous sentence is stored as context information as information corresponding to the meta information, and different tags are assigned to the original sentence according to the context information. . This mechanism makes it possible to translate the original text according to the context.

<Constitution>
Referring to FIG. 9, PBSMT system 400 according to the third embodiment learns a model for machine translation using a bilingual sentence in bilingual corpus 220, and stores model parameters and the like in a model storage unit. A translation processing 414 is output by performing PBSMT on the learning processing unit 412 to be stored in 410 and the input sentence 226 using a model for translation composed of model parameters and the like stored in the model storage unit 410. Machine translation device 416.

  The learning processing unit 412 receives the parallel translation reading unit 250 as in FIG. 6 and the parallel translation original from the parallel translation reading unit 250, and sets different start tags and end tags depending on the context of the sentence at the beginning and end of the original sentence. A translation processing unit 254 and a source text processing unit 440, which are the same as those in FIG. A learning data storage for storing the word string of the original sentence to which the tag is added and the translation sentence to which the translation sentence processing unit 254 outputs the same tag as that of the conventional tag and corresponding to each other and storing them as learning data Model learning for performing learning of a model for PBSMT using the learning data stored in the learning unit 442 and the learning data storage unit 442 and storing the model parameters and the like in the model storage unit 410 And a 444.

  The original sentence processing unit 440 indicates whether or not the original sentence being processed is a negative question sentence based on the result of the parsing by the morphological analysis unit 260, the syntax analysis unit 372, the pre-ordering unit 264, and the syntax analysis unit 372. Context information storage unit 450 that stores information as context information, and the context information stored in context information storage unit 450 is shifted and stored after processing of one sentence, and is output as context information obtained from the preceding original sentence For the word string of the original sentence after the pre-sorting output from the context information storage unit 452 and the pre-sorting unit 264, different tags are added according to the context information of one sentence before stored in the one-sentence context information storage unit 452 A tag assigning unit 454 for giving and outputting to the learning data storage unit 442.

  The machine translation device 416 indicates whether or not the input sentence 226 obtained from the output of the morphological analysis unit 280, the syntax analysis unit 382, the pre-rearrangement unit 284, and the syntax analysis unit 382 shown in FIG. Context information storage unit 470 for storing context information, and every time the machine translation device 416 processes one sentence, the context information stored in the context information storage unit 470 is shifted and stored as the previous context information. The previous sentence context stored in the previous sentence context information storage unit 472 at the beginning and end of the word string of the input sentence 226 rearranged by the previous rearrangement unit 284 A tag assignment unit 474 that assigns and outputs different tags according to information, and a tagged word string output by the tag assignment unit 474 as input, and a model parameter of translation stored in the model storage unit 410 Performing PBSMT with reference to chromatography data or the like and a PBSMT device 288 for outputting the translated sentence 414.

<Operation>
The PBSMT system 400 operates as follows.

  When learning the model, the bilingual sentence reading unit 250 takes out the bilingual sentences one by one from the bilingual corpus 220, and the original sentence is sent to the morpheme analyzing unit 260 of the original sentence processing unit 440, and the translated sentence is sent to the tag adding unit 274 of the translated sentence processing unit 254 Give each. In this embodiment, different tags are attached to the original text depending on the context. Accordingly, the bilingual sentences stored in the bilingual corpus 220 are ordered, and the bilingual sentence reading unit 250 must read the bilingual sentences from the bilingual corpus 220 in the order.

  The morpheme analysis unit 260 and the syntax analysis unit 372 perform morpheme analysis and syntax analysis on the original sentence, respectively, and the syntax analysis result is given to the pre-ordering unit 264. The syntax analysis unit 372 outputs context information indicating whether or not the sentence is a negative question sentence from the result of the syntax analysis. The context information storage unit 450 stores this context information. The pre-rearrangement unit 264 performs pre-rearrangement on the original word string using the result of the syntax analysis by the syntax analysis unit 372, and gives the rearranged word string to the tag adding unit 454. The tag assignment unit 454 assigns, to the word string output from the pre-rearrangement unit 264, a start tag and an end tag that differ depending on the previous sentence context information stored in the previous sentence context information storage unit 452, at the beginning and end of the sentence, respectively. To the tag assignment unit 454. When the first sentence is processed, nothing is stored in the previous sentence context information storage unit 452, and therefore it is assumed that the previous sentence was a plain text.

  When performing such processing, when processing a sentence extracted from a completely different document, the context information obtained from the last sentence of the preceding document is used as the context information of the first sentence of the next document. Is not preferred. Therefore, it is necessary to clear the context information stored in the previous sentence context information storage unit 452 every time the document changes.

  The translated sentence processing unit 254 adds the same tag as the conventional tag to the translated sentence and outputs it. The learning data storage unit 442 outputs the word sequence of the original sentence after the pre-rearrangement with the context information of the previous sentence output from the tag addition unit 454 and the word string of the translation output from the tag addition unit 274. Store them in association with each other.

  When the learning data becomes available in the learning data storage unit 442, the model learning unit 444 starts learning a model for translation using the learning data. The model parameters of the learned model are stored in the model storage unit 410.

  When translating the input sentence 226, the machine translation device 416 operates as follows. Since the context information is also used during translation, the input sentence 226 given to the machine translation device 416 must be given to the machine translation device 416 according to the order in which the sentences appear in the document.

  When the input sentence 226 is given, the morpheme analysis unit 280 performs morpheme analysis and gives the obtained morpheme sequence to the syntax analysis unit 382. The syntax analysis unit 382 performs syntax analysis on the morpheme string and outputs the syntax analysis result to the pre-ordering unit 284. This parsing result includes information indicating whether or not the sentence is a negative question sentence. The context information storage unit 470 stores this information. The prior rearrangement unit 284 rearranges the order of the word strings constituting the input sentence 226 using the syntax analysis result given from the syntax analysis unit 382 prior to translation so as to be close to the order of the word strings in the translation target language. It replaces and gives to the tag provision part 474. FIG. The tag assigning unit 474 reads the previous context information stored in the previous context information storage unit 472 and assigns different start tags and end tags to the beginning and end of the input word string according to the context information. And output.

  The PBSMT device 288 performs PBSMT by applying a model for translation composed of model parameters and the like stored in the model storage unit 410 to the tagged word string output from the tag addition unit 474, and translates the translated sentence. 414 is output. When the output of the translated sentence 414 is completed, the context information stored in the context information storage unit 470 is shifted to the previous sentence context information storage unit 472 and can be used as the previous sentence context information.

<Effects of the present embodiment>
According to the present embodiment, in the translation phase, context information indicating whether or not the previous sentence is a negative question sentence is stored in the previous sentence context information storage unit 472. By assigning a tag corresponding to the context information to the word string and using it as an input to the PBSMT device 288, it is possible to appropriately translate depending on the context, such as when the previous sentence is a negative question sentence or not. It becomes possible.

  In the third embodiment, only whether or not the previous sentence is a negative question sentence is used as context information. However, the present invention is not limited to such an embodiment. A series of sentences existing before a certain sentence may be grouped together, and a tag corresponding to the class may be assigned to the subsequent sentence. As a class, affirmation / denial, question / phrase, etc. can be used, or a combination of these can be used.

  In each of the above-described embodiments, PBSMT is used as the machine translation engine. However, the present invention is not limited to such an embodiment. Even with other machine translation systems, the same effects as those of the above embodiments can be obtained as long as the machine translation system uses a model that learns by statistically processing parallel translations.

  In the first to third embodiments, tags representing meta information are attached to the beginning and end of a sentence. However, the present invention is not limited to an embodiment in which meta information is attached to such a position. In short, it is only necessary to add a tag so that a part to be translated can be specified by meta information. In that case, the portion that needs to be translated corresponds to the original text. The fourth embodiment described below corresponds to such an example.

[Fourth Embodiment]
<Constitution>
As described in the fourth embodiment below, when a next tag is encountered after a certain tag is attached, it is considered that the range in which the meta information by the previous tag affects has ended. In that case, the end tag of the meta information can be omitted. Further, when the end of the sentence to be translated is reached, the end tag can be similarly omitted by interpreting that the range affected by the meta information is over.

  In the first embodiment, a tag is selected based on grammar type information determined from the result of parsing the original sentence. In the second embodiment, a tag is selected based on meta information given in advance to the original text or meta information obtained by analyzing the original text. In the third embodiment, the grammar type of the previous sentence is stored as context information as information corresponding to the meta information, and different tags are assigned to the original sentence according to the context information. In the fourth embodiment described below, a tag for specifying the language of the translation destination is added to the head of the original text as meta information. When learning the model, by adding meta information that specifies the other language of the parallel translation to one of the parallel translations, and by adding meta information that specifies the language of the translation destination at the beginning of the input source text at the time of translation, Translation into a plurality of languages can be performed using one model.

  Referring to FIG. 10, translation system 500 according to the present embodiment reads a multilingual parallel corpus 510 including parallel translations related to a combination of many languages, and reads each parallel translation from multilingual parallel corpus 510, and sequence-to- A learning processing unit 512 that performs learning of an NN that performs sequence-type translation and an NN parameter storage unit 514 that stores parameters of the NN learned by the learning processing unit 512 are included. In the present embodiment, it is assumed that each bilingual translation stored in the multilingual parallel corpus 510 has no information indicating its language. The NN used in the present embodiment has the same configuration as that using the LSTM described in Non-Patent Document 3.

  The learning processing unit 512 indicates the language of the second sentence at the beginning of the first sentence among the parallel sentence reading part 540 that reads each parallel translation from the multilingual parallel corpus 510 and the parallel sentence read by the parallel sentence reading part 540. A first sentence processing unit 542 that outputs a tag, and a second sentence that is output by adding a tag indicating the language of the first sentence to the beginning of the second sentence among the parallel translations read by the parallel sentence reading unit 540. A sentence processing unit 544; a learning data generation unit 546 that generates and outputs learning data by pairing the first sentence output from the first sentence processing unit 542 and the second sentence output from the second sentence processing unit 544; A learning data storage unit 548 that stores the output learning data and an NN learning unit 550 that performs learning of the NN 552 using each parallel translation data stored in the learning data storage unit 548 are included.

  In this embodiment, each of the translations stored in the multilingual parallel corpus 510 is a pair of a sentence in one language and a sentence in another language. However, the present invention is not limited to such an embodiment. For example, it may be a corpus in which translation groups that are translated into each other in three or more languages are collected. In such a case, for example, the bilingual sentence reading unit 540 may select any two sentences from the translation group as bilingual translations and give them to the first sentence processing unit 542 and the second sentence processing unit 544. Therefore, there are at least two types of tags used in the present embodiment, and there are as many as the number of parallel sentence languages stored in the multilingual parallel corpus 510.

  The first sentence processing unit 542 includes a language identifying unit 580 that identifies the language of the first sentence in the parallel translation read by the parallel translation reading unit 540 and outputs information specifying the language. Similarly, the second sentence processing unit 544 includes a language identifying unit 590 that identifies the language of the second sentence and outputs information specifying the language. The first sentence processing unit 542 further includes a tag adding unit 582 that adds a tag indicating the language of the second sentence output by the language identifying unit 590 to the first sentence of the parallel translation read by the parallel sentence reading unit 540 and outputs the tag. Including. Similarly, the second sentence processing unit 544 attaches a tag indicating the language of the first sentence output by the language identifying unit 580 to the second sentence of the parallel translation read by the parallel sentence reading unit 540 and outputs the tag. Further included.

  The tag assignment unit 582 and the tag assignment unit 592 illustrated in FIG. 10 have the same configuration, and both are realized by a computer program in this embodiment. Referring to FIG. 11, for example, a program that implements tag assigning unit 582 declares variable STR from step 630, and reads the first sentence of the parallel translation to be processed from the memory location where the output of parallel sentence reading unit 540 is stored. Step 632 for reading, Step 634 for reading the language tag indicating the language of the second sentence from the memory location where the output of the language identification unit 590 is stored, the language tag read in Step 634, and the first read in Step 632 A step 636 for substituting a character string obtained by combining a sentence and a symbol <EOS> indicating the end of the sentence into a variable STR, and a step 638 for outputting the stored contents of the variable STR to the learning data generation unit 546 are included.

  In the program that implements the tag assignment unit 592, the first sentence may be replaced with the second sentence and the second sentence may be replaced with the first sentence in FIG. 11.

  Referring to FIG. 10, learning data generation unit 546 generates learning data from the output of first sentence processing unit 542 and the output of second sentence processing unit 544. Specifically, the learning data generation unit 546 has learning data in which the output of the first sentence processing unit 542 is the original text, the output of the second sentence processing unit 544 is the translated sentence, and the output of the second sentence processing unit 544 is the original text. Learning data with the output of the single sentence processing unit 542 as a translation is generated and stored in the learning data storage unit 548. The learning data generation unit 546 can prepare bi-directional learning data related to the combination of the first sentence and the second sentence for the translation to be processed.

  The NN learning unit 550 has a function of performing learning of the NN 552 using learning data stored in the learning data storage unit 548. This learning can be performed in the same manner as the technique described in Non-Patent Document 3.

  Specifically, the outline of learning described in Non-Patent Document 3 is as follows. It is assumed that the original text of the parallel translation for learning includes words A, B, and C, and the translated text includes W, X, Y, and Z. All of these end with a sentence ending symbol <EOS>. Referring to FIG. 12, for example, first, words A, B, and C of an input sentence are input to NN in order, and NN is learned by back propagation using these as teacher signals. For the symbol <EOS> indicating the end of the input sentence, NN learning is performed using the beginning W of the word of the translated sentence as a teacher signal. In the following, NN learning is performed using words X, Y, and Z of the translated sentence as inputs, and using the next word Y, Z and the symbol <EOS> indicating the end of the translated sentence as a teacher signal. This process is performed for all parallel translations.

  The learning performed by the NN learning unit 550 on the NN 552 in the present embodiment is exactly the same. The only difference is that a tag indicating the language of the other party of the translation is attached to the head of each sentence to be input.

  Returning to FIG. 10, the machine translation device 518 using the NN performs machine translation using the NN on the input sentence 516 and outputs the translation sentence 520. The machine translation device 518 has an input / output device 600 that accepts user input in an interactive manner and a tag that indicates a translation destination language input by the input / output device 600 in order to determine a translation destination language. A tag that is connected to the storage unit 602 and the target language storage unit 602, receives the input sentence 516, reads a tag read from the target language storage unit 602 at the top, and adds a sentence ending symbol <EOS> to the end and outputs the tag The translation unit 604 and the input sentence 516 with a tag output from the tag addition unit 604 are translated by an NN having a parameter stored in the NN parameter storage unit 514 and having the same configuration as that of the NN 552. NN translation engine 606 for output.

  In this embodiment, the user designates a language to be translated using the input / output device 600, but the present invention is not limited to such an embodiment. For example, the language selected as the language of the user interface set in a device (smart phone, computer, etc.) in which the machine translation device 518 is incorporated may be used.

<Operation>
The translation system 500 whose configuration has been described above operates as follows. There are two phases in the operation of translation system 500. The first is an NN learning phase, and the second is a translation phase by the machine translation device 518.

  During learning, the learning processing unit 512 of the translation system 500 operates as follows. A multilingual parallel corpus 510 stores a large number of parallel translations related to a combination of a large number of languages. The parallel translation reading unit 540 sequentially extracts the parallel translations from the multilingual parallel corpus 510 and provides the first sentence of each parallel translation to the first sentence processing unit 542 and the second sentence to the second sentence processing unit 544. The language identifying unit 580 of the first sentence processing unit 542 identifies the language of the first sentence, and stores a tag indicating the language in a predetermined memory location. Similarly, the language identifying unit 590 of the second sentence processing unit 544 identifies the language of the second sentence and stores the tag indicating the language in a predetermined memory location. When the first sentence of the parallel translation is received, the tag of the language identified by the language identification unit 590 is read from a predetermined memory location, and the tag is added to the beginning of the first sentence and the sentence end symbol <EOS> is attached to the end. To the learning data generation unit 546. Similarly, when the tag adding unit 592 receives the second sentence of the parallel translation from the parallel sentence reading unit 540, the tag indicating the language of the first sentence is read from a predetermined memory location, and the tag is added to the end of the second sentence. Is appended with a sentence end symbol <EOS>, and is given to the learning data generation unit 546.

  The learning data generation unit 546 uses the first sentence received from the tag assignment unit 582 as the original sentence, the learning data having the second sentence received from the tag assignment unit 592 as the translated sentence, and the second sentence received from the tag assignment unit 592. The original text and learning data having the first sentence received from the tag assigning unit 582 as a translated sentence are generated and stored in the learning data storage unit 548. The parallel sentence reading unit 540, the first sentence processing unit 542, the second sentence processing unit 544, and the learning data generation unit 546 generate a large number of learning data in this way and store them in the learning data storage unit 548.

  When a sufficient number of learning data is generated in the learning data storage unit 548, the NN learning unit 550 learns the NN 552 using the learning data. This learning method is as described above. In the learning, when the NN552 parameter satisfies a certain end condition, the learning of the NN552 is finished, and the parameters that define the function of the NN552 at that time are stored in the NN parameter storage unit 514, and the learning of the NN is finished. This parameter is set to NN included in the translation engine 606.

  During translation, the machine translation device 518 operates as follows. Prior to translation, the user operates the input / output device 600 to specify the language of the translation destination. The target language storage unit 602 stores a tag indicating the designated language.

  When the input sentence 516 is input to the machine translation device 518 and translation is requested, the tag assigning unit 604 reads a tag indicating the language of the translation destination from the target language storage unit 602 and assigns the tag to the head of the input sentence 516. To do. Further, the tag assigning unit 604 assigns a sentence end symbol <EOS> to the end of the input sentence 516 and inputs it to the translation engine 606.

  The translation engine 606 gives each word of the input sentence 516 as input to the NN in which the learning parameters are set. When an end sentence <EOS> at the end of the input sentence 516 is given as an input, the word obtained at the output of the NN becomes the first word of the translated sentence. Thereafter, the word string thus obtained is given to the NN input, and the obtained output is sequentially connected to obtain a translation word string for the input sentence 516. When the sentence end symbol <EOS> is obtained as the output of NN, the translation engine 606 finishes the translation, concatenates the word strings obtained so far, and outputs it as a translated sentence 520.

<Effects of the present embodiment>
According to the translation system 500 according to the fourth embodiment, different tags are given to the sentence heads depending on the language of the translation destination. In the NN, these tags are also considered as input words to the translation engine NN. Therefore, in an NN learned by parallel translation of a plurality of languages using such tags, a single NN can perform translation between the plurality of languages. When multiple languages have a common property, even if there are a small number of parallel translations that contain sentences in a particular language, such specific language can be learned by learning with parallel translations of languages that have other common properties. The language translation accuracy can also be improved. In addition, in this case, there is no need to change the configuration of the translation engine itself by the NN, and only a tag indicating the language of the translation destination is added to the head of each sentence as preprocessing at the time of learning and translation. Therefore, the accuracy of machine translation can be improved with a simple configuration.

  In the above-described fourth embodiment, at the time of learning, the language identifying unit identifies the language for both the first sentence and the second sentence of the parallel translation. If each bilingual translation stored in the multilingual bilingual corpus 510 has information identifying the language, there is no need to provide a language identification unit, and the translated language is indicated using the attached information. What is necessary is just to specify a tag. Preprocessing may be performed such that a tag indicating the language of the partner sentence in the pair is attached to the head of each parallel sentence in the multilingual parallel corpus 510.

  In the above-described embodiment, LSTM NN is used as a translation engine. However, the present invention is not limited to such an embodiment. Even when an RNN using a cell other than LSTM is used, the same learning as that of the fourth embodiment can be expected because similar learning is performed.

[Realization by computer]
The machine translation system, the learning processing unit, and the machine translation apparatus according to the above embodiments can be realized by computer hardware and a computer program executed on the computer hardware. FIG. 13 shows the external appearance of the computer system 930, and FIG. 14 shows the internal configuration of the computer system 930.

  Referring to FIG. 13, the computer system 930 includes a computer 940 having a memory port 952 and a DVD (Digital Versatile Disc) drive 950, a keyboard 946, a mouse 948, and a monitor 942.

  14, in addition to the memory port 952 and the DVD drive 950, the computer 940 boots up with a CPU (Central Processing Unit) 956, a bus 966 connected to the CPU 956, the memory port 952, and the DVD drive 950. A read only memory (ROM) 958 that stores programs and the like, and a random access memory (RAM) 960 that is connected to the bus 966 and stores program instructions, system programs, work data, and the like are included. The computer system 930 further includes a network interface (I / F) 944 that provides a connection to a network that enables communication with other terminals. The network I / F 944 may be connected to the Internet 968.

  A computer program for causing the computer system 930 to function as each of the functional units constituting the machine translation system, the learning processing unit, or the machine translation device of each of the above-described embodiments is a DVD 962 mounted on the DVD drive 950 or the memory port 952. Alternatively, it is stored in the removable memory 964 and further transferred to the hard disk 954. Alternatively, the program may be transmitted to the computer 940 through the network I / F 944 and stored in the hard disk 954. The program is loaded into the RAM 960 when executed. The program may be loaded directly into the RAM 960 from the DVD 962, the removable memory 964, or the network I / F 944.

  This program includes a plurality of instructions for causing the computer 940 to function as each functional unit of the machine translation system, the learning processing unit, or the machine translation device according to each of the above embodiments. Some of the basic functions required to perform this operation are provided by operating system (OS) or third party programs running on the computer 940, or modules of various programming toolkits installed on the computer 940. The Therefore, this program does not necessarily include all functions necessary for realizing the machine translation system, the learning processing unit, or the machine translation apparatus of this embodiment. The program calls the appropriate function or the appropriate program tool in the programming tool kit in a controlled manner so as to obtain a desired result. It is only necessary to include an instruction for realizing a function as a machine translation device. The operation of computer system 930 is well known. Therefore, it does not repeat here.

  Note that various corpora are stored in the hard disk 954 in the above embodiment, and are expanded in the RAM 960 as appropriate. All model parameters for translation are stored in the RAM 960. The finally optimized model parameters and the like are stored from the RAM 960 to the hard disk 954, the DVD 962, or the removable memory 964. Alternatively, the model parameter may be transmitted to another device via the network I / F 944 or may be received from another device.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are included. Including.

62, 82 Word order conversion 64, 68, 84, 88, 182, 192 Word string 66, 86, 180, 190 Tag assignment processing 110, 220 Bilingual corpus 114, 258, 364, 444 Model learning unit 118, 226, 344, 516 Input sentence 120, 230, 348, 416, 518 Machine translation device 122, 228, 346, 414, 520 Translated sentence 140, 260, 280 Morphological analyzer 142, 372, 382 Syntax analyzer 144, 264, 284 Pre-rearranger 146, 274, 374, 454, 474, 582, 592, 604 Tag assignment unit 148, 288 PBSMT device 184, 194 Translation by PBSMT 210, 320, 400 PBSMT system 222, 342, 410 Model storage unit 240 Parallel corpus with meta information 250, 540 Bilingual reading Part 252,360,440 original processing unit 254 translation processing unit 256,362,442 learning data storage unit 262,282 grammar type determination unit (syntax analysis section)
266, 286 Tag assignment unit by grammar type 224, 340, 412, 512 Learning processing unit 370, 380 Meta information separation unit 384 Meta information tag assignment unit 450, 470 Context information storage unit 452, 472 Previous sentence context information storage unit 510 Multilingual parallel corpus 514 NN parameter storage unit 542 First sentence processing unit 544 Second sentence processing unit 546 Learning data generation unit 548 Learning data storage unit 550 NN learning unit 552 NN
602 Target language storage unit 606 Translation engine by NN

Claims (8)

Meta-information specifying means for specifying meta-information about translation;
Meta information corresponding tag insertion means for inserting a tag corresponding to the meta information specified by the meta information specifying means at a predetermined position of the original text of translation;
A machine translation device that receives as input the original text with the tag attached,
A plurality of predetermined types are defined as the meta information, and the meta information corresponding tag insertion unit selects the tag according to the type of the meta information.   The meta information corresponding tag insertion means is configured to specify a range in which translation using the meta information is performed in the original text, and a first tag corresponding to the meta information at a start position and an end position of the range. The machine translation device according to claim 1, further comprising range specifying tag insertion means for inserting the second tag and the second tag, respectively. The meta information specifying means includes:
Morphological analysis means for morphological analysis of the original text;
Syntax analysis means for performing syntax analysis of the original text that has been morphologically analyzed by the morpheme analysis means;
The grammar type output means for outputting, as the meta information of the original sentence, information indicating the grammatical type of the original sentence, obtained from the result of syntactic analysis of the original sentence by the syntactic analysis means. Item 3. The machine translation device according to Item 2. The original text has the meta information related to the translation of the original text,
The meta information specifying means includes meta information separating means for separating the meta information attached to the original text from the original text and supplying the meta information corresponding tag insertion means to the meta information corresponding tag inserting means. The machine translation device described.   The meta information includes the grammatical type of the original sentence, scene information about a scene where the original sentence is uttered, speaker information about a speaker who utters the original sentence, and a sentence translated by the machine translation means prior to the original sentence. 5. The machine translation apparatus according to claim 1, wherein the machine translation apparatus is selected from a group consisting of grammatical types of preceding texts.   The machine translation device according to claim 1, wherein the machine translation unit is a phrase-based machine translation unit. The meta information specifying means includes means for specifying a translation destination language of the original text of the translation as meta information,
The machine translation device according to claim 1, wherein the meta information corresponding tag insertion means includes means for inserting a tag indicating the translation language specified by the meta information into a predetermined position of the original text. A computer program that causes a computer to function as the machine translation device according to any one of claims 1 to 7.

Images