JP5410334B2 - Word order conversion device, machine translation statistical model creation device, machine translation device, word order conversion method, machine translation statistical model creation method, machine translation method, program - Google Patents

Description

  In the present invention, a sentence written in a non-subscript postfix language such as English (a language in which the main prefix is not often postfixed) (source language) (source language) The present invention relates to a language translation technology that translates into (target language) (target language).

  Many Asian languages, such as Japanese, have the property that the head (or modifier) is often located after the modifier, and is called a head postfix language. For example, if you translate the English word “I respect you.” Into Japanese, “I respect you.” Here, “respect” is the head of “you”. Looking at the word order, it can be seen that “respect” is positioned before “you” in English, whereas “respect” is positioned after “you” in Japanese.

  In recent years, the Internet has been developed, and Internet users have increased opportunities to view foreign language documents. Along with this, the demand for translation is also increasing, and machine translation services are available on the Internet free of charge or for a fee. These translation services are roughly classified into rule-based machine translation (RBMT) and statistical machine translation (SMT). Some translation services can translate between several tens of languages.

  RBMT is realized by a person programming a translation dictionary and translation rules. At that time, the programmer is familiar with the grammar of both the source language and the target language, thinks about rules for changing the word order or replacing with corresponding words, and writes down the rules in a form that can be executed as a program.

  However, in RBMT, program maintenance becomes more difficult as the number of rules increases. Even if the target language is Japanese, there are many programmers who are familiar with both English and Japanese grammar, but there are programmers who are familiar with both non-English source and Japanese grammar. The number is small. Therefore, recently, there has been a lot of research on SMT that automatically acquires translation rules by statistical processing from a large amount of document data (translational corpus) actually translated.

  In SMT, humans do not have to directly manage translation rules, but high-quality translation rules cannot be obtained unless a large number of actual translated examples (parallel corpus) are prepared. In addition, even if a large number of parallel corpora are prepared, a huge amount of calculation is required.

  Thus, each translation technique has advantages and disadvantages. Therefore, the framework of SMT that simply handles words as translation units can be expanded to handle multiple word sequences (phrases) (phrase-based SMT), and the phrases can be hierarchized (hierarchical phrase-based). Improvements such as SMT) have been proposed.

  For example, Patent Document 1 discloses an improved technique related to a method for automatically acquiring translation rules in phrase-based SMT. Patent Document 2 discloses an improved technique related to the calculation amount in hierarchical phrase-based SMT. In particular, in the technique disclosed in Patent Document 2, the efficiency of solution search is improved by adding restrictions to the form of translation rules generated in the hierarchical phrase-based SMT.

  On the other hand, attempts have been made to perform statistical processing after parsing the source language or the target language, and Non-Patent Documents 1 and 2 disclose trends in this field. In particular, in Non-Patent Document 2, in order to reduce the size of a solution search space for SMT, a solution search space reduction technique called ITG (Inversion Transduction Grammar) and its extension IST (Imposing Source Tree) -ITG Such a technique is disclosed.

  In addition, many people have shown various rules of thumb regarding how to rearrange the word order. For example, Non-Patent Documents 3 and 4 show some empirical rules that serve as an index for changing the word order in English-Korean translation, including moving the main word last in verb phrases and adjective phrases. The specific rule is also included.

JP 2007-328483 A JP 2008-15844 A

Akira Shioda, “The role of syntax in the fusion of rule-based translation and statistical translation”, Japio 2009 YEARBOOK, Japan Patent Information Organization, 2009, pp.286-289 Eiichiro Sumida, “A New Approach to Introduce Structural Constraints in Statistical Translation”, Japio 2008 YEARBOOK, Japan Patent Information Organization, 2008, pp.266-271 Gumwon Hong, Seung-Wook Lee and Hae-Chang Rim, “Bridging Morpho-Syntactic Gap between Source and Target Sentences for English-Korean Statistical Machine Translation”, Proceedings of the ACL-IJCNLP-2009, Singapore, pp. 233-236, 2009 Peng Xu, Jaeho Kang, Michael Ringgaard, and Franz Och: “Using a Dependency Parser to Improve SMT for Subject-Object-Verb Languages”, Proceedings of the NAACL-HLT-2009, pp. 245-253, 2009.

  However, none of the above-mentioned conventional techniques can realize a translation with sufficient quality. The reason is that, for example, in the translation from English to Japanese, there is a problem that the word order has to be largely changed. In RBMT, it is difficult to realize high accuracy by comparing the order of words in English and Japanese, finding the case where the order of words is changed, and defining the rules for when and how to change the order. is there.

On the other hand, in SMT, it is statistically determined which word and which word should be replaced. However, for example, if there are 30 words in a sentence, the word rearrangement method is theoretically 30! ≒ 10 ^32.4 . Performing such a huge solution search for translation of only one sentence takes too much time and is unrealistic.

  In addition, the solution search space has been gradually reduced by the techniques disclosed in Non-Patent Documents 1 and 2, Patent Documents 1 and 2, and so on. In this case, it is determined by a statistical language model (single language model). In other words, the probability value of which word comes after a certain word is obtained, and the word order with the highest probability is given priority. In this word model determination, when a new word is included, the good word order is discarded. There is a possibility.

  The rule of thumb disclosed in Non-Patent Document 3 is only for English-Korean translation and cannot be applied to other translations such as English-Japanese translation or German-Korean translation as it is. In addition, the technique disclosed in Non-Patent Document 4 has a problem that, for example, when applied to English-Japanese translation, a complicated sentence including a plurality of sections may not be in Japanese word order.

  Therefore, the present invention has been made in view of the current state of machine translation, and the postfix type of a sentence in a source language (for example, English) which is a non-subject postfix language is used by using the result of parsing. The task is to rearrange the language in the order of the target language (for example, Japanese).

In order to solve the above-mentioned problem, the present invention according to claim 1 arranges words of a sentence of a source language that is a non-subject postfix language in order of words of a target language that is a postfix postscript language with respect to a syntactic main subject. A word-order conversion device for replacing the syntax tree information indicating the hierarchical structure of nodes corresponding to the words constituting the sentence, and the child node related to each node, created by a syntax analysis performed in advance on the source language sentence. A storage unit that stores main node node information indicating which of the main node is included, and word order invariant node information that indicates a node that should not be changed in word order, including when each node has a parallel structure and a mathematical expression; For the sentence, referring to the syntax tree information, the head node information, and the word order invariant node information, starting from the root node of the structure tree shown in the syntax tree information, all the structure trees For nodes, among the child nodes, the node other than the node that should not alter the word order, by repeating the process of moving the head-node to the last position of the child node, the target word of the sentence of the source language And a processing unit that rearranges the words in order of language.

According to this invention, with respect to the syntactic main word, the word order should not be changed among the child nodes of all the nodes starting from the root node of the structural tree for the words of the source language sentence which is a non-subscript postfix language. By repeating the process of moving the main word node to the last position of the child node for nodes other than the node, it is possible to rearrange the word order in the target language of the main word postfix language.

Further, the present invention according to claim 2 relates to a statistical machine used for translating a sentence in a source language, which is a non-subscript postfix language, into a sentence in a target language, which is a postfix language, with respect to a syntactic head A statistical model creation device for machine translation that creates a statistical model for translation, the storage storing a bilingual corpus of the source language sentence and the target language sentence, and a monolingual corpus of the target language sentence A syntactic analysis unit that parses the sentence in the source language of the bilingual corpus, a word order conversion device according to claim 1, and a word order conversion by the word order conversion device after the syntax analysis by the syntax analysis unit A main postfix translation bilingual corpus that creates a bilingual corpus that is a bilingual corpus consisting of the source language sentence and the target language sentence, the main postfix translation bilingual corpus, and the object Using monolingual corpora of words, characterized in that it comprises a statistical model creation unit for creating the statistical model.

  According to this invention, a statistical model is obtained using a well-known statistical machine translation technique, except that a word order conversion device according to claim 1 converts a word order of a source language sentence which is a non-subscript postfix language. Can be created.

Further, the present invention according to claim 3 is a machine translation device that translates a sentence in a source language that is a non-subscript postfix language into a sentence in a target language that is a postfix language for a syntactic main word, 3. A storage unit for storing the statistical model according to claim 2, a syntax analysis unit for parsing the source language sentence to be translated, and converting a word order of words of the source language sentence that has been parsed. 1, a statistical machine translation unit that translates the sentence into the target language by applying the statistical model stored in the storage unit to the converted source language sentence; It is characterized by providing.

  According to this invention, the statistical machine translation unit refers to the statistical model according to claim 2 for the source language sentence converted by the word order conversion device according to claim 1, thereby enabling enormous word order conversion. Translation into a target language sentence can be performed without performing a search or the like.

  Further, the present invention is a program for causing a computer to execute the method using each of the above devices. According to this invention, by installing this program in the computer, the computer can execute the method of each device described above.

  According to the present invention, words of a sentence in a source language that is a non-subscript postfix language can be rearranged in the order of words in the target language of the postfix language using a syntactic analysis result.

It is a block diagram of the word order conversion apparatus of this embodiment. It is a figure which shows the example which converts an English sentence into a postfix main sentence English sentence. It is a figure which shows the example which converts a German sentence into a head postfix German sentence. It is a flowchart which shows the flow of a process of a word order conversion apparatus. It is a figure which shows the example of word order conversion. It is a figure which shows the example of word order conversion. It is a block diagram of the statistical model production apparatus for machine translation of this embodiment. It is a block diagram of the machine translation apparatus of this embodiment. It is a figure which shows the example of an output of the analysis result of an English sentence. It is a figure which shows the example of an output of the analysis result of an English sentence.

  Hereinafter, with respect to a mode for carrying out the present invention (hereinafter referred to as “embodiment”), refer to the drawings for the first embodiment, the second embodiment, and the third embodiment (the drawings other than the referenced drawings are also appropriately referred to). The details will be described.

<First Embodiment>
As shown in FIG. 1, the word order conversion device 100 according to the first embodiment is a computer device, and includes a processing unit 101 and a storage unit 195. Although not shown, the word order conversion apparatus 100 also includes an input / output interface.

  The processing unit 101 is realized by, for example, a CPU (Central Processing Unit) and a memory, and includes an initial setting unit 110, a terminal node determination unit 120, a child node enumeration unit 130, a word order exchange unit 140, a verb determination unit 150, and an argument label assignment unit. 160, a word replacement unit 170, a node selection unit 180, and a particle symbol addition unit 190. These operations will be described later.

  The storage unit 195 is realized by, for example, a ROM (Read Only Memory), an HDD (Hard Disk Drive), and the like. The processing target text, the operation program of the processing unit 101, various setting values, and the processing target language text Parse tree information created by parsing in advance, head node information indicating which head node is a child node of each node in the syntax tree, and a parallel structure or expression of each node in the syntax tree Word order invariant node information indicating nodes that should not change the word order, and sentences after word order conversion are stored.

  In the following, a case where the source language is English will be described as an example. The operation outline of the word order conversion apparatus 100 will be described. The word order conversion apparatus 100 outputs the input English sentence as a main part postfix English sentence rearranged in the word order of a main part postfix language such as Japanese or Korean.

  It is assumed that the input sentence has been parsed in advance, and the result is expressed as a syntax tree (syntax tree information). When the input sentence has not been parsed, it is necessary to analyze the syntax by a syntax analysis unit (shown by a broken line in FIG. 1) to form a syntax tree. The syntax analysis unit is included in the word order conversion apparatus 100. Or other devices may be provided.

  The English parsing is, for example, the document “Feature forest models for probabilistic HPSG parsing,” (Y. Miyao and J. Tsujii: Computational Linguistics, vol. 34, no. 1, pp. 35-80, 2008) ”(below , Referred to as “English parsing document”), and techniques such as enju that are publicly available on the web can be used.

  For example, when an English sentence “John saw a beautiful girl yesterday.” Is input, an English syntax tree output by the syntax analysis unit is as shown in FIG. And the word postfix English sentence which the word order conversion apparatus 100 of this embodiment converts and outputs the syntax tree is as shown in FIG.2 (b).

  A node ID (IDentifier) is assigned to each node of the syntax tree. A number written in a node (indicated by a circle) in FIG. 2 is a node ID. The node shown in gray is the main node for the node connected immediately above. A node represented in white is a node other than the main letter. For example, when attention is paid to a node having a node ID of 3 (hereinafter referred to as node 3, the same applies to other nodes), this node 3 has node 4 and node 11 as child nodes, and the main word is node 4.

Although not shown in the figure, it is assumed that a label indicating the following meaning is given to each node of the syntax tree obtained as a result of the syntax analysis.
-ID of the node that is the main word
-ID of the subject node (only for nodes corresponding to verbs)
-ID of the target node (only for nodes corresponding to transitive verbs)

  FIG. 8 shows the result of analyzing the English sentence “John saw a beautiful girl yesterday.” In enju in the XML (eXtensible Markup Language) format. The node ID of the syntax tree in FIG. 2 is described in the id attribute in the <cons> tag (lines 2 to 4, 8 to 10, 14, 15, 18, 19, 22, 28, 29) in FIG. Yes. In FIG. 8, id is c0 to c12, but in FIG. 2, c is omitted and only numbers are shown. In the analysis result of “enju”, the following information is given to each node, and this is used to give a label to each node.

(1) ID (head) of main node in child node
Information “head” is described in a <cons> tag and indicates a head attribute. That is, for example, in the case of the node 3 in FIG. 2A, since head = “c4” is described in the eighth line in FIG. 8, the node 4 becomes the main node for the node 3. Similarly, from the 14th line in FIG. 8, the node 8 becomes the main node for the node 6.

(2) ID of the node that is the first argument of the node (arg1)
The information “arg1” indicates the subject in the case of a verb. For example, in FIG. 2A, when the saw XML tag (11th and 12th lines in FIG. 8) connected to the node 5 (c5) is seen, c1 is described in arg1. Therefore, if the head of the node 1 (c1) is traced (see the third line in FIG. 8), the node 2 (John) is reached, so that the label “subject” is given to the node 2.

(3) ID of the node that is the second argument of that node (arg2)
The information “arg2” indicates the object in the case of a transitive verb and the complement in the case of a be verb. For example, in FIG. 2A, looking at the XML tag of the saw connected to the node 5 (c5) (11th and 12th lines in FIG. 8), c6 is described in arg2. Therefore, if the head of the node 6 (c6) is traced (see the 14th and 18th lines in FIG. 8), the node 10 (girl) is reached, and therefore, the label “object” is given to the node 10.

  In the word order conversion apparatus 100 of the present embodiment, processing is started from the highest node (root node) of the input syntax tree and connected to a processing target node (hereinafter referred to as “processing target node”). Find the node that is the main character from the child nodes that are in the list, and move that node to the end of the child node. Note that the word order invariant node information in the storage unit 195 is referred to, and if the child node has a parallel structure or a mathematical expression, the main word node is not moved. By repeating this for all child nodes, word order conversion is performed.

  Hereinafter, word order conversion processing will be described with reference to FIGS. 1, 4, and 8. After that, a specific example of word order conversion will be described with reference to FIGS. 5A and 5B.

<Initial setting unit 110>
The initial setting unit 110 receives (reads) the parsed English sentence (syntax tree) from the syntax analysis unit (step S100), selects the root node of the grammatical tree of the English sentence as a processing target node, and displays the node in the node list. The processing target node is recorded (step S110).

<Termination node determination unit 120>
The termination node determination unit 120 determines whether the processing target node is a termination node (a node having no child node) (step S120). If Yes, the process proceeds to step S150, and if No, the process proceeds to step S130. When XML format output is selected in enju, the node with the <tok> tag is the terminal node (lines 5, 11, 16, 20, 23, and 30 in FIG. 8), and <cons The one with tag is not a terminal node.

<Child node enumeration unit 130>
The child node enumeration unit 130 creates a “child node list” that is a list of child nodes of the processing target node (step S130). For example, if the XML output of enju is read by the python xml.dom.minidom library, child nodes can be enumerated by childNodes.

<Word order exchanging unit 140>
If the processing target node is not a parallel structure or a mathematical expression, the word order exchange unit 140 moves the node ID corresponding to the main node in the “child node list” to the end of the “child node list” (step S140). In the XML output of enju, the ID of the main node is recorded in the head attribute of the processing target node. In the case of a parallel structure, the attribute of cat or xcat of the <cons> tag corresponding to the processing target node is “COOD”. For example, in the English sentence “Mary saw John and Bob.”, The part of “John and Bob” has a parallel structure.

  FIG. 9 shows the result of analyzing an English sentence “Mary saw John and Bob.” With enju, and xcat = “COOD” is described on the 20th line. By following the head of this node (id = “c19”), “John” is reached (see the 20th to 23rd lines in FIG. 9). Whether or not the expression is a mathematical expression can be determined by a criterion such as whether an equal sign or an inequality sign is included. Then, the node ID portion of the processing target node in the node list is replaced with the rewritten “child node list” (step S141).

  Here, the reason for considering the structure of the child node is as follows. Some well-known parsers determine A as a main character in the case of a parallel structure (coordination) such as A and B or A or B. At this time, if the node corresponding to the main node is simply moved to the end of the child node list without considering the parallel structure, the word order is converted to B and A or B or A. Even though A and B are equal, such a change is not a faithful translation. Therefore, when the child node has a parallel structure, the problem caused by such a syntax analyzer can be solved by not changing the word order. Also, mathematical expressions such as m <n + 1 do not change the word order because the meaning changes if the word order is changed.

<Verb determination unit 150>
The verb determining unit 150 determines whether or not the processing target node is a verb in order to apply the argument label assigning unit 160 (step S150). If Yes, the process proceeds to step S160, and if No, step S170 is performed. Proceed to

<Argument label giving unit 160>
The argument label assigning unit 160 performs preprocessing for generating particles that do not exist in English. When the processing target node is a verb, the argument label assigning unit 160 labels the word that is the subject / object. In the case of “enju”, node IDs corresponding to the arg1 and arg2 attributes given to the verb are described. However, since these node IDs do not directly correspond to words, it is necessary to add a process of tracing head.

That is, the terminal node (word) can be reached as follows.
・ Starting from the node arg1 of the processing target node (verb) and following head, the terminal node (word) that is the subject can be reached.
-Starting from the node arg2 of the processing target node, the head can be traced to reach the terminal node (word) that is the object.

  Then, labels va1 and va2 indicating the subject and object are attached to these terminal node IDs (step S160). For example, in the case of node 5 (saw) in FIG. 2, since arg1 = c1, arg2 = c6 (see the 12th line in FIG. 8), va1 is the head of node1 and head of node6 (node8) A label va2 is given to the girl which is the head (node 10).

  In addition, when attaching a label, it may be determined whether or not the input sentence is passive. If it is passive, the result of the parser may be corrected and the main and objective labels may be assigned. For example, in a sentence such as “A was stolen by B”, depending on the parser, A may be determined to be the object and B to be the subject. By assigning a label with the subject, B as the object, a translation that is faithful to the original can be obtained.

  Also, when the input sentence is “John bought a toy that was popular in Japan”, toy is the object of bought and the subject of was. In such a case, the noun (toy) that has been modified is only labeled by a case outside the modified clause. In other words, since this toy is the object of the verb bought outside the combined modifier, it is labeled as an object.

<Word replacement unit 170>
The word replacement unit 170 replaces the processing target node in the node list with the corresponding word (step S170). At this time, if there are labels of va1 and va2 in the node list, these labels remain associated with the words that have already been replaced (see (14) in FIG. 5B).

<Node selection unit 180>
The node selection unit 180 determines whether or not the node ID is included in the node list (step S180). If Yes, the node ID closest to the head of the node list is selected as the processing target node (step S181). ), The process proceeds to step S120. If No in step S180, the process proceeds to step S190.

<Particulate symbol addition unit 190>
The particle symbol adding unit 190 adds the particle equivalent words “_va1” and “_va2” indicating the possibility that the word corresponding to the particle is included immediately after the word labeled va1 and va2 in the node list (step S1). S190), the words in the node list are output in order from the front (step S200). In many cases, the main subject in Japanese is “ha” rather than “ga”, so “_va1” in the main clause may be replaced with another word like “_va0” ( See below for examples). However, it is difficult to properly use "ga" and "ha", "wo" and "ni", etc., so these particle equivalents are not actually fixedly replaced with Japanese words, It depends on the processing of statistical machine translation which particle to replace with.
With the above processing, English sentences in which words are rearranged after the main word are output.

  Next, the operation of the above-described word order conversion apparatus 100 will be specifically described on the assumption that the syntax tree obtained by the syntax analysis unit is as shown in FIG. Here, how the node list is updated is shown in FIGS. 5A and 5B. The numbers in () below ((1), (2),...) Correspond to the numbers in () in FIGS. 5A and 5B ((1), (2),...). doing. In FIG. 5A and FIG. 5B, the portion of the node list whose contents are updated as compared with the previous step is underlined.

(1) First, node 0 is selected as a processing target node (step S110 in FIG. 4). The node list is [Node ID: 0].

(2) Since a child node exists in node 0 (since it is not a terminal node), the processing is transferred from the terminal node determination unit 120 (step S120 in FIG. 4) to the child node enumeration unit 130 (step S130 in FIG. 4). To do. In the child node enumeration unit 130, the child node list of node 0 [node ID: 1, node ID: 3] is set as a child node list (step S130 in FIG. 4). In the word order exchange unit 140, since the child node structure is not a parallel structure or a mathematical expression, the main word node (node 3) in the child node list is moved to the end of the child node list (step S140 in FIG. 4). However, in this case, there is no need to move because the head node is already the last. Therefore, the node list is replaced with [node ID: 1, node ID: 3] (step S141 in FIG. 4).

(3) Next, the node selection unit 180 selects node 1 as a processing target node (step S181 in FIG. 4), and repeats the same processing. The only child node of node 1 is node 2. Therefore, it is not necessary to exchange the word order in the word order exchange unit 140, and the node list becomes [node ID: 2, node ID: 3].

(4) Next, node 2 is selected as a processing target node. Since node 2 has no child node, the process proceeds to the verb determination unit 150 (step S150 in FIG. 4). Further, since node 2 is not a verb (No in step S150 in FIG. 4), the process proceeds to the word replacement unit 170 (step S170 in FIG. 4). Since the word corresponding to the node 2 is John, the node list is [word: “John”, node ID: 3] (step S170 in FIG. 4).

(5) Next, the node 3 is selected as a processing target node. The child node list is [node ID: 4, node ID: 11]. The word order exchange unit 140 moves the main word node (node 4) to the end of the child node list. Therefore, the node list reflecting the child node list is [word: “John”, node ID: 11, node ID: 4].

(6) Subsequently, the node 11 is selected as a processing target node. Since the child node is only the node 12, the node 11 is replaced with the node 12, and the node list becomes [word: “John”, node ID: 12, node ID: 4].

(7) Next, the node 12 is selected as a processing target node. Since the node 12 has no child node, it is replaced with a word, and the node list becomes [word: “John”, word: “yesterday”, node ID: 4].

(8) When node 4 is selected as a processing target node and the same processing is repeated, the node list becomes [word: “John”, word: “yesterday”, node ID: 6, node ID: 5].

(9) The node 6 is selected as a processing target node. As a result, the node list is [word: “John”, word: “yesterday”, node ID: 7, node ID: 8, node ID: 5].

(10) Node 7 is selected as a processing target node. Since node 7 has no child node, it is replaced with a word, and the node list becomes [word: “John”, word: “yesterday”, word: “a”, node ID: 8, node ID: 5]. .

(11) Node 8 is selected as a processing target node, and the node list is [word: “John”, word “yesterday”, word: “a”, node ID: 9, node ID: 10, node ID: 5]. Become.

(12) Similarly, since node 9 is selected as a processing target node and replaced with a word, the node list is [word: “John”, word: “yesterday”, word: “a”, word: “beautiful”, node ID: 10, node ID: 5].

(13) Since node 10 is selected as a processing target node and replaced with a word, the node list is [word: “John”, word: “yesterday”, word: “a”, word: “beautiful”, word: “girl” ", Node ID: 5].

(14) Next, the node 5 is selected as a processing target node. Since node 5 is a verb (Yes in step S150 in FIG. 4), when its attributes are examined in the argument label assigning unit 160, arg1 = c1, arg2 = c6 is written (see lines 10 to 12 in FIG. 8). . Therefore, when the attribute of the node 1 is examined, since head = c2 (see the third line in FIG. 8), the node 2 reaches “John” (see the fourth and fifth lines in FIG. 8). Therefore, the label “va1” is assigned to “John” (step S160 in FIG. 4). Since node 6 is head = c8 (see the 14th line in FIG. 8) and node 8 is head = c10 (see the 18th line in FIG. 8), it reaches “girl” (see the 22nd and 23rd lines in FIG. 8). ), The label “va2” is given to “girl” (step S160 in FIG. 4). Therefore, the node list is [word: “John”: va1, word: “yesterday”, word: “a”, word: “beautiful”, word: “girl”: va2, word: “saw”]. Here, va1 is used as a subject and va2 is used as a label indicating an object.

(15) Lastly, the particle symbol adding unit 190 adds the particle equivalent words “_va1 (ga)” and “_va2 ()” immediately after the word given the labels va1, va2, and the node list [[ John ”,“ _va1 ”,“ yesterday ”,“ a ”,“ beautiful ”,“ girl ”,“ _va2 ”,“ saw ”] (step S190 in FIG. 4). Furthermore, if “_va1 (ga)” is replaced with “_va0 (ha)”, the node list will be [“John”, “_va0”, “yesterday”, “a”, “beautiful”, “girl”, “_va2” , “Saw”] (step S190 in FIG. 4).

  Through the above processing, the English sentence “John saw a beautiful girl yesterday.” Is converted into a main part postfix English sentence “John _va0 yesterday a beautiful girl _va2 saw.”. This makes it possible to sequentially translate into Japanese, “John saw a beautiful girl yesterday.”

  Even when more complicated English sentences are input, they can be converted in the order of words close to Japanese. For example, the English sentence “John went to the police because Mary lost his wallet.” Is converted to “John _va0 Mary _va1 his wallet _va2 lost because the police to went.” Sequentially translating this into Japanese, "John went to the police because Mary lost his wallet," and it turns out that the accuracy of word order conversion is high.

  Similarly, many English sentences can be arranged in a word order close to Japanese. This is possible because Japanese is a language with a regular nature ("postfix postfix language") that "places the head at the end".

  As described above, according to the word order conversion apparatus 100 of the present embodiment, the syntax tree for the sentence in the source language (English or the like) is examined, and the head node (qualifier) is placed behind the node (modifier) other than the main word. By moving, each word of the sentence in the source language can be rearranged in the order of the target language, which is the postfix word. At that time, it is possible to perform more accurate word order conversion by not changing the word order of the nodes whose word order should not be changed, such as a parallel structure or a mathematical expression.

  The word order conversion apparatus 100 can be applied not only to English but also to other languages. For example, the case of German (German) will be described. FIG. 3 is an explanatory view similar to FIG. That is, FIG. 3A is a diagram showing an input syntax tree when the source language is German and the target language is Japanese. FIG. 3B is a diagram illustrating a result of converting the syntax tree into a head suffix.

  In German, when there is an auxiliary verb, the verb moves to the end of the sentence. When word order conversion is performed on “Albert muste jeden Tag Arbeiten.” Shown in FIG. 3A, “Albert (_va0) jeden Tag Arbeiten muste.” Shown in FIG. Sequentially translating this into Japanese, “Albert had to work every day.” It turns out that the accuracy of word order conversion is high.

Second Embodiment
Next, a statistical model creation apparatus for machine translation using the word order conversion apparatus 100 of the first embodiment will be described. In the second embodiment, the case where the source language is English and the target language is Japanese will be described as an example. However, instead of Japanese, the target language is a language such as Korean having the grammar of the suffix It is good. Also, the source language may be a language such as German instead of English.

  As shown in FIG. 6, the statistical model creation apparatus 200 for machine translation according to the second embodiment is a computer device, and includes an English-Japanese parallel corpus 210, a Japanese monolingual corpus 220, a syntax analysis unit 230, a word order. A conversion unit 100a (corresponding to the word order conversion device 100 of the first embodiment. Also serving as a “postfix postfix translation corpus creation unit” in the claims), a postfix postfix translation corpus 240, and a statistical model creation unit 250 are provided. Yes.

The machine translation statistical model creation apparatus 200 uses English and Japanese bilingual corpus 210 of learning data used to create a statistical model (translation rule) for translation in the conventional statistical machine translation (SMT) technique. After the main part is converted into the main part postfix by parsing each sentence by the method such as the English parsing literature described above by the syntax analysis unit 230 and performing the word order conversion by the word order conversion unit 100a. A bilingual corpus 240 is created.

  Then, the statistical model creation unit 250 uses the main part postfix bilingual corpus 240 and the Japanese monolingual corpus 220 as learning data, and performs statistics using a well-known statistical machine translation technique SMT (see Patent Documents 1 and 2). A model 260 can be created (learned).

<Third Embodiment>
Next, a machine translation apparatus using the word order conversion apparatus 100 of the first embodiment and the statistical model 260 of the second embodiment will be described. As shown in FIG. 7, the machine translation device 300 according to the third embodiment is a computer device, and includes a syntax analysis unit 230, a word order conversion unit 100 a (corresponding to the word order conversion device 100 according to the first embodiment), statistical machine translation. Unit 320 and statistical model 260 (created in the second embodiment).

  In the machine translation apparatus 300, the input English sentence is parsed by the syntax analysis unit 230, and then the word order conversion unit 100 a performs word order conversion to create a main part postfix English sentence. Then, using the statistical model 260 created (learned) by the machine translation statistical model creation apparatus 200 according to the second embodiment, the statistical machine translation unit 320 translates the postfix English sentence into Japanese. The statistical machine translation unit 320 can use a well-known rule-based machine translation (RBMT) or statistical machine translation (SMT) technique as it is.

  As described above, according to the machine translation apparatus 300, when the target language is a postfix word language for translation between languages such as English-Japanese translation and the like that are greatly different in word order, each sentence is brought close to the target language in advance. Therefore, it is possible to translate almost verbatim without searching for a large number of word orders, and it is possible to obtain a natural word order translation result at high speed.

  In this method, the creator of the source language syntax analysis unit need only have knowledge of the grammar of the source language, and does not need to know the grammar of the target language. If the syntax analysis can be performed, it is possible to approximate the word order of the target language only by applying the sentence (syntax tree) after the syntax analysis to the word order conversion unit 100a.

  In addition, in the technologies disclosed in Non-Patent Documents 3 and 4, basically only predicates such as verbs and adjectives are postfixed. For this reason, complicated sentences including a cause clause are written in Japanese. Can't be in the order of words. For example, Non-Patent Document 4 shows that “because” comes to the beginning of a sentence. For this reason, it is far from the word order of Japanese and Korean. On the other hand, according to the present embodiment, by performing the postfix processing without limiting the part of speech, the “because” comes last in the “because” clause, and the “because” is placed at the same position as the Japanese “So”. .

Although description of this embodiment is finished above, the aspect of the present invention is not limited to these.
For example, as the word order invariant node information, an administrator such as the word order conversion apparatus 100 may freely set other than the parallel structure and the mathematical expression.
In addition, the configuration of each device and the specific contents of the processing can be appropriately changed without departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 100 Word order conversion apparatus 100a Word order conversion part 101 Processing part 110 Initial setting part 120 Terminal node determination part 130 Child node enumeration part 140 Word order exchange part 150 Verb determination part 160 Argument label provision part 170 Word substitution part 180 Node selection part 190 Addition of particle symbol Section 195 Storage section 200 Machine translation statistical model creation device 210 English-Japanese bilingual corpus 220 Japanese monolingual corpus 230 Parsing section 240 Head-postfix bilingual corpus 250 Translation statistical model creation section 260 Statistical model 300 Machine translation Device 320 Statistical Machine Translation Department

Claims (7)

A word order conversion device that rearranges words in a source language sentence that is a non-subscript postfix language in the order of words in a target language that is a postfix language, with respect to a syntactic head letter,
About the sentence of the source language, the tree structure information indicating the hierarchical structure of the nodes corresponding to the words constituting the sentence and the main node among the child nodes related to each node, which is created by the syntax analysis performed in advance. A storage unit for storing main node information and word order invariant node information indicating a node that should not be changed in word order, including a parallel structure and a mathematical expression among the nodes;
For the sentence, referring to the syntax tree information, the head node information, and the word order invariant node information, starting from the root node of the structural tree indicated in the syntax tree information, all the nodes of the structural tree Among the nodes other than the node whose word order should not be changed, by repeating the process of moving the main word node to the last position of the child node, the words of the source language sentence are rearranged in the target language word order. A processing unit;
A word order conversion device comprising: For machine translation that creates a statistical model for statistical machine translation that is used to translate a sentence in the source language that is a non-subscript postfix language into a sentence in the target language that is a postfix language A statistical model creation device,
A bilingual corpus of the source language sentence and the target language sentence, and a storage unit that stores a monolingual corpus of the target language sentence;
A parsing unit that parses the source language sentence of the parallel corpus;
The word order conversion device according to claim 1;
A main part posttranslation that creates a main part posttranslational corpus that is a bilingual corpus consisting of the source language sentence converted by the word order conversion device after the syntax analysis by the syntax analysis unit and the target language sentence Corpus creation department,
A statistical model creation unit that creates the statistical model using the head-postfix bilingual corpus and the target language monolingual corpus;
A statistical model creation device for machine translation, comprising: A machine translation device that translates a sentence in a source language, which is a non-subscript postfix language, into a sentence in a target language, which is a postfix language, with respect to a syntactic subject,
A storage unit that stores the statistical model according to claim 2;
A parsing unit that parses the source language sentence to be translated;
The word order conversion device according to claim 1, wherein the word order of words in the sentence of the source language that has been analyzed is converted,
A statistical machine translation unit that translates the sentence into the target language by applying the statistical model stored in the storage unit to the converted source language sentence;
A machine translation device comprising: A word order conversion method by a word order conversion apparatus that rearranges words in a source language sentence that is a non-subject postfix language in the order of words in a target language that is a postfix word,
The word order conversion device includes:
About the sentence of the source language, the tree structure information indicating the hierarchical structure of the nodes corresponding to the words constituting the sentence and the main node among the child nodes related to each node, which is created by the syntax analysis performed in advance. A storage unit that stores main node information, word order invariant node information indicating a node that should not be changed in word order, including a parallel structure and a mathematical expression among the nodes, and a processing unit. ,
The processor is
For the sentence, referring to the syntax tree information, the head node information, and the word order invariant node information, starting from the root node of the structural tree indicated in the syntax tree information, all the nodes of the structural tree Among the nodes other than the node whose word order should not be changed, the word of the sentence in the source language is rearranged in the word order of the target language by repeating the process of moving the main word node to the last position of the child node. A word order conversion method characterized by that. For machine translation that creates a statistical model for statistical machine translation that is used to translate a sentence in the source language that is a non-subscript postfix language into a sentence in the target language that is a postfix language A statistical model creation method for machine translation by a statistical model creation device,
The machine translation statistical model creation device comprises:
A bilingual corpus of the source language sentence and the target language sentence, a storage unit that stores a monolingual corpus of the target language sentence, a word order conversion device according to claim 4, a syntax analysis unit, The main-postfix bilingual corpus creation unit and the statistical model creation unit
The parsing unit parses the source language sentence in the bilingual corpus,
The postscript bilingual corpus creation unit is a bilingual corpus consisting of a sentence in the source language and a sentence in the target language, which are converted into a word order by the word order conversion device after the syntax analysis by the syntax analysis unit. Create a bilingual corpus,
The statistical model creation unit creates the statistical model by using the prefix postfix parallel corpus and the monolingual corpus of the target language. The statistical model creation method for machine translation, A machine translation method using a machine translation device that translates a sentence in a source language that is a non-subscript postfix language into a sentence in a target language that is a postfix language with respect to a syntactic subject,
The machine translation device includes:
A storage unit that stores the statistical model according to claim 5, a syntax analysis unit, a word order conversion device according to claim 4, and a statistical machine translation unit,
The parsing unit parses the source language sentence to be translated;
The word order conversion device converts the word order of words in the sentence of the source language that has been parsed,
The statistical machine translation unit performs translation into the target language sentence by applying the statistical model stored in the storage unit to the converted source language sentence. .   The program for making a computer perform the method of any one of Claims 4-6.

Images