JP2007018462A - Machine translation device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem of a conventional machine translation device incapable of performing high-quality translation at a high speed. <P>SOLUTION: Based on an original language tree structure model having original language tree structure information using a long unit phrase, which consists of a plurality of words and includes no non-terminal symbol, as one node, a target language tree structure model having target language tree structure information using a long unit phrase, which consists of a plurality of words and include no non-terminal symbol, as one node, a tree structure mapping model having mapping information, which shows correspondence between the original language structure and the target language tree structure information, and a mapping probability, and a language model having one or more word appearance probabilities serving as information about a probability concerned with appearance of a word in a second language, syntax analysis is carried out on a received sentence for acquiring a translation model probability and a language model probability. Based on these two probabilities, a syntax tree of an output is evaluated, and a translation sentence to be outputted is decided. In this way, this machine translation device can perform high-quality translation at a high speed. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a machine translation device that translates received text into another language.

  The conventional first machine translation apparatus employs a translation algorithm called “Phrase-based SMT” (see, for example, Non-Patent Document 1). “Phrase-based SMT” is a statistical translation based on a phrase, and translation is performed in units of a plurality of words (phrases).

  The conventional second machine translation apparatus employs a translation algorithm called “syntax transfer system MT” (for example, see Non-Patent Document 2). In the machine transfer of the syntax transfer method, an input sentence is parsed, and translation is performed by converting the obtained syntax tree into an output syntax tree.

Further, as a third conventional technique, there is a technique for automatically acquiring a translation model that can be used in the machine translation apparatus. Rules included in such a translation model (generic name of source language tree structure model, target language tree structure model, tree structure mapping model) are automatically generated from a corpus using a hierarchical phrase alignment method (see Non-Patent Document 3). Can be extracted. The probabilities of these models can be calculated by processing such as counting the frequency of use of rules in the corpus and calculating the relative frequency. With this third technique, the source language tree structure model, the target language tree structure model, and the tree structure mapping model can be automatically acquired.
Philipp Koehn, Franz J. Och, and Daniel Marcu: Statistical Phrase-Based Translation.HLT-NAACL 2003: Main Proceedings, 2003, pp.127-133 Kurose and others, multilingual spoken language translation using constituent boundary analysis, natural language processing, Vol. 6, No. 5, 1999, pp.63-91 Kenji Imamura, Hierarchical phrase alignment fused with syntax analysis, Natural language processing, Vol. 9, No. 5, 2002, pp. 23-42

However, in the conventional first machine translation apparatus, the translated sentence is not correct unless the phrase order is adjusted. In the translation method of this machine translation apparatus, the order of phrases is adjusted on a flat structure and verified by a language model. For this reason, there is a problem that a syntactically incorrect translation may be output.
In the second conventional machine translation apparatus, when a plurality of structures are obtained as input analysis results, the ambiguity is resolved using the semantic distance of words. For this reason, there is a problem that a thesaurus is required.

The machine translation device according to the first aspect of the invention is information relating to the tree structure of the first language, which is the language of the original sentence to be translated, and a long unit phrase consisting of a plurality of words and not including a non-terminal symbol is regarded as one node. Source language having at least one source language tree structure record having source language tree structure probabilities that are information indicating the probability of the tree structure corresponding to the source language tree structure information. This is information about the source language tree structure model storage section that stores the tree structure model and the tree structure of the second language that is the language of the sentence of the translation result. One target language tree structure record including target language tree structure information including information on a tree structure as one node and a target language tree structure probability which is information indicating a probability of a tree structure corresponding to the target language tree structure information. Purpose of having A target language tree structure model storage unit that stores a word tree structure model, mapping information that is information indicating correspondence between source language tree structure information and target language tree structure information, and a correspondence probability indicated by the mapping information. A tree structure mapping model storage unit storing a tree structure mapping model having one or more tree structure mapping records having a mapping probability which is information indicating, and a word appearance probability which is information on the probability of word appearance in the second language A language model storage unit that stores one or more language models, a reception unit that receives a sentence in a first language, a sentence that is received by the reception unit, and a partial or entire tree structure of the sentence An input sentence syntax analysis unit that sequentially obtains tree structure information as information, and one or more source language trees corresponding to the tree structure information obtained by the input sentence syntax analysis unit And the one or more target language tree structure information and one or more mapping probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit are acquired from the source language tree structure model storage unit. One or more target language tree structure probabilities obtained from the storage unit and corresponding to each of the one or more target language tree structure information are acquired from the target language tree structure model storage unit, and the acquired target language tree structure information is included in the acquired target language tree structure information. On the basis of one or more word appearance probabilities of two or more words constituting the target language tree structure information from the language model storage unit, the acquired source language tree structure probability, the acquired mapping probability, A tree structure mapping probability calculation unit for calculating an evaluation value of an output syntax tree based on the acquired target language tree structure probability and the acquired word appearance probability; and the tree structure mapping Based on the evaluation value calculated by the probability calculation unit, a part or all of the syntax tree of the second language to be output is determined, and an output that is a sentence of all or part of the second language based on the determined syntax tree A machine translation apparatus comprising: a maximum likelihood sequence search unit that acquires information; and an output unit that outputs a second language sentence having one or more pieces of output information acquired by the maximum likelihood sequence search unit.
With this configuration, high-quality and high-speed translation becomes possible.

Further, in the machine translation device of the second invention, in contrast to the first invention, the tree structure mapping probability calculation unit includes one or more source languages corresponding to the tree structure information obtained by the input sentence syntax analysis unit. Source language tree structure probability acquisition means for acquiring a tree structure probability from the source language tree structure model storage unit, one or more target language tree structure information corresponding to the tree structure information obtained by the input sentence syntax analysis unit, and one or more Mapping probability acquisition means for acquiring the mapping probabilities from the tree structure mapping model storage unit, and one or more target language tree structure probabilities corresponding to each of the one or more target language tree structure information stored in the target language tree structure model Based on the acquired target language tree structure probability acquisition means and the acquired target language tree structure information, the one or more word appearance probabilities of two or more words constituting the target language tree structure information Based on the word appearance probability acquisition means acquired from the model storage unit, the source language tree structure probability, the mapping probability, the target language tree structure probability, and the word appearance probability, an evaluation value of the output syntax tree is calculated. It is a machine translation apparatus provided with an evaluation value calculation means.
With this configuration, high-quality and high-speed translation becomes possible.

The machine translation device according to the third aspect of the present invention provides two or more evaluation values calculated by the tree structure mapping probability calculation unit and an output syntax tree corresponding to the evaluation values, relative to the first and second aspects of the invention. A merge unit that combines evaluation values corresponding to the same syntax tree, and the maximum likelihood sequence search unit is obtained by combining the evaluation value calculated by the tree structure mapping probability calculation unit and the merge unit. A machine that determines part or all of the syntax tree of the second language to be output based on the evaluated value and acquires output information that is a sentence of all or part of the second language based on the determined syntax tree It is a translation device.
With this configuration, high-quality and high-speed translation becomes possible.
Further, in the machine translation device of the fourth invention, in contrast to the second and third inventions, the evaluation value calculation means includes the source language tree structure probability, the mapping probability, the target language tree structure probability, This is a machine translation device that calculates a product of word appearance probabilities and uses the product as an evaluation value of an output syntax tree.
With this configuration, high-quality and high-speed translation becomes possible.

The machine translation device according to the fifth aspect of the present invention provides the maximum likelihood sequence search unit according to any of the first to fourth aspects, wherein the maximum likelihood sequence search unit has a maximum evaluation value calculated by the tree structure mapping probability calculation unit. This is a machine translation device that determines a tree as a syntax tree to be output, and obtains output information that is a sentence of all or part of the second language based on the determined syntax tree.
With this configuration, high-quality and high-speed translation becomes possible.

  The machine translation apparatus according to the present invention enables high-quality and high-speed translation.

Hereinafter, embodiments of a machine translation apparatus and the like will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in embodiment performs the same operation | movement, description may be abbreviate | omitted again.
(Embodiment 1)
FIG. 1 is a block diagram of a machine translation apparatus according to the present embodiment.

The machine translation apparatus includes a source language tree structure model storage unit 11, a target language tree structure model storage unit 12, a tree structure mapping model storage unit 13, a language model storage unit 14, a reception unit 15, an input sentence syntax analysis unit 16, a tree structure A mapping probability calculation unit 17, a merge unit 18, a maximum likelihood sequence search unit 19, and an output unit 20 are provided. The machine translation apparatus also includes, for example, a keyboard 302 and a mouse 303 as input means. Furthermore, the machine translation apparatus includes a display 304 as output means, for example.
The tree structure mapping probability calculation unit 17 includes source language tree structure probability acquisition means 171, mapping probability acquisition means 172, target language tree structure probability acquisition means 173, word appearance probability acquisition means 174, and evaluation value calculation means 175.

  The source language tree structure model storage unit 11 is information on the tree structure of the first language that is the language of the original sentence to be translated, and associates one non-terminal signal with one or more terminal signals or non-terminal signals. Source language tree structure information including information on the tree structure, and a source language having at least one source language tree structure record having a source language tree structure probability that is information indicating the probability of the tree structure corresponding to the source language tree structure information Stores a tree structure model. A non-terminal symbol is a variable with a syntax label. The terminator is the word itself. The source language tree structure information includes the following types. For example, the source language tree structure information has two or more terminal symbols as child nodes, and the source language long unit phrase tree structure information which is information including a rule whose parent node is a non-terminal symbol, and one or more non-terminal symbols Source language syntax tree structure information, which is a rule indicating that the parent node is a non-terminal symbol, having zero or more terminal symbols as child nodes, and one terminal symbol as a child node, and the parent node is non-terminal There is source language word unit tree structure information which is information indicating a rule which is a symbol. The source language tree structure probability is a probability that a non-terminal symbol of a parent node generates a non-terminal symbol string or a terminal symbol string of a child node in the source language. Also, the information of the tree structure in which one or more terminal signals or non-terminal signals are associated with one non-terminal signal is a tree-structured information having a long unit phrase including a plurality of words and not including a non-terminal symbol as one node. Contains information. The data structure of the source language tree structure model does not matter. An example of the source language tree structure model will be described later. The source language tree structure model storage unit 11 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The target language tree structure model storage unit 12 is information regarding the tree structure of the second language that is the language of the sentence of the translation result, and is a tree in which one or more terminal signals or non-terminal signals are associated with one non-terminal signal. Target language tree structure information including structure information, and target language tree structure records having one or more target language tree structure records having target language tree structure probabilities that are information indicating the probability of the tree structure corresponding to the target language tree structure information Stores the model. The target language tree structure information has the following types. For example, the target language tree structure information has two or more terminal symbols as child nodes, and the target language long unit phrase tree structure information which is information including a rule whose parent node is a non-terminal symbol, and one or more non-terminal symbols And the target language syntax tree structure information, which is a rule indicating that the parent node is a non-terminal symbol, having zero or more terminal symbols as child nodes, and one terminal symbol as a child node, and the parent node is non-terminal There is target language word unit tree structure information which is information indicating a rule which is a symbol. The target language tree structure probability is a probability that a non-terminal symbol of a parent node generates a non-terminal symbol string or a terminal symbol string of a child node in the target language. Also, the information of the tree structure in which one or more terminal signals or non-terminal signals are associated with one non-terminal signal is a tree structure information including a long unit phrase including a plurality of words and not including a non-terminal symbol as one node. Contains information. The data structure of the target language tree structure model does not matter. An example of the target language tree structure model will be described later. The target language tree structure model storage unit 12 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The tree structure mapping model storage unit 13 is a tree having mapping information that is information indicating the correspondence between the source language tree structure information and the target language tree structure information, and mapping probability that is information indicating the correspondence probability indicated by the mapping information. A tree structure mapping model having one or more structure mapping records is stored. Here, the mapping probability is information indicating the probability that the source language tree structure information corresponds to the target language tree structure information. The data structure of the tree structure mapping model does not matter. An example of the tree structure mapping model will be described later. The tree structure mapping model storage unit 13 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The language model storage unit 14 stores a language model having one or more word appearance probabilities, which is information about the probability of word appearance in the second language. The word appearance probability is, for example, information on the probability related to the continuous appearance of two or more words in the second language. The word appearance probability is information indicating the probability that the second word appears after the first word in the second language, the second word, and the first word, for example. The language model includes, for example, one or more records having a first word, a second word, and a word appearance probability. The language model storage unit 14 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The accepting unit 15 accepts a sentence in the first language. The text input means may be anything such as a numeric keypad, keyboard, mouse or menu screen. The accepting unit 15 can be realized by a device driver for input means such as a numeric keypad and a keyboard, control software for a menu screen, and the like.

  The input sentence syntax analysis unit 16 parses the sentence received by the reception unit 15 and sequentially obtains tree structure information that is information about a part or all of the tree structure of the sentence. The input sentence syntax analysis unit 16 normally parses a sentence using the source language tree structure model in the source language tree structure model storage unit 11. However, although the input sentence syntax analysis unit 16 preferably uses a source language tree structure model, the sentence may be parsed using another tree structure model, or the sentence may be parsed by other means. You may analyze. In addition, since the process which parses a sentence and acquires tree structure information sequentially is a process in a well-known technique, detailed description is abbreviate | omitted. The input sentence syntax analysis unit 16 can usually be realized by an MPU, a memory, or the like. The processing procedure of the input sentence syntax analysis unit 16 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

The tree structure mapping probability calculation unit 17 obtains one or more source language tree structure probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit 16 from the source language tree structure model storage unit 11, and the input sentence syntax analysis unit One or more target language tree structure information corresponding to the tree structure information obtained by 16 and one or more mapping probabilities are obtained from the tree structure mapping model storage unit 13 and 1 corresponding to each of the one or more target language tree structure information. The above target language tree structure probabilities are acquired from the target language tree structure model storage unit 12, and based on the acquired target language tree structure information, one or more words of two or more words constituting the target language tree structure information The appearance probability is acquired from the language model storage unit 14, and based on the acquired source language tree structure probability, the acquired mapping probability, the acquired target language tree structure probability, and the acquired word appearance probability, To calculate the evaluation value of Bunki. The tree structure mapping probability calculation unit 17 can usually be realized by an MPU, a memory, or the like. The processing procedure of the tree structure mapping probability calculation unit 17 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).
The source language tree structure probability acquisition unit 171 acquires one or more source language tree structure probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit 16 from the source language tree structure model storage unit 11.
The mapping probability acquisition unit 172 acquires one or more target language tree structure information and one or more mapping probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit 16 from the tree structure mapping model storage unit 13.

  The target language tree structure probability acquisition unit 173 acquires one or more target language tree structure probabilities corresponding to each of the one or more target language tree structure information from the target language tree structure model storage unit 12. The one or more target language tree structure information here is the target language tree structure information acquired by the mapping probability acquisition means 172.

  Based on the acquired target language tree structure information, the word appearance probability acquisition unit 174 acquires one or more word appearance probabilities of two or more words constituting the target language tree structure information from the language model storage unit 14. The word appearance probability acquisition unit 174 may acquire, for example, the probability of a word bigram model described later.

  The evaluation value calculation means 175 calculates an evaluation value of the output syntax tree based on the source language tree structure probability, the mapping probability, the target language tree structure probability, and the word appearance probability. The evaluation value calculation means 175 calculates a translation model probability by a product of the source language tree structure probability, mapping probability, and target language tree structure probability, calculates a language model probability by a product of one or more word appearance probabilities, and It is preferable to calculate the evaluation value based on the translation model probability and the language model probability. Furthermore, it is preferable that the evaluation value calculation means 175 calculates the evaluation value by the product of the translation model probability and the language model probability.

The source language tree structure probability acquisition unit 171, the mapping probability acquisition unit 172, the target language tree structure probability acquisition unit 173, the word appearance probability acquisition unit 174, and the evaluation value calculation unit 175 can usually be realized by an MPU, a memory, or the like. The processing procedure of the source language tree structure probability acquisition unit 171 and the like is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).
The merge unit 18 synthesizes evaluation values corresponding to the same syntax tree in the two or more evaluation values calculated by the tree structure mapping probability calculation unit 17 and the output syntax tree corresponding to the evaluation value. Combining is, for example, calculating the sum of two or more evaluation values. Further, the composition is, for example, grouping (linking or the like) two or more evaluation values corresponding to one output syntax tree as one evaluation value. Specifically, the merge unit 18 temporarily stores a buffer (for example, an output word string list and a probability) using, for example, tree structure information (for example, a word string list) of the obtained output syntax tree as a key. Search). In this buffer, output tree structure information (word string list) and evaluation value (for example, probability) obtained by the tree structure mapping probability calculation unit 17 are stored. Then, the merging unit 18 determines whether or not the obtained tree structure information (for example, a word string list) exists in the buffer. When the merge unit 18 determines that the obtained tree structure information (for example, word string list) exists, the evaluation value (for example, probability) corresponding to the obtained tree structure information (for example, word string list). As a result, the obtained evaluation value (for example, probability) is added. When there are two or more evaluation values (for example, probabilities) corresponding to one word list, the sum of the translation model probabilities (evaluation values) is the translation probability (evaluation value) of the word list. For example, “(sum of two or more translation model probabilities) × language model probability” is the probability (evaluation value) of the word list. When the merging unit 18 determines that the obtained tree structure information (for example, word string list) does not exist, the obtained tree structure information (for example, word string list) and the obtained evaluation value (for example, , Probability) as a pair. The merging unit 18 can be usually realized by an MPU, a memory, or the like. The processing procedure of the merge unit 18 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The maximum likelihood sequence search unit 19 determines a part or all of the syntax tree of the second language to be output based on the evaluation value calculated by the tree structure mapping probability calculation unit 17, and based on the determined syntax tree Get output information that is part or all of a sentence in two languages. The maximum likelihood sequence search unit 19 determines a syntax tree that outputs a syntax tree having the maximum evaluation value calculated by the tree structure mapping probability calculation unit 17, and a part or all of the second language is based on the determined syntax tree. It is preferable to obtain output information that is a sentence. The maximum likelihood sequence search unit 19 can be usually realized by an MPU, a memory, or the like. The processing procedure of the maximum likelihood sequence search unit 19 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

The output unit 20 outputs a second language sentence having one or more pieces of output information acquired by the maximum likelihood sequence search unit 19. Here, the output is a concept including display on a display, printing on a printer, output by speech synthesis, transmission to an external device, and the like. The output unit 20 may be considered as including or not including an output device such as a display or a speaker. The output unit 20 can be realized by output device driver software, or output device driver software and an output device.
The first language and the second language may be anything such as Japanese, English, Chinese, Korean. However, the first language and the second language are different languages.
Next, the operation of the machine translation apparatus will be described with reference to the flowcharts of FIGS.
(Step S201) The reception unit 15 determines whether or not a sentence in the first language has been received. If a sentence is accepted, the process goes to step S202, and if no sentence is accepted, the process returns to step S201.

  (Step S202) The input sentence syntax analysis unit 16 uses the source language tree structure model (specifically, source language tree structure information) in the source language tree structure model storage unit 11 to parse the sentence received in step S201. Then, tree structure information, which is information about the tree structure of part or all of the sentence, is obtained. Note that the sentence parsing process in this step is a process of constructing a large syntax tree bottom-up by using previously read words one after another.

  (Step S203) The tree structure mapping probability calculation unit 17 determines whether or not a new subtree (new tree structure information) has been acquired in Step S202. If a new subtree can be acquired, the process goes to step S204. If a new subtree cannot be acquired, the process goes to step S205.

  (Step S204) The tree structure mapping probability calculation unit 17 performs an evaluation process. The evaluation process is a process for calculating an evaluation value of the output syntax tree. The evaluation value of the output syntax tree is information indicating the probability that output information (translated text) will be output. The process returns to step S202. When the process returns to step S202, in step S202, the previous word constituting the subtree created immediately before is also acquired, and processing is performed so as to form a large syntax tree by bottom-up. Details of the evaluation process will be described with reference to the flowchart of FIG.

  (Step S205) The maximum likelihood sequence search unit 19 determines whether or not the end of the sentence is reached. If it is the tail, it goes to step S206, and if it is not the tail, it goes to step S210. If the determination in step S205 is not the end, this sentence has two or more subtrees.

  (Step S206) The maximum likelihood sequence search unit 19 performs partial processing from a plurality of translation candidates (a set of one or more syntax trees in the second language) registered on the buffer in the parsing and evaluation processes described above. A subtree sequence having the smallest number of trees (syntax trees) is obtained.

(Step S207) The maximum likelihood sequence search unit 19 determines a word sequence having the maximum probability from the output word list of each subtree. In this case, if there are two or more identical output word lists, the sum of the probabilities corresponding to the output word list is calculated, and when determining the word string that maximizes the sum of the probabilities, the sum of the probabilities is calculated. To be compared.
(Step S208) The maximum likelihood sequence search unit 19 concatenates the word sequences of one or more subtrees determined in step S206.

(Step S209) The output unit 20 outputs the word string connected in step S208. The word string is a second language sentence (translation result) having one or more output information acquired by the maximum likelihood string search unit 19. The process ends.
(Step S210) The maximum likelihood sequence search unit 19 skips to the next word. Skipping to the next word is a process of shifting the pointer of the word to be parsed. The process returns to step S202.
Next, the evaluation process in step S204 will be described in detail using the flowchart in FIG.

(Step S301) The tree structure mapping probability calculation unit 17 obtains the rule θ _e of all target language tree structure models corresponding to the rule θ _f of the highest node of the constructed subtree from the target language tree structure model storage unit 12. get. That is, here, one or more (θ _f , θ _e ) pairs are acquired. The rules are source language tree structure information, target language tree structure information, and the like.
(Step S302) The tree structure mapping probability calculation unit 17 substitutes 1 for a counter i.

(Step S303) The tree structure mapping probability calculation unit 17 determines whether or not the i-th (θ _f , θ _e ) pair exists. If the i-th (θ _f , θ _e ) pair exists, the process goes to step S304, and if not, the process returns to the upper function.

(Step S304) The tree structure mapping probability calculation unit 17 constructs an output subtree (target language tree structure information) using θ _e of the i-th (θ _f , θ _e ) pair. The tree structure mapping probability calculation unit 17 constructs an output subtree as follows. That is, since the input sentence is analyzed bottom-up by the input sentence syntax analysis unit 16, the non-terminal symbols of the child nodes of θ _f in the input syntax tree are already known. Then, the output syntax tree (and the output word string list) is already known for the non-terminal symbols of the child nodes of θ _e . The output syntax tree of children nodes of theta _e, theta _e when embedding the non-terminal symbols, compares the syntax labels highest output syntax tree of children nodes, the syntax label non-terminal symbol, theta all child nodes of _e embedded only if they match the non-terminal symbols, the output syntax tree to the top of the parent node of the theta _e, tree mapping probability calculation unit 17 generates.
(Step S305) The tree structure mapping probability calculation unit 17 determines whether or not an output subtree has been constructed in step S304.
(Step S306) The tree structure mapping probability calculation unit 17 expands the constructed subtree into a word string list.
(Step S307) The tree structure mapping probability calculation unit 17 calculates the translation model probability of the subtree. The process of calculating the translation model probability will be described in detail with reference to the flowchart of FIG.
(Step S308) The tree structure mapping probability calculation unit 17 calculates a language model probability of the subtree. The process of calculating the language model probability will be described in detail with reference to the flowchart of FIG.

(Step S309) The tree structure mapping probability calculation unit 17 calculates the probability of the output word string using the translation model probability calculated in step S307 and the language model probability calculated in step S308. Normally, the tree structure mapping probability calculation unit 17 calculates the probability of the output word string by “translation model probability × language model probability”. The probability of the output word string is the probability that the input word string is translated into the output word string.
(Step S310) The merge unit 18 searches the buffer (the output word string list and the buffer in which the probability is temporarily stored) using the word string list obtained in step S306 as a key.

  (Step S311) The merge unit 18 determines whether or not the word string list obtained in step S306 exists in the buffer. If the word string list exists, the process goes to step S312. If the word string list does not exist, the process goes to step S313.

(Step S312) The merge unit 18 adds the probability obtained in step S309 as the probability corresponding to the word string list obtained in step S306. Go to step S314. When there are two or more probabilities corresponding to one word list, the sum of the translation model probabilities becomes the translation probability of the word list. Then, “(sum of two or more translation model probabilities) × language model probability” is the probability of the word list.
(Step S313) The merging unit 18 registers the word string list obtained in Step S306 and the probability obtained in Step S309 as a pair.
(Step S314) The tree structure mapping probability calculation unit 17 increments the counter i by one. The process returns to step S303.
Next, the process of calculating the translation model probability in step S307 will be described in detail using the flowchart of FIG.
(Step S401) The source language tree structure probability acquisition unit 171 uses the source language tree structure to acquire the source language tree structure probability from the source language tree structure model storage unit 11.

(Step S402) The mapping probability acquisition means 172 uses the source language tree structure (source language tree structure information) and the output subtree (target language tree structure information) to determine the mapping probability to the tree structure mapping model storage unit 13. Get from.
(Step S403) The target language tree structure probability acquisition unit 173 acquires the target language tree structure probability from the target language tree structure model storage unit 12 using the output subtree.

(Step S404) The evaluation value calculation means 175 calculates a translation model probability of the output syntax tree based on the source language tree structure probability, the mapping probability, and the target language tree structure probability. The evaluation value calculation means 175 preferably calculates the translation model probability by the product of the source language tree structure probability, the mapping probability, and the target language tree structure probability. Return to upper function.
In the flowchart of FIG. 4, it goes without saying that the order of acquiring the source language tree structure probability, the mapping probability, and the target language tree structure probability does not matter.

Next, the process of calculating the language model probability in step S308 will be described in detail using the flowchart of FIG. The language model of the language model probability calculated in this flowchart is a word bigram model.
(Step S501) The tree structure mapping probability calculation unit 17 substitutes 1 for a counter i.
(Step S502) The tree structure mapping probability calculation unit 17 determines whether or not the i-th word exists. If the i-th word exists, the process goes to step S503, and if the i-th word does not exist, the process goes to step S504.

(Step S503) The tree structure mapping probability calculation unit 17 acquires the (i-1) th word and the i-th word. When i is “1”, the tree structure mapping probability calculation unit 17 acquires “<S>” and the first word. “<S>” is a sentence start symbol indicating a start of a sentence. Go to step S505.
(Step S504) The tree structure mapping probability calculation unit 17 acquires the (i-1) th word, “</ S>”. “</ S>” is a sentence end symbol indicating the end of the sentence.

  (Step S505) The word appearance probability acquisition means 174 selects “<S>, i-th word”, “(i−1) -th word, i-th word” or “(i−1) -th word, < Information corresponding to “/ S>” is searched from the language model storage unit 14.

(Step S506) The word appearance probability acquisition means 174 selects “<S>, i-th word”, “(i−1) -th word, i-th word” or “(i−1) -th word, < It is determined whether or not “/ S>” exists in the language model storage unit 14. If it exists in the language model storage unit 14, the process goes to step S507, and if not, the process goes to step S508.
(Step S507) The word appearance probability acquisition unit 174 acquires the corresponding word appearance probability from the language model storage unit 14, and temporarily accumulates it. Go to step S509.
(Step S508) The word appearance probability acquisition means 174 temporarily stores the word appearance probability as a predetermined value. It is assumed that the predetermined value is held by the word appearance probability acquisition unit 174.
(Step S509) The tree structure mapping probability calculation unit 17 increments the counter i by 1.

(Step S510) The tree structure mapping probability calculation unit 17 determines whether it is the last. If it is the last, the process goes to step S511, and if it is not the last, the process returns to step S502. Whether or not it is the last can be determined by whether or not “</ S>” appears.
(Step S511) The tree structure mapping probability calculation unit 17 calculates a product of one or more temporarily stored word appearance probabilities. The product of the one or more word appearance probabilities is the language model probability. Return to upper function.

In the flowchart of FIG. 5, the word bigram model is used as the language model, but other language models such as the word trigram model and the part-of-speech trigram model may be used. Since other language models are well-known techniques, detailed description thereof is omitted.
The translation method of the machine translation apparatus described in the above flowchart is the following translation method.

なお、数式１において、Ｐ（ｅ）を言語モデル確率、Ｐ（ｆ｜ｅ）を逆方向翻訳モデル確率、Ｐ（ｅ｜ｆ）を順方向翻訳モデル確率、Ｐ（ｆ｜ｅ）Ｐ（ｅ｜ｆ）を単に翻訳モデル確率と言う。
つまり、本具体例において、翻訳結果の文章を取得することは、言語モデル確率と翻訳モデル確率の積が最大の単語列を取得することである。
以下、翻訳モデル確率の算出方法について、数式を用いて説明する。 That is, this translation method is a kind of statistical translation. Statistical translation is a method of performing translation by searching an output word string e that maximizes a probability from all possible combinations when an input word string f is given. The search result is expressed by the following Equation 1. In Equation 1, “argmax” indicates that an output word string that maximizes the probability is acquired. The output word string that maximizes the probability is a sentence (second language sentence) as a translation result.

In Equation 1, P (e) is a language model probability, P (f | e) is a backward translation model probability, P (e | f) is a forward translation model probability, and P (f | e) P (e | F) is simply referred to as a translation model probability.
That is, in this specific example, acquiring the translation result sentence means acquiring the word string having the maximum product of the language model probability and the translation model probability.
Hereinafter, a method for calculating the translation model probability will be described using mathematical expressions.

The statistical transfer of the syntax transfer method assumes a syntax tree of the source language and the target language (represented as F and E, respectively, and generates word strings f and e) as hidden variables in the translation model, and maps the tree structures to each other. By doing so, a translated sentence is generated.
In this machine translation measure, the translation model is decomposed into a source language (input) tree structure model, a target language (output) tree structure model, and a forward / reverse tree structure mapping model.
Specifically, it is expressed by Equation 2.

  Here, P (F | f) is a tree structure model probability of the source language. P (E | e) is a tree structure model probability of the target language. P (E | F) is a forward tree structure mapping model probability. P (F | E) is the backward tree structure mapping model probability. P (E | F) P (F | E) is simply referred to as a tree structure mapping model probability.

However, since the entire syntax tree cannot be converted, the conversion is performed in units of context-free grammar rules constituting the syntax tree. A context-free grammar rule unit is a rule for generating a non-terminal symbol or terminal symbol string of a child node for a non-terminal symbol of a parent node. For example, when the context-free grammar rule constituting the source language syntax tree E is θ _e and the context-free grammar rule constituting the target language syntax tree F is θ _f , the probabilities of the respective models are expressed by the following formulas 3 to 3. 5 is calculated. Formula 3 is a formula for calculating the tree structure model probability of the source language. Formula 4 is a formula for calculating the tree structure model probability of the target language. Expression 5 is an expression for calculating the tree structure mapping model probability.

さらに、本機械翻訳装置において、例えば、言語モデルは、単語ｂｉｇｒａｍモデルを用いる。つまり、単語ｂｉｇｒａｍモデルは、数式７により、確率が算出される。

In this machine translation apparatus, when translation is performed, the translation model probability is recursively calculated. For example, when the highest node N _i of a certain subtree includes j subtrees immediately below it, the following equation 6 is used.

Furthermore, in the machine translation apparatus, for example, a word bigram model is used as the language model. That is, the probability of the word bigram model is calculated according to Equation 7.

In the machine translation apparatus, the sentence (output word string list) in the second language is evaluated by the above-described mathematical formula, and, for example, an output word string list having the largest evaluation value is output. This output word string list is a translation result sentence (second language sentence).
Hereinafter, a specific operation of the machine translation apparatus in the present embodiment will be described.

FIG. 6 is an example of a source language tree structure model stored in the source language tree structure model storage unit 11 of the machine translation apparatus. The source language tree structure model includes one or more source language tree structure records having attribute values such as a name, source language tree structure information, and a source language tree structure probability “P (θ _f )”. The name is information for identifying a set of source language tree structure information and source language tree structure probabilities. The source language tree structure information has a parent node and a child node string. That is, the source language tree structure information constitutes a tree structure by parent nodes and child node strings. The source language tree structure information includes source language length unit phrase tree structure information which is information including a rule having two or more terminal symbols as child nodes and a parent node being a non-terminal symbol. The source language long unit phrase tree structure information is, for example, source language tree structure information corresponding to the names “SRC-102”, “SRC-103”, and “SRC-105”. For example, “3.67e-3” in the source language tree structure probability “P (θ _f )” of FIG. 6 is “3.67 × 10 ⁻³ ”. The source language tree structure probability “P (θ _f )” is a probability that a parent node generates a child node sequence in the source language.

FIG. 7 is an example of the target language tree structure model stored in the target language tree structure model storage unit 12 of the machine translation apparatus. The target language tree structure model includes one or more target language tree structure records having an attribute value of name, target language tree structure information, and target language tree structure probability “P (θ _e )”. The name is information for identifying a set of target language tree structure information and target language tree structure probability. Similar to the source language tree structure information, the target language tree structure information includes a parent node and a child node string. The target language tree structure probability “P (θ _e )” is the probability that the corresponding target language tree structure information will appear in the target language.

FIG. 8 is an example of a tree structure mapping model stored in the tree structure mapping model storage unit 13. The main tree structure mapping model has one or more tree structure mapping records having mapping information and mapping probability “P (θ _e | θ _f ) P (θ _f | θ _e )”. The mapping information includes information on the name of the target language and the name of the source language. That is, the mapping information is information indicating the correspondence between the tree structure of the target language identified by two names and the tree structure of the source language. The mapping probability “P (θ _e | θ _f ) P (θ _f | θ _e )” is the product of the forward tree structure mapping model probability and the backward tree structure mapping model probability.

FIG. 9 is an example of a language model stored in the language model storage unit 14. This language model is a word that is information indicating the first word (w ^i-1 _e ), the second word (w ⁱ _e ), and the probability that the second word will appear after the first word. It holds at least one word appearance probability having an appearance probability (P (w ⁱ _e | w ⁱ⁻¹ _e ), where “<S>” is a sentence start symbol and “</ S>” is It is a sentence end symbol.

In this situation, the machine translation apparatus performs syntax transfer machine translation as shown in FIG. 10, for example. Here, it is assumed that the machine translation apparatus performs Japanese-English translation. FIG. 10 parses a Japanese sentence “Do the bus leave at 12:00”, obtains a Japanese syntax tree F, and then maps the Japanese syntax tree F to the English syntax tree E? It shows that it goes and outputs an English sentence. In FIG. 10, “S” is a sentence, “NP” is a noun phrase, “NOUN” is a noun, “VP” is a verb phrase, “PP” is an adverb phrase, “V” is a verb, and “P” is a particle. “NUM” indicates a number. In FIG. 10, “SQ” indicates a question sentence, “NN” indicates a noun, “VB” indicates a verb, “IN” indicates a preposition, and “CD” indicates a number. (Is it correct?)
This machine translation apparatus includes a rule (long unit phrase tree structure information) that includes a plurality of words and does not include a non-terminal symbol in a translation model in such syntax transfer type machine translation.
Then, for example, it is assumed that the user inputs a sentence in the first language “Does the bus leave at 12:00” from the keyboard? And the reception part 15 receives the sentence of the first language "Is a bus appearing at 12:00?"
Hereinafter, in such a case, a process when the machine translation apparatus performs partial translation of “out at 12:00” will be described.
The machine translation apparatus holds the long unit phrase tree structure information shown in FIG. 11 and the translation model probabilities and language model probabilities corresponding thereto.

Next, the input sentence syntax analysis unit 16 parses “out at 12:00” using the source language tree structure model (specifically, source language tree structure information) of FIG. Then, the input sentence syntax analysis unit 16 detects that the SRC-002, SRC-003, and SRC-103 are applicable, and that any of them constitutes a VP for “12:00”, and SRC-002, SRC- 003, SRC-103 is acquired. The three subtrees “SRC-002”, “SRC-003”, and “SRC-103” are information (names) indicating the input subtree.
Then, the tree structure mapping probability calculation unit 17 performs the following processing on the three input subtrees (source language tree structure information).

First, the tree structure mapping probability calculation unit 17 sets the target language tree structure model corresponding to the rule θ _f (specified here by the name of the source language tree structure model) used for the top node of the input subtree. The rule θ _e (identified here by the name of the target language tree structure model) is obtained from the tree structure mapping model of FIG. Θ _e corresponding to SRC-002 is TRG-002, TRG-003, θ _e corresponding to SRC-003 is TRG-002, TRG-003, and θ _e corresponding to SRC-103 is TRG-104.

Here, the tree structure mapping probability calculation unit 17 sets the (θ _f , θ _e ) pairs “(SRC-002, TRG-002)”, “(SRC-002, TRG-003)”, “(SRC-003, TRG). -002) "" (SRC-003, TRG-003) "" (SRC-103, TRG-104) ".
Next, the tree structure mapping probability calculation unit 17 performs the following processing for each of the five sets (θ _f , θ _e ).

First, the tree structure mapping probability calculation unit 17 constructs a subtree using θ _e (steps 1 and 2 in FIG. 12). In this example, the three subtrees in FIG. 12 are constructed. (SRC-002, TRG-002 ) set, and in the case of (SRC-003, TRG-003 ) set, because there is no matching children subtrees in theta _e, subtree output is not built, And (1) in FIG. 12 is a subtree corresponding to the set of (SRC-002, TRG-003), (2) is a subtree corresponding to the set of (SRC-003, TRG-002), (3) Is a subtree corresponding to a set of (SRC-103, TRG-104).

  Next, the word string list of the output subtrees (1) to (3) in FIG. 12 is expanded to obtain a new output word list (step 3 in FIG. 12). Then, for each output word string, the translation model probability and the language model probability are calculated using Equation 6 and Equation 7. Here, the language model probability of the subtree is calculated without a sentence start / end symbol, and for a word string not shown in FIG. 9, for example, a probability “1.0e-7” is assigned.

For example, when the set of (SRC-002, TRG-003) is used, the output word string list of “12 o'clock” is “at 12 o'clock”, “at noon”, “to noon”, and “output” The output word string list of “leave” is “leave” and “start”. Therefore, when the combination is expanded and the translation model probability and the language model probability are calculated, the output word string list of FIG. 13 is obtained.
In FIG. 13, “input syntax tree” is an example of source language tree structure information. The “output syntax tree” is an example of “target language tree structure information”.
Similarly, in the case of the set of (SRC-003, TRG-002) and the set of (SRC-103, TRG-104), the output word string list of FIG. 14 is obtained.
Through the above processing, the tree structure mapping probability calculation unit 17 can obtain the output word string list of FIGS.

  Next, if the output word string list having the same input word string, the syntax label of the input subtree, and the syntax label of the output subtree has already been registered in the buffer, the tree structure mapping probability calculation unit 17 outputs both output words. Merge column lists. If the same output word string exists at the time of merging, the sum of translation model probabilities is calculated according to Equation 2 and registered in the buffer.

Here, for example, when the size of the output word string list exceeds a certain value (denoted as N, assuming 3 here) as a result of merging, the tree structure mapping probability calculation unit 17 calculates the translation model probability as Only the top N products of language model probabilities are left and registered. Such processing enables high-speed translation processing.
In the case of the above example, as a result of merging, the output word string list shown in FIG. 15 is obtained. For example, only the top three are registered in the buffer (step 4 in FIG. 12).

As a result of the above processing, a new partial translation result of “Exit at 12 o'clock” is obtained. By recursively repeating this for the entire input sentence, a partial translation result of the input sentence “Is the bus exit at 12 o'clock” is obtained. be able to. In the final output word string list, the language model probabilities are recalculated including the sentence start and end symbols, and the output word string with the maximum product with the translation model probability is output as the translation result of the input sentence. To do.
As described above, according to the present embodiment, translation can be performed without distinguishing words and phrases (a plurality of words).
Further, according to the present embodiment, the order of phrases and words can be hierarchically changed, and a syntactically correct translated sentence can be output. As a result, translation quality is improved.

  Further, according to the present embodiment, even when a plurality of candidates are obtained in the syntax tree, based on the probabilities based on the source language / target language tree structure model and tree structure mapping model automatically obtained from the corpus. An optimal output word string can be constructed. Therefore, a thesaurus or the like is not necessary.

  Further, according to the specific example of the present embodiment, the example of the process of step S206 (a process of selecting an output subtree having the smallest number of subtrees) has not been described. Such processing is, for example, the following processing. When "I'm sorry, does the bus leave at 12:00" is entered, the two subtrees with the minimum number of subtrees are "I'm sorry" and "Do the bus leave at 12:00"? Then, output word strings such as “excuse me”, “will the bus leave at 12 o'clock”, and “will the bus leave at noon” are obtained. Then, when an output word string having the maximum probability is obtained from each subtree and connected and output, “excuse me will the bus leave at 12 o'clock” is obtained.

Furthermore, in the present embodiment, even when there are a large number of output word strings, only word strings having higher probabilities are left out of the input word strings and the output word strings having the same syntax label of the input / output subtree. If done, candidates can be reduced appropriately and the translation speed is improved.
In the specific example of the present embodiment, the word bigram model is used as the language model, but other language models such as the word trigram model and the part-of-speech trigram model may be used.

  Furthermore, the processing in the present embodiment may be realized by software. Then, this software may be distributed by software download or the like. Further, this software may be recorded on a recording medium such as a CD-ROM and distributed. Note that the software that implements the information processing apparatus according to the present embodiment is the following program. In other words, this program is information about the tree structure of the first language, which is the language of the original sentence to be translated, on the computer, and is a tree with a long unit phrase consisting of a plurality of words and not including a nonterminal symbol as one node. Source language tree structure information including one or more source language tree structure records having source language tree structure probabilities that are information indicating the probability of the tree structure corresponding to the source language tree structure information. Information on the tree structure of the second language, which is the language of the sentence of the translation result, storing the model, and information on the tree structure with a long unit phrase consisting of multiple words and not including non-terminal symbols as one node A target language tree structure model having at least one target language tree structure record having a target language tree structure probability that is information indicating the probability of the tree structure corresponding to the target language tree structure information. A tree structure mapping record having mapping information which is information indicating the correspondence between the source language tree structure information and the target language tree structure information, and mapping probability which is information indicating the correspondence probability indicated by the mapping information And a language model having one or more word appearance probabilities having word appearance probabilities that are information on probabilities related to successive appearances of two or more words in the second language, An accepting step for accepting a sentence in a first language; an input sentence syntactic analyzing step for sequentially obtaining a tree structure information that is information on a tree structure of a part or all of the sentence by parsing the sentence accepted in the accepting step; , One or more source language tree structure probabilities corresponding to the tree structure information obtained in the input sentence syntax analysis step are acquired, and the input sentence syntax analysis One or more target language tree structure information corresponding to the tree structure information obtained in step 1 and one or more mapping probabilities are acquired, and one or more target language tree structures corresponding to each of the one or more target language tree structure information are acquired. Probability is acquired, one or more word appearance probabilities of two or more words constituting the target language tree structure information are acquired based on the acquired target language tree structure information, and the acquired source language tree structure probability is acquired A tree structure mapping probability calculation step for calculating an evaluation value of an output syntax tree based on the acquired mapping probability, the acquired target language tree structure probability, and the acquired word appearance probability; and the tree structure mapping probability Based on the evaluation value calculated in the calculation step, a part or all of the syntax tree of the second language to be output is determined, and a part or all of the text of the second language is determined based on the determined syntax tree. And a program for executing an output step of outputting a sentence in a second language having one or more pieces of output information acquired in the maximum likelihood sequence search step.

The tree structure mapping probability calculation step includes a source language tree structure probability acquisition step for acquiring one or more source language tree structure probabilities corresponding to the tree structure information obtained in the input sentence syntax analysis step, and the input sentence syntax One or more target language tree structure information corresponding to the tree structure information obtained in the analysis step, a mapping probability acquisition step for acquiring one or more mapping probabilities, and one or more corresponding to each of the one or more target language tree structure information A target language tree structure probability acquisition step of acquiring the target language tree structure probability of the target language tree, and one or more word appearances of two or more words constituting the target language tree structure information based on the acquired target language tree structure information A word appearance probability acquisition step of acquiring a probability from the language model storage unit; the source language tree structure probability; the mapping probability; the target language tree structure probability; Based on the fine said word occurrence probabilities may be provided with an evaluation value calculation step of calculating an evaluation value of the syntax tree output.
Preferably, the evaluation value calculating step calculates a product of the source language tree structure probability, the mapping probability, the target language tree structure probability, and the word appearance probability, and uses the product as an evaluation value of an output syntax tree. It is.

  Further, in the maximum likelihood sequence search step, a syntax tree having the maximum evaluation value calculated in the tree structure mapping probability calculation step is determined as an output syntax tree, and one of the second languages is determined based on the determined syntax tree. It is preferable to obtain output information that is a part or all of a sentence.

  In each of the above embodiments, each process (each function) may be realized by centralized processing by a single device (system), or by distributed processing by a plurality of devices. May be.

  FIG. 16 shows the external appearance of a computer that executes the programs described in this specification to realize the information processing apparatuses according to the various embodiments described above. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 16 is an overview diagram of the computer system 300, and FIG. 17 is a block diagram of the system 300.

  In FIG. 16, a computer system 300 includes a computer 301 including a FD (Flexible Disk) drive and a CD-ROM (Compact Disk Read Only Memory) drive, a keyboard 302, a mouse 303, and a monitor 304.

  In FIG. 17, in addition to the FD drive 3011 and the CD-ROM drive 3012, the computer 301 includes a CPU (Central Processing Unit) 3013, a bus 3014 connected to the CPU 3013, the CD-ROM drive 3012 and the FD drive 3011, and a boot. A ROM (Read-Only Memory) 3015 for storing a program such as an up program, and a RAM (Random Access Memory) connected to the CPU 3013 for temporarily storing instructions of the application program and providing a temporary storage space 3016 and a hard disk 3017 for storing application programs, system programs, and data. Although not shown here, the computer 301 may further include a network card that provides connection to a LAN.

  A program that causes the computer system 300 to execute the functions of the information processing apparatus according to the above-described embodiment is stored in the CD-ROM 3101 or FD 3102, inserted into the CD-ROM drive 3012 or FD drive 3011, and further stored in the hard disk 3017. May be forwarded. Alternatively, the program may be transmitted to the computer 301 via a network (not shown) and stored in the hard disk 3017. The program is loaded into the RAM 3016 at the time of execution. The program may be loaded directly from the CD-ROM 3101, the FD 3102 or the network.

The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 301 to execute the functions of the information processing apparatus according to the above-described embodiment. The program only needs to include an instruction portion that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 300 operates is well known and will not be described in detail.
Further, the computer that executes the program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.
The present invention is not limited to the above-described embodiments, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, the machine translation apparatus according to the present invention has an effect that high-quality and high-speed translation is possible, and is useful as a machine translation apparatus or the like.

Block diagram of machine translation apparatus in embodiment 1 Flow chart for explaining the operation of the machine translation apparatus A flowchart for explaining the operation of the evaluation process A flowchart for explaining the process of calculating the translation model probability Flowchart explaining processing for calculating same language model probability Diagram showing an example of the original language tree structure model Diagram showing an example of the same language tree structure model Diagram showing an example of the same tree structure mapping model Diagram showing an example of the same language model Diagram explaining machine translation of the same syntax transfer method The figure which shows the initial state of the same machine translation device The figure which shows the subtree which the same machine translation device constructs The figure which shows the output word sequence list which the machine translation apparatus comprises The figure which shows the output word sequence list which the machine translation apparatus comprises The figure which shows the output word sequence list which the machine translation apparatus comprises External view of a computer that implements the machine translation device Block diagram of the machine translation device

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Source language tree structure model storage part 12 Target language tree structure model storage part 13 Tree structure mapping model storage part 14 Language model storage part 15 Reception part 16 Input sentence syntax analysis part 17 Tree structure mapping probability calculation part 18 Maximum likelihood sequence search part 19 Output unit 171 Source language tree structure probability acquisition means 172 Mapping probability acquisition means 173 Target language tree structure probability acquisition means 174 Word appearance probability acquisition means 175 Evaluation value calculation means

Claims (6)

Source language tree structure that is information about the tree structure of the first language that is the language of the original sentence to be translated, and that contains information about the tree structure that consists of a plurality of words and does not include a non-terminal symbol and has a long unit phrase as one node Information and a source language tree storing a source language tree structure model having one or more source language tree structure records having source language tree structure probabilities that are information indicating the probability of the tree structure corresponding to the source language tree structure information A structural model storage;
Information on the tree structure of the second language, which is the language of the sentence of the translation result, the target language tree structure information including information on the tree structure consisting of a plurality of words and a long unit phrase not including a non-terminal symbol as one node; , A target language tree structure model storing a target language tree structure model having one or more target language tree structure records having a target language tree structure probability that is information indicating the probability of the tree structure corresponding to the target language tree structure information A storage unit;
Tree structure having one or more tree structure mapping records having mapping information which is information indicating correspondence between source language tree structure information and target language tree structure information and mapping probability which is information indicating correspondence probability indicated by the mapping information A tree structure mapping model storage unit storing a mapping model;
A language model storage unit storing a language model having one or more word appearance probabilities, which is information on the probability of word appearance in the second language;
A reception unit that accepts sentences in the first language;
An input sentence syntax analysis unit that sequentially parses the sentence received by the reception unit and sequentially obtains tree structure information that is information on a part or all of the tree structure of the sentence;
One or more source language tree structure probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit are acquired from the source language tree structure model storage unit and correspond to the tree structure information obtained by the input sentence syntax analysis unit One or more target language tree structure information and one or more mapping probabilities are acquired from the tree structure mapping model storage unit, and one or more target language tree structure probabilities corresponding to each of the one or more target language tree structure information are obtained. Acquired from the target language tree structure model storage unit, and stores one or more word appearance probabilities of two or more words constituting the target language tree structure information based on the acquired target language tree structure information in the language model Based on the acquired source language tree structure probability, the acquired mapping probability, the acquired target language tree structure probability, and the acquired word appearance probability. A tree structure mapping probability calculation unit for calculating a value,
Based on the evaluation value calculated by the tree structure mapping probability calculation unit, a part or all of the second language to be output is determined, and based on the determined syntax tree, part or all of the second language is determined. A maximum likelihood sequence search unit for acquiring output information that is a sentence;
A machine translation apparatus comprising: an output unit that outputs a sentence in a second language having one or more pieces of output information acquired by the maximum likelihood sequence search unit. The tree structure mapping probability calculator
Source language tree structure probability acquisition means for acquiring one or more source language tree structure probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit from the source language tree structure model storage unit;
Mapping probability acquisition means for acquiring one or more target language tree structure information and one or more mapping probabilities corresponding to the tree structure information obtained by the input sentence syntax analysis unit from the tree structure mapping model storage unit;
Target language tree structure probability acquisition means for acquiring one or more target language tree structure probabilities corresponding to each of the one or more target language tree structure information from the target language tree structure model storage unit;
Based on the acquired target language tree structure information, word appearance probability acquisition means for acquiring one or more word appearance probabilities of the two or more words constituting the target language tree structure information from the language model storage unit;
2. The machine according to claim 1, further comprising: an evaluation value calculation means for calculating an evaluation value of an output syntax tree based on the source language tree structure probability, the mapping probability, the target language tree structure probability, and the word appearance probability. Translation device. In the syntax tree of the output corresponding to the two or more evaluation values calculated by the tree structure mapping probability calculation unit and the evaluation value, the merge structure unit further includes a merging unit that combines the evaluation values corresponding to the same syntax tree,
The maximum likelihood sequence search unit includes:
Based on the evaluation value calculated by the tree structure mapping probability calculation unit and the evaluation value obtained by combining by the merge unit, a part or all of the syntax tree of the second language to be output is determined, and the determined syntax tree The machine translation apparatus according to claim 1, wherein output information that is a part or all of a sentence in the second language is acquired based on. The evaluation value calculation means includes
4. The machine according to claim 2, wherein a product of the source language tree structure probability, the mapping probability, the target language tree structure probability, and the word appearance probability is calculated, and the product is used as an evaluation value of an output syntax tree. Translation device. The maximum likelihood sequence search unit includes:
A syntax tree having the maximum evaluation value calculated by the tree structure mapping probability calculation unit is determined as a syntax tree to be output, and output information that is a sentence of a part or all of the second language based on the determined syntax tree. The machine translation apparatus according to claim 1, which is acquired. On the computer,
Source language tree structure that is information about the tree structure of the first language that is the language of the original sentence to be translated, and that contains information about the tree structure that consists of a plurality of words and does not include a non-terminal symbol and has a long unit phrase as one node A source language tree structure model having at least one source language tree structure record having a source language tree structure probability that is information and information indicating the probability of the tree structure corresponding to the source language tree structure information;
Information on the tree structure of the second language, which is the language of the sentence of the translation result, the target language tree structure information including information on the tree structure consisting of a plurality of words and a long unit phrase not including a non-terminal symbol as one node; A target language tree structure model having at least one target language tree structure record having a target language tree structure probability, which is information indicating the probability of the tree structure corresponding to the target language tree structure information,
Tree structure having one or more tree structure mapping records having mapping information which is information indicating correspondence between source language tree structure information and target language tree structure information and mapping probability which is information indicating correspondence probability indicated by the mapping information Contains the mapping model,
Storing a language model having one or more word appearance probabilities having a word appearance probability, which is information on the probability of consecutive occurrences of two or more words in the second language;
A reception step for accepting sentences in the first language;
Parsing the sentence received in the reception step, the input sentence syntax analysis step to sequentially obtain the tree structure information that is information about a part or all of the tree structure of the sentence;
One or more source language tree structure probabilities corresponding to the tree structure information obtained in the input sentence syntax analysis step are acquired, and one or more target language tree structure information corresponding to the tree structure information obtained in the input sentence syntax analysis step And one or more mapping probabilities, one or more target language tree structure probabilities corresponding to each of the one or more target language tree structure information, and the target language tree structure information based on the acquired target language tree structure information One or more word appearance probabilities of two or more words constituting language tree structure information are acquired, the acquired source language tree structure probability, the acquired mapping probability, the acquired target language tree structure probability, and the acquisition A tree structure mapping probability calculation step for calculating an evaluation value of the output syntax tree based on the word appearance probability,
Based on the evaluation value calculated in the tree structure mapping probability calculation step, a part or all of the second language to be output is determined, and based on the determined syntax tree, part or all of the second language is determined. A maximum likelihood sequence search step for obtaining output information that is a sentence;
A program for executing an output step of outputting a sentence in a second language having one or more output information acquired in the maximum likelihood sequence search step.

Images