JP2007072663A - Example translation device and example translation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an example translation device capable of performing translation from an original language to an objective language by use of a parallel translation example corpus. <P>SOLUTION: This example translation device has: a reception part 11 receiving document information; a parallel translation example corpus storage part 12 storing the parallel translation example corpus associating and having a plurality of original language examples and a plurality of objective language examples that are translations of the original language examples in the objective language; a selection part 13 selecting at least two original language examples similar to the document information from the plurality of original language examples of the parallel translation example corpus; a translation candidate information production part 14 producing translation candidate information showing a candidate of the translation of the document information by use of at least two objective language examples corresponding to at least the two selected original language examples; a translation document information production part 15 selecting the candidate of the translation satisfying a prescribed condition from the translation candidate information, and producing translation document information showing the selected candidate of the translation; and an output part 16 outputting the translation document information. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an example translation apparatus that performs machine translation.

With the progress of computers and the increase of language resources, natural language processing technology using corpus has been developed. Also in machine translation, which is one of natural language processing, an example translation using a parallel translation example has been proposed and an approach using a request and an example has been studied (for example, see Non-Patent Documents 1 to 5).
Makoto Nagao. A framework of a mechanical translation between Japan and English by analog principal. Proc. of the international NATO symposium on Artificial and human intelligence, pp. 173-180, 1984. Satoshi Sato, Makoto Nagao. Toward memory-based translation. Procs of the 13th conference on Computational linguistics, pp. 247-252, 1990. Satoshi Sato. MBT2: a method for combining fragments of examples in example-based translation. Artificial intelligence, Vol. 75, pp. 31-49, 1995. Eiichiro Sumita. Example-based machine translation using DP-matching between word sequences. DDMT workshop of 39th ACL, pp. 1-8, 2001 Eiji Aramaki et al., Translation selection using Japanese-English alignment confidence and Japanese similarity for example-based translation. Natural language processing, Vol. 11, no. 1, pp. 107-123, 2004.

  In the conventional method, it is assumed that word alignment and phrase alignment are established between the source language and the target language. However, in an actual bilingual corpus, especially a spoken language corpus, alignment is often difficult due to omissions and idiomatic expressions. In addition, in the case of a sentence using different expressions on the source language side even though the two sentences are the same on the target language side, the example may not be used well.

  The present invention has been made to solve the above-described problems, and an example that enables translation from a source language to a target language using a bilingual example corpus without necessarily using word alignment or phrase alignment. An object is to provide a translation device or the like.

  In order to achieve the above object, an example translation apparatus according to the present invention differs from a source language in a receiving unit that receives document information that is information indicating a document in the source language, and a plurality of source language examples that are examples in the source language. From the bilingual example corpus storage unit storing the bilingual example corpus that is information having a plurality of target language examples that are translations of the source language examples in the target language, and a plurality of source language examples of the bilingual example corpus, A selection unit that selects two or more source language examples similar to the document indicated by the document information received by the reception unit, and two or more target language examples corresponding to the two or more source language examples selected by the selection unit are used. A translation candidate information creation unit that creates translation candidate information that is information indicating translation candidates of document information received by the reception unit, and a translation candidate information created by the translation candidate information creation unit. Select a translation candidate that satisfies the condition, and create a translation document information creation unit that creates translation document information that is information indicating the selected translation candidate, and output the translation document information created by the translation document information creation unit Part.

  With such a configuration, machine translation from the source language to the target language can be executed using the parallel translation example corpus. Therefore, machine translation can be performed without necessarily using word alignment or phrase alignment.

  In the example translation apparatus according to the present invention, the selection unit calculates a similarity between the plurality of source language examples and the document indicated by the document information received by the reception unit, and selects two or more source language examples selected. The source language example having a high degree of similarity may be selected so that the entire document indicated by the document information can be covered with words common to the document indicated by the document information.

  With such a configuration, it is possible to select a source language example closely related to the document indicated by the document information, and to perform more appropriate translation by performing translation using the target language example corresponding to the source language example. Can be.

In the example translation apparatus according to the present invention, the selection unit may use a ratio of words common to two documents as a similarity between the two documents.
In the example translation apparatus according to the present invention, the selection unit may use a DP (Dynamic Programming) matching value of two documents as the similarity between the two documents.

  In the example translation apparatus according to the present invention, the selection unit may include a document indicated by the document information received by the reception unit and a source of a parallel translation example corpus when the source language is a language that is not divided into words. A process of dividing a language example into words may be performed.

  With such a configuration, even if the source language is a language that is not segmented for each word such as Japanese or Chinese, the document or source language example indicated by the document information is segmented for each word. Thus, the similarity between the two can be calculated.

In the example translation apparatus according to the present invention, the translation candidate information creation unit obtains two or more target language examples corresponding to the two or more source language examples selected by the selection unit from the bilingual example corpus, A translation candidate indicating the adjacent relationship between the common node and the single node, with the common word between the two or more target language examples as one common node and the other words in the acquired two or more target language examples as one single node Information may be created.
With such a configuration, translation candidates can be created with possible combinations of words in the target language example.

Moreover, in the example translation apparatus according to the present invention, the translation candidate information creation unit has a two-gram probability of an arrangement of two words related by an adjacent relationship other than the adjacent relationship of the node indicated by the translation candidate information equal to or higher than a threshold value. In this case, an adjacency indicating the arrangement of the two words may be added to the translation candidate information.
With such a configuration, the number of translation candidates can be increased, and the possibility of obtaining a correct translation result can be further increased.

  In the example translation apparatus according to the present invention, the translation candidate information creation unit corresponds to two or more source language examples selected by the selection unit when the target language is a language that is not divided for each word. Two or more target language examples may be divided into words.

  With such a configuration, even if the target language is a language that is not segmented for each word such as Japanese or Chinese, the common node and Translation candidate information can be created by the adjacency relationship of single nodes.

In the example translation apparatus according to the present invention, the translation document information creation unit calculates a product of N-gram probabilities for each translation candidate indicated by the translation document information, and the translation product having the maximum N-gram probability product is calculated. The candidate may be selected as a translation candidate that satisfies a predetermined condition.
With such a configuration, the most probable translation candidate is used as the translation result, and the possibility of obtaining a correct translation result can be increased.

  In the example translation apparatus according to the present invention, the parallel example corpus includes a component alignment that is an alignment of components that constitute a sentence, and the translated document information creation unit is received by the receiving unit. Translation candidates that satisfy a predetermined condition may be selected based on the presence or absence of alignment by the component alignment between the document indicated by the document information and the translation candidate document.

  With such a configuration, it is possible to increase the possibility of obtaining a correct translation result by creating translation document information so that the alignment result is reflected for the aligned components. .

In the example translation apparatus according to the present invention, the translation document information creation unit may correspond to a translation candidate document and the source language example for one of the two or more source language examples selected by the selection unit. If the similarity to the language example is not within a predetermined range with respect to the similarity between the document indicated by the document information received by the receiving unit and the source language example, the translation candidate may not be selected. .
With such a configuration, it is possible to prevent selection of inappropriate translation candidates as a result of translation.

In the example translation apparatus according to the present invention, the translated document information creation unit may use a word ratio common to the two documents as the similarity between the two documents.
In the example translation apparatus according to the present invention, the translated document information creation unit may use a value of DP (Dynamic Programming) matching between two documents as the similarity between the two documents.

  According to the example translation apparatus and the like according to the present invention, machine translation from the source language to the target language can be performed using the parallel translation example corpus.

  Hereinafter, an example translation apparatus according to the present invention will be described using embodiments. Note that, in the following embodiments, the components given the same reference numerals are the same or equivalent, and repetitive description may be omitted.

(Embodiment 1)
An example translation apparatus according to Embodiment 1 of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration of an example translation apparatus 1 according to the present embodiment. In FIG. 1, an example translation apparatus 1 according to the present embodiment includes a reception unit 11, a parallel example corpus storage unit 12, a selection unit 13, a translation candidate information creation unit 14, a translation document information creation unit 15, and an output. Part 16.

  The accepting unit 11 accepts document information. Here, the document information is information indicating a source language document. The document information is information indicating a document of one sentence or two sentences, for example. In the present embodiment, it is assumed that the document information is information indicating a single sentence document. For example, the reception unit 11 may receive document information input from an input device (for example, a keyboard, a mouse, or a touch panel), or may receive document information transmitted via a wired or wireless communication line. Alternatively, document information read from a predetermined recording medium (for example, an optical disk, a magnetic disk, a semiconductor memory, etc.) may be received. In the present embodiment, it is assumed that the accepting unit 11 reads document information stored in the recording medium 21. Here, the recording medium 21 can be realized by, for example, a semiconductor memory, a magnetic disk, an optical disk, or the like. The recording medium 21 may be included in the example translation apparatus 1 or may be outside the example translation apparatus 1. The accepting unit 11 may or may not include a device for accepting (for example, a modem or a network card). In addition, the reception unit 11 may be realized by hardware, or may be realized by software such as a driver that drives a predetermined device.

  The translation example corpus storage unit 12 stores a translation example corpus. Here, the bilingual example corpus is information having a plurality of source language examples that are examples in the source language and a plurality of target language examples that are translations of the source language examples in the target language in association with each other. The example translation apparatus 1 performs translation from a source language to a target language. The target language is a language different from the original language. The language of the source language and the target language is not limited. The source language and the target language are, for example, Japanese, Chinese, Korean, English, French, German, Russian, and the like. In this embodiment, a case where the source language is Japanese and the target language is English will be described. The parallel translation example corpus may be, for example, a BTEC (Basic Travel Expression Corpus) that is a corpus of a basic expression of travel conversation created by the applicant (ATR), or may be another corpus. The bilingual example corpus has 10,000 or more, usually several tens of thousands, pairs of source language examples and target language examples. Needless to say, for example translation, it is better to have a larger number of pairs of source language examples and target language examples. The bilingual example corpus storage unit 12 can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, or the like). The process of storing the parallel example corpus in the parallel example corpus storage unit 12 does not matter. For example, the parallel example corpus may be stored in the parallel example corpus storage unit 12 via a recording medium, and the parallel example corpus transmitted via a communication line or the like is stored in the parallel example corpus storage unit 12. Alternatively, the bilingual example corpus input via the input device may be stored in the bilingual example corpus storage unit 12. The storage in the parallel example corpus storage unit 12 may be temporary storage in the RAM of the parallel example corpus read from an external storage device or the like, or may be long-term storage in a hard disk or the like.

The selection unit 13 selects two or more source language examples similar to the document indicated by the document information received by the receiving unit 11 from a plurality of source language examples of the parallel translation example corpus. As a more specific process, the selection unit 13 may select a source language example as follows. That is, the selection unit 13 first calculates the similarity between a plurality of source language examples and the document indicated by the document information received by the receiving unit 11. Then, the selection unit 13 can use the common source language examples having high similarity so that the entire document indicated by the document information can be covered by words common to the two or more selected source language examples and the document indicated by the document information. May be selected. Note that the selection unit 13 may use the ratio of words common to the two documents as the similarity between the two documents. The selection unit 13 may calculate the ratio of common words using, for example, the following equation.

  Note that the ratio of words common to two documents may be calculated by an expression other than Expression (1). For example, an equation in which the coefficient “2” of the numerator in the equation (1) is “1” may be used. Further, the selection unit 13 may use a DP (Dynamic Programming) matching value between two documents as the similarity between the two documents. Since DP matching is conventionally known, its description is omitted. Note that when the source language is a language that is not divided for each word, the selection unit 13 may perform processing for dividing the document or the source language example indicated by the document information received by the reception unit 11 for each word. This process is performed by morphological analysis, for example. Languages that are not divided into words are, for example, Japanese and Chinese.

  The translation candidate information creation unit 14 creates translation candidate information using two or more target language examples corresponding to the two or more source language examples selected by the selection unit 13. Here, the translation candidate information is information indicating candidates for translation of the document information received by the receiving unit 11. Translation candidate information is created by finding possible combinations of word sequences in the target language example. As a more specific process, the translation candidate information creation unit 14 may create translation candidate information as follows. The translation candidate information creation unit 14 acquires two or more target language examples corresponding to the two or more source language examples selected by the selection unit 13 from the parallel translation example corpus. The translation candidate information creation unit 14 sets the common word between the acquired two or more target language examples as one common node, and sets the other words in the acquired two or more target language examples as one single node. Translation candidate information indicating adjacency relationships between single nodes may be created. The “common word” may be a matching word or a similar word. In the present embodiment, it is assumed that the common word is a matching word. When the target language is a language that is not divided for each word, the translation candidate information creation unit 14 performs two or more target language examples corresponding to the two or more source language examples selected by the selection unit 13 for each word. You may perform the process divided into. This process is performed by morphological analysis, for example.

The translation document information creation unit 15 selects a translation candidate that satisfies a predetermined condition from the translation candidate information created by the translation candidate information creation unit 14, and creates translation document information that is information indicating the selected translation candidate. . The translated document information is information indicating a document obtained by translating the document information received by the receiving unit 11. As a more specific process, the translation document information creation unit 15 may create translation document information as follows. The translation document information creation unit 15 calculates a product of N-gram probabilities for each translation candidate indicated by the translation candidate information, and selects a translation candidate having the maximum N-gram probability product as a translation candidate satisfying a predetermined condition. You may choose. Here, the corpus used when calculating the N-gram probability may be a bilingual example corpus stored in the bilingual example corpus storage unit 12 or a corpus stored in a recording medium (not shown). . Further, “N” of the N-gram probability may be, for example, N = 2, N = 3, or a value greater than that. Moreover, the translation document information creation unit 15 may calculate the N-gram probability, or the translation document information creation unit 15 may use the N-gram probability calculated by another component, another device, or the like. Since the N-gram probability is conventionally known, the description thereof is omitted. The translation document information creation unit 15 determines whether the similarity between the candidate document for translation and the target language example corresponding to the source language example for any one of the two or more source language examples selected by the selection unit 13 is a reception unit. When the similarity between the document indicated by the document information received by the document 11 and the source language example is not within a predetermined range, the translation candidate may not be selected. Specifically, the document indicated by the document information receiving unit 11 has received the J _0, selector 13 selects the source language example J _{i (where,} i is an integer .M of M 1 has selecting section 13 The number of selected source language examples is an integer of 2 or more), the target language example corresponding to the source language example J _i is E _i (where i is an integer from 1 to M), and the translation candidate is E _0. Then, the translation document information creation unit 15 may select any one of translation candidates E ₀ satisfying the following expression for each of i from 1 to M.

  The translated document information creation unit 15 may use the ratio of words common to the two documents (for example, the above formula (1)) as the similarity between the two documents, or the DP of the two documents. A matching value may be used.

  The output unit 16 outputs the translation document information created by the translation document information creation unit 15. Here, the output may be, for example, display on a display device (for example, a CRT or a liquid crystal display), transmission via a communication line to a predetermined device, printing by a printer, or output to a recording medium. It may be accumulated. In the present embodiment, it is assumed that the output unit 16 accumulates translated document information in the recording medium 22. Here, the recording medium 22 can be realized by, for example, a semiconductor memory, a magnetic disk, an optical disk, or the like. The recording medium 22 may be included in the example translation apparatus 1 or may be outside the example translation apparatus 1. The output unit 16 may or may not include an output device (for example, a display device or a printer). The output unit 16 may be realized by hardware, or may be realized by software such as a driver that drives these devices.

  The bilingual example corpus storage unit 12 and the recording media 21 and 22 may be realized by the same recording medium or may be realized by different recording media. In the former case, the region of the recording medium in which the parallel example corpus is stored becomes the parallel example corpus storage unit 12.

  Next, the operation of the example translation apparatus 1 according to the present embodiment will be described using the flowchart of FIG. The flowchart in FIG. 2 is a process after the example translation is started.

  (Step S101) The receiving unit 11 determines whether document information has been received. If the document information is received, the process proceeds to step S102. If not, the process of step S101 is repeated until the document information is received.

  (Step S102) The selection unit 13 selects two or more source language examples similar to the document indicated by the document information received by the receiving unit 11 from the bilingual example corpus. Details of this process will be described later.

  (Step S103) The translation candidate information creation unit 14 uses the two or more target language examples corresponding to the two or more source language examples selected by the selection unit 13 to select translation candidate of the document information received by the receiving unit 11. The translation candidate information shown is created.

  (Step S104) The translation document information creation unit 15 selects a translation candidate satisfying a predetermined condition from the translation candidates indicated by the translation candidate information created by the translation candidate information creation unit 14, and translate document information indicating the translation candidate. Create Details of this process will be described later.

  (Step S105) The output unit 16 outputs the translation document information created by the translation document information creation unit 15. Then, the operation of the example translation apparatus 1 ends. In addition, when performing example translation about several document information, the example translation apparatus 1 may repeatedly perform the process from step S101 to step S105.

  Next, the operation of the example translation apparatus 1 according to this embodiment will be described using a specific example. In this specific example, it is assumed that the document indicated by the document information in the source language is “Please give me a bed ticket to Glasgow”. Further, the parallel example corpus stored in the parallel example corpus storage unit 12 is assumed to be as shown in FIG. In FIG. 3, in the bilingual example corpus, a Japanese source language example related to travel conversation is associated with an English target language example that is a translation thereof.

  First, it is assumed that the user specifies document information “Please give me a bed ticket to Glasgow” stored in the recording medium 21 and inputs an instruction to translate the document to the example translation apparatus 1 (this The input may be performed using, for example, a mouse). Then, the reception unit 11 reads the document information “Please give me a bed ticket to Glasgow” from the recording medium 21, and passes it to the selection unit 13 and the translation document information creation unit 15 (step S101).

  As shown in the flowchart of FIG. 4, the selection unit 13 selects a source language example (step S102 in the flowchart of FIG. 2). First, the flowchart of FIG. 4 will be described.

  (Step S201) The selection unit 13 calculates the similarity between the document indicated by the document information received by the reception unit 11 and the source language example included in the parallel translation example corpus. Note that, as described above, the document received by the receiving unit 11 when the source language is a language that is not divided into words and the morphological analysis is not performed on the document indicated by the source language document information or the source language example. A process of dividing the document indicated by the information into words is performed.

(Step S202) The selection unit 13 sorts the source language examples in descending order of similarity based on the calculated similarity.
(Step S203) The selection unit 13 sets the counter I to “1”.

  (Step S204) The selection unit 13 selects the Ith source language example. Here, the I-th means the I-th from the ones with a high similarity as a result of sorting in step S202.

  (Step S205) The selection unit 13 determines whether the document information can be covered by the source language examples selected so far. In other words, the selection unit 13 determines whether or not the entire document indicated by the document information can be covered with words common to the source language examples selected so far and the document indicated by the document information. Here, the word common to the document A and the document B may be common including the word sequence, or may be common regardless of the arrangement of the words. In the former case, for example, in document A, word a and word b are arranged in the order, and in document B, word b and word a are arranged in order. Is a word common to both documents, but both words a and b are determined not to be common words in both documents. On the other hand, in the latter case, in the same situation as described above, both words a and b are common words in both documents. In this specific example, the former case is assumed. The common word may be a matching word or a similar word. In this specific example, it is assumed that the common word is a matching word. Also, being able to cover the entire document with a plurality of words means that the document can be created by arranging the plurality of words. If it is determined that the cover is available, the process ends. If it is determined that the cover is not available, the process proceeds to step S206.

  (Step S206) The selection unit 13 increments the counter I by 1. Then, the process returns to step S204.

  If document information cannot be covered even if all source language examples are selected, it is determined that an error has occurred, and the example translation process may be terminated, or the processes in and after step S103 may be continued. .

と算出される。同様にして、用例ＩＤ２〜４についても類似度を算出すると、

となる（ステップＳ２０１）。なお、用例ＩＤが「５」以降の用例についても、同様にして類似度を求めるものとする。 Returning to the specific processing, the selection unit 13 first reads the source language example from the parallel example corpus shown in FIG. 3, and reads the source language example and the document information “Glassgo sleeper bed” received from the reception unit 11. The morphological analysis is performed on “Please give me a ticket”, and the similarity between the document information “Please give me a bed ticket to Glasgow” and each source language example is calculated using Equation (1). Examples ID "1" of the source language example (This is referred to as J _1. Examples ID is assumed to be the same for example after "2") and the document (This is referred to as J ₀₎ indicated by the document information and Since there are six common words in both documents, "Until", "No", "O", "Request", "Shi", "Masu",

Is calculated. Similarly, when similarity is calculated for example IDs 2 to 4,

(Step S201). It should be noted that the similarities are obtained in the same manner for the examples with the example ID “5” and thereafter.

  When the source language examples are sorted so that the selection unit 13 is in descending order of similarity, the example IDs are “2”, “4”, “1”, “3”, and the example ID is “5” or later. Is the fifth or later (step S202). Then, the selection unit 13 selects the source language examples until the document information can be covered with words common to the document information “Please give me a bed ticket to Glasgow” in order from the first source language example after sorting. (Steps S203 to 205). FIG. 5 is a diagram illustrating a source language example of the example IDs 1 to 4. In each source language example, words that are common to document information are underlined. As can be seen from FIG. 5, the document information is covered by the example IDs 1 to 4. Therefore, after selecting the example with the example ID “3”, the selection unit 13 determines that the document information can be covered with the selected source language example (step S205), and ends the selection of the source language example (step S102). ). Then, the selection unit 13 passes the example IDs “1”, “2”, “3”, and “4” of the selected source language examples to the translation candidate information creation unit 14 and the translation document information creation unit 15.

Upon receiving the example IDs “1”, “2”, “3”, “4” of the selected source language example from the selection unit 13, the translation candidate information creation unit 14 translates the target language example corresponding to the example ID. Read from the example corpus. Then, the translation candidate information creation unit 14 sets a common word among the read target language examples as one common node. Specifically, as shown in FIG. 6, words that are common between the target language examples are grouped together (in FIG. 6, the words surrounded by a square are common words). Here, the matching word is a common word. In FIG. 6, the target language example whose example ID is “i” is “E _i ”. As a method for finding common words in this way, a multiple alignment technique often used in DNA analysis can be used. Since multiple alignment is conventionally known, the description thereof is omitted. For details of multiple alignment, refer to Reference 1 below.

(Reference 1) Naruki Miyake, Kaoru Kinhisa. Human genome project and knowledge information processing. Baifukan, 1995.

  Next, the translation candidate information creation unit 14 sets a word that is not common among the target language examples as one single node. Then, the translation candidate information creation unit 14 gives each node a node ID that can uniquely identify each node (including both a common node and a single node). For example, “Can” is a node ID “N001”, “I” is a node ID “N002”, “′ d” is a node ID “N003”, “would” is a node ID “N004”, and so on. Is done. Thereafter, the translation candidate information creation unit 14 creates translation candidate information indicating the adjacent relationship between the common node and the single node (step S103). This translation candidate information is information indicating a word candidate before a certain word and a word candidate behind. FIG. 7 is a diagram showing translation candidate information. In the translation candidate information, the node ID, the word of the node identified by the node ID, the node ID before the node identified by the node ID, and the node ID after the node identified by the node ID Are associated. Here, the node ID “N000” indicates the start node, and the node ID “N999” (not illustrated in FIG. 7) indicates the end node. For example, since the node ID before the word “Can” of the node ID “N001” is “N000” and the subsequent node ID is “N002”, the node before “Can” is the start node and the It can be seen that the node of “I” is “I”.

  FIG. 8 is a diagram illustrating the translation candidate information in FIG. 7 in an easy-to-understand manner. With the translation candidate information shown in FIG. 7, as shown in FIG. 8, the possible word sequences from the start node to the end node can be known from the translation candidate information. Translation document information is created by selecting a sequence of words from the translation candidate information (step S104).

  Next, creation of the translated document information will be described. The translated document information creating unit 15 creates translated document information (step S104 in the flowchart in FIG. 2) as shown in the flowchart in FIG. First, the flowchart of FIG. 9 will be described.

(Step S301) The translated document information creation unit 15 sets the counter J to “1”.
(Step S302) The translation document information creation unit 15 uses the example IDs “1”, “2”, “3”, “4” (J) that the J-th translation candidate indicated by the translation candidate information has received from the selection unit 13. In the equation (2), it is determined whether or not the above equation (2) is satisfied for i = 1 to 4). Here, Δ _lower and Δ _upper are set to “0.2”, respectively. In addition, the similarity is calculated using Equation (1). Further, when calculating the similarity, the translated document information creating unit 15 uses the document information received from the accepting unit 11 and uses the example ID of the selected source language example received from the selecting unit 13. Read the example from the parallel example corpus. It is assumed that morphological analysis is performed on the source language example and the document indicated by the document information when calculating the similarity. Or you may use the result of the morphological analysis which the selection part 13 performed with respect to the document which the example for a source language and document information show. If the expression (2) is satisfied, the process proceeds to step S303. Otherwise, the process proceeds to step S306.

  (Step S303) The translation document information creation unit 15 calculates a product of N-gram probabilities for the Jth translation candidate. Here, it is assumed that a product of 3-gram probabilities is calculated. For example, the translated document information creation unit 15 multiplies the 3 gram probability in units of three words while shifting words one by one from the head of the Jth candidate sentence. Since the value of the product of 3 gram probabilities is generally very small, the translated document information creation unit 15 may calculate “−log (product of 3 gram probabilities)”. This log may be a natural logarithm or a common logarithm. Also in the following description, the product of N-gram probabilities may be obtained by taking “−log” as the product of N-gram probabilities. Further, the product value of N-gram probabilities (including the value obtained by taking “−log”) may be normalized with respect to the number of words. For example, normalization may be performed by dividing the product of N-gram probabilities by the number of words.

  (Step S304) The translation document information creation unit 15 calculates the product of the N-gram probabilities stored so far in the recording medium (not shown) of the translation document information creation unit 15 and the product of the N-gram probabilities calculated in Step S303. Compare. If the N-gram probability product calculated in step S303 is larger than the stored N-gram probability product, the process proceeds to step S305. Otherwise, the process proceeds to step S306.

(Step S305) Information indicating the candidate for the Jth translation and the N-gram probability calculated in Step S303 are overwritten and recorded on the recording medium (not shown). If nothing is stored in the recording medium, it is recorded as it is. The information indicating the Jth translation candidate recorded may be any information as long as the information can reproduce the Jth translation candidate document. For example, the information may be “J-th”, may be a sequence of node IDs for indicating a sequence of translation candidate words, or may be information indicating a translation candidate document itself.
(Step S306) The translated document information creation unit 15 increments the counter J by one.

  (Step S307) The translation document information creation unit 15 refers to the translation candidate information and determines whether there is a Jth translation candidate. If there is a Jth translation candidate, the process returns to step S302. If not, all translation candidates have been considered, and the process proceeds to step S308.

  (Step S308) The translation document information creation unit 15 reads information indicating translation candidates stored in the above-described recording medium (not shown), creates translation document information based on the information, and creates the created translation document information. It passes to the output unit 16. If a translation candidate document itself is stored in the recording medium (not shown), reading the document is also used to create translation document information.

  In the flowchart of FIG. 9, the case where the product of N-gram probabilities is obtained only for translation candidates satisfying Expression (2) has been described. However, the product of N-gram probabilities is calculated for all translation candidates, and Expression (2 ) And a translation candidate having a maximum N-gram probability product value may be selected.

となることからも、そのような考えが妥当であると思われる。 Here, the reason why the translation document information creation unit 15 does not use translation candidates that do not satisfy Expression (2) in the creation of translation document information will be briefly described. In general, if there is a document in the source language and a translation of the document in the target language, the similarity between the document in the source language and the source language example, and the translation of the document in the target language and the example in the target language It is considered that the degree of similarity should be relatively close. Therefore, final translation document information is created from the translation candidates satisfied by the expression (2). The document information "Glasgow I'd like a ticket of bed until" the translation in English is "I'd like a sleeping car ticket to Glasgow, please. " As it is, and the translation of this English and _{E 0} ,

Therefore, such an idea seems to be appropriate.

  The output unit 16 accumulates the translation document information received from the translation document information creation unit 15 in the recording medium 22 (step S105). In this way, a series of example translation processing is completed.

  In this specific example, the case where the selection of the source language example is completed when the document information can be covered by the selected source language example has been described using the flowchart of FIG. Even if the information can be covered, extra source language examples may be selected. Further, in the flowchart of FIG. 4, the case where the source language examples are selected in order from the one with the highest similarity has been described. However, the one with the highest similarity is skipped (for example, every other piece or every second piece). ) Source language examples may be selected. Further, the selection unit 13 randomly reads the source language example from the parallel translation example corpus, and has a value (for example, 0.3, etc.) in which the similarity between the read source language example and the document information is determined in advance. (If not limited to a value), the source language example may be selected, and such selection processing may be repeated until the document information can be covered by the selected source language example.

Further, in this specific example, the case where the common node is a word that matches between the target language examples has been described. However, as described above, a word that is similar between the target language examples may be used as the common node. For example, “would” and “′ d” may be determined to be similar and may be a common node. Further, whether the word w ₁ and the word w ₂ are similar may be determined by the following formula.

Here, simChar is a surface layer match (character match), and simPos is a function for finding part-of-speech match, and takes “1.0” when matched and “0.0” when not matched. SimSem is a semantic similarity and takes a value of “0.0” to “1.0”, and the value is larger as the similarity is higher. If the value of simWord is equal to or greater than a predetermined value (for example, 0.5, but not limited to this value), the translation candidate information creation unit 14 determines that both words are similar. In the above formula (3), a case where λ ₂ , λ ₃ = 0 and a predetermined value is “1.0” is the specific example. Since simChar, simPos, and simSem are conventionally known, the description thereof is omitted.

  Further, in this specific example, as shown in FIG. 7, the description has been given of the case where nodes before and after the word are indicated in the translation candidate information. However, in the translation candidate information, only the node before or after the word is included. May be shown. Even in such a case, it is possible to know the arrangement of words in translation candidates.

  Further, in this specific example, the case where the translation document information is created using the translation candidate satisfying the expression (2) has been described. However, this is an example, and the translation document information creation unit 15 determines whether the translation candidate is a translation candidate. It is not necessary to determine whether or not Expression (2) is satisfied. In that case, for example, in the flowchart of FIG. 9, the process of step S302 may be omitted, and the process may proceed from step S301 to step S303.

  In this example, a candidate for translation was created assuming that a comma, period, question mark (question mark) is a single word, but a comma, period, question mark, exclamation mark (exclamation mark), punctuation marks, etc. May be treated as a single word or not.

In this example, the source language is Japanese and the target language is English. However, as long as the source language and the target language are different from each other, both may be arbitrary languages. Needless to say, it is not limited to English.
In this specific example, the case where Δ _lower and Δ _upper are the same value has been described, but it is needless to say that both values may be different.

  As described above, in the example translation apparatus 1 according to the present embodiment, translation from the source language to the target language can be performed by using the parallel translation example corpus instead of using word alignment or phrase alignment.

  In creating the translation candidate information, the translation candidate information creating unit 14 may further expand the translation candidate information created as described above. For example, the translation candidate information creation unit 14 determines the arrangement of the two words when the 2-gram probability of the arrangement of the two words related by the adjacent relationship other than the adjacent relationship of the node indicated by the translation candidate information is greater than or equal to the threshold value. The adjacent relationship shown may be added to the translation candidate information. Specifically, in the translation candidates shown in FIG. 8, two words that are related by an adjacent relationship between arbitrary nodes not included in FIG. 8, for example, an adjacent relationship between the node “like” and the node “seat” If the 2-gram probability of “like-seat” is equal to or greater than the threshold value “0.1”, an adjacency indicating the arrangement of the two words may be added to the translation candidate information. An appropriate value can be arbitrarily set for the threshold value. Further, a predetermined restriction may be added instead of determining the adjacency relationship between arbitrary nodes. For example, a node that can be traced from a certain node (referred to as node A) by the adjacent relationship included in the translation candidate information from the node A, is a node after the node A, and K nodes from the node A The 2 gram probability is compared with the threshold value as described above only for the adjacent relationship with nodes within (with K being a predetermined integer of 2 or more), and the other adjacent relationships are added to the translation candidate information. It is not necessary to perform. As the value of K, for example, K = 3 may be used. FIG. 10 is a diagram illustrating an example of translation candidate information to which an adjacent relationship indicating the arrangement of two words is added. This is an adjacency in which an adjacency indicated by a broken line is added.

  In addition, the parallel translation example corpus may include a component alignment that is an alignment of components that constitute a sentence. In that case, the translation document information creation unit 15 may create translation document information using the component element alignment. Here, the component which is an element which comprises a sentence is a word, a phrase, etc., for example. Thus, the component alignment may be, for example, word alignment, phrase alignment, or both.

  In the case where the component alignment is included in the bilingual example corpus, the translated document information creation unit 15 determines whether or not there is alignment by the component alignment between the document indicated by the document information received by the receiving unit 11 and the translation candidate document. Based on the above, a translation candidate that satisfies a predetermined condition may be selected. For example, the translation document information creation unit 15 is included in the translation candidates depending on whether or not there is alignment by component alignment between the document indicated by the document information received by the reception unit 11 and the translation candidate document. A score may be assigned to the element, and a translation candidate in which the product of N-gram probabilities and the score satisfy a predetermined condition may be selected. More specifically, when “Glasgow” and “Glasgow” are aligned by the word alignment included in the bilingual example corpus, the translation document information creation unit 15 is included in the translation candidates. When a score “2.0” is assigned to “Glasgow”, which is an element (in this case, a word), and a product of N-gram probabilities is calculated for translation candidates including the node “Glasgow” (for example, FIG. In step S303), the score “2.0” given to “Glasgow” may be multiplied, and the result of the product may be compared with the result of the product of other N-gram probabilities. By doing so, it is possible to increase the probability that translation candidates including the aligned constituent elements are selected by the translation document information creation unit 15 and finally become translation document information. In the case of the above description, the translation result including “Glasgow” has a larger product of N-gram probabilities than the translation candidates including “London”, “San Francisco”, etc., and the English translation of “Glasgow” It is possible to increase the possibility that “Glasgow” is included in the translated document. Moreover, the translation document information creation unit 15 calculates the product of N-gram probabilities for all translation candidates, and translation candidates including the constituent elements aligned by the constituent element alignment are included in the candidates. The translation document information may be created by adding a score to the product of N-gram probabilities for each component and selecting the maximum value of the results. Needless to say, the value of the score to be given to the aligned component can be arbitrarily set. Also, instead of assigning a score to the aligned component, select a translation candidate including the aligned component that maximizes the product of N-gram probabilities. It may be.

  In the present embodiment, the case where the similarity between documents is calculated using DP matching has been described. However, the similarity between documents may be calculated using an edit distance other than DP matching.

  In the above embodiment, each process or each function may be realized by centralized processing by a single device or a single system, or may be distributedly processed by a plurality of devices or a plurality of systems. It may be realized by doing. The example translation apparatus 1 may be a stand-alone apparatus or a server in a server / client system.

  In the above embodiment, each component may be configured by dedicated hardware, or a component that can be realized by software may be realized by executing a program. For example, each component can be realized by a program execution unit such as a CPU reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory. The software that realizes the example translation apparatus in the above embodiment is the following program. In other words, the program includes a reception step of receiving document information, which is information indicating a document in the source language, and a plurality of source language examples that are examples of the source language stored in the parallel translation example corpus storage unit. The document information received in the receiving step from the plurality of source language examples of the parallel translation example corpus, which is information having a plurality of target language examples that are translations of the source language examples in the target language different from the source language, Document information received in the receiving step using a selection step for selecting two or more source language examples similar to a document and two or more target language examples corresponding to the two or more source language examples selected in the selection step A translation candidate information creation step for creating translation candidate information that is information indicating translation candidates of the translation, and the translation candidate information created in the translation candidate creation step A translation document information creating step for creating a translation document information that is information indicating the selected translation candidate and a translation document information created in the translation document information creation step. And an output step for outputting.

  In the above program, the reception step for receiving information and the output step for outputting information do not include at least processing performed only by hardware, for example, processing performed by a modem or an interface card in the output step. .

  Further, this program may be executed by being downloaded from a server or the like, and a program recorded on a predetermined recording medium (for example, an optical disk such as a CD-ROM, a magnetic disk, a semiconductor memory, or the like) is read out. May be executed by

  Further, the computer that executes this program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  FIG. 11 is a schematic diagram showing an example of the external appearance of a computer that executes the program and realizes the example translation apparatus 1 according to the embodiment. The above-described embodiment can be realized by computer hardware and a computer program executed on the computer hardware.

  In FIG. 11, a computer system 100 includes a computer 101 including a CD-ROM (Compact Disk Read Only Memory) drive 105, an FD (Flexible Disk) drive 106, a keyboard 102, a mouse 103, and a monitor 104.

  FIG. 12 is a diagram illustrating a computer system. In FIG. 12, in addition to the CD-ROM drive 105 and the FD drive 106, a computer 101 includes a CPU (Central Processing Unit) 111, a ROM (Read Only Memory) 112 for storing a program such as a bootup program, A CPU (Random Access Memory) 113 that is connected to the CPU 111 and temporarily stores application program instructions and provides a temporary storage space, a hard disk 114 that stores application programs, system programs, and data, a CPU 111 and a ROM 112. Etc. to each other. The computer 101 may include a network card (not shown) that provides connection to the LAN.

  A program for causing the computer system 100 to execute the function of the example translation apparatus 1 according to the above embodiment is stored in the CD-ROM 121 or the FD 122, inserted into the CD-ROM drive 105 or the FD drive 106, and stored in the hard disk 114. May be forwarded. Instead, the program may be transmitted to the computer 101 via a network (not shown) and stored in the hard disk 114. The program is loaded into the RAM 113 at the time of execution. The program may be loaded directly from the CD-ROM 121, the FD 122, or the network.

The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 101 to execute the functions of the example translation apparatus 1 according to the above-described embodiment. The program may include only a part of an instruction that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 100 operates is well known and will not be described in detail.
Further, the present invention is not limited to the above-described embodiment, 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, according to the example translation apparatus and the like according to the present invention, translation can be performed using the parallel example corpus, which is useful as an apparatus for performing machine translation.

The block diagram which shows the structure of the example translation apparatus by Embodiment 1 of this invention. The flowchart which shows operation | movement of the example translation apparatus by the embodiment The figure which shows an example of the example corpus for parallel translation in the embodiment The flowchart which shows an example of the process which selects the example for source languages in the embodiment The figure which shows an example of the example for the selected source language in the embodiment The figure which shows an example of the multiple alignment in the embodiment The figure which shows an example of the translation candidate information in the embodiment The figure which shows an example of the arrangement | sequence of the node of the word shown by the translation candidate information in the embodiment The flowchart which shows an example of the process which produces the translation document information in the embodiment The figure which shows an example of the arrangement | sequence of the node of the word shown by the translation candidate information to which the adjacency relation of the word in the same embodiment was added Schematic diagram showing an example of the appearance of the computer system in the embodiment The figure which shows an example of a structure of the computer system in the embodiment

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Example translation apparatus 11 Reception part 12 Parallel example corpus memory | storage part 13 Selection part 14 Translation candidate information creation part 15 Translation document information creation part 16 Output part 21,22 Recording medium

Claims (9)

A reception unit that receives document information that is information indicating a document in the source language;
A parallel translation in which a bilingual example corpus, which is information having a plurality of source language examples that are examples in the source language, and a plurality of target language examples that are translations of the source language examples in a target language different from the source language is stored Example corpus storage unit,
A selection unit that selects two or more source language examples similar to the document indicated by the document information received by the reception unit from a plurality of source language examples of the parallel example corpus;
A translation candidate that creates translation candidate information, which is information indicating translation candidates of document information received by the receiving unit, using two or more target language examples corresponding to the two or more source language examples selected by the selection unit. An information creation department;
A translation document information creation unit that selects translation candidates satisfying a predetermined condition from the translation candidate information created by the translation candidate information creation unit, and creates translation document information that is information indicating the selected translation candidates;
An example translation apparatus comprising: an output unit that outputs translation document information created by the translation document information creation unit. The selection unit includes:
The degree of similarity between the plurality of source language examples and the document indicated by the document information received by the receiving unit is calculated, and the document is expressed by a word common to the two or more selected source language examples and the document indicated by the document information. 2. The example translation apparatus according to claim 1, wherein a source language example having a high degree of similarity is selected so that the entire document indicated by the information can be covered. The selection unit includes:
The example translation apparatus according to claim 2, wherein a ratio of words common to the two documents is used as the similarity between the two documents. The translation candidate information creation unit
Two or more target language examples corresponding to two or more source language examples selected by the selection unit are acquired from the parallel example corpus, and a common word between the acquired two or more target language examples is defined as one common node. The example of any one of claims 1 to 3, wherein translation candidate information indicating adjacency relation between a common node and a single node is created with another word in the obtained two or more target language examples as one single node. Translation device. The translation candidate information creation unit
When the 2-gram probability of the two-word sequence related by the adjacent relationship other than the adjacent relationship of the node indicated by the translation candidate information is greater than or equal to the threshold value, the adjacent relationship indicating the two-word sequence is added to the translation candidate information The example translation apparatus according to claim 4. The translated document information creation unit
The product of N-gram probabilities is calculated for each translation candidate indicated by the translation document information, and the translation candidate having the maximum N-gram probability product is selected as a translation candidate satisfying a predetermined condition. The example translation apparatus according to claim 5. The bilingual example corpus also includes a component alignment that is an alignment of components that are components of a sentence,
The translated document information creation unit
The translation candidate satisfying a predetermined condition is selected based on the presence or absence of alignment by the component element alignment between the document indicated by the document information received by the reception unit and the translation candidate document. Example translation device. A parallel translation in which a bilingual example corpus, which is information having a plurality of source language examples that are examples in the source language, and a plurality of target language examples that are translations of the source language examples in a target language different from the source language is stored An example translation method that is processed using an example corpus storage unit, a reception unit, a selection unit, a translation candidate information creation unit, a translation document information creation unit, and an output unit,
A receiving step in which the receiving unit receives document information that is information indicating a document in a source language;
A selection step in which the selection unit selects two or more source language examples similar to the document indicated by the document information received in the receiving step, from a plurality of source language examples of the parallel example corpus;
The translation candidate information creation unit is information indicating translation candidates of the document information received in the reception step using two or more target language examples corresponding to the two or more source language examples selected in the selection step. A translation candidate information creation step for creating translation candidate information;
The translation document information creation unit selects a translation candidate satisfying a predetermined condition from the translation candidate information created in the translation candidate creation step, and creates translation document information that is information indicating the selected translation candidate Document information creation step;
An example translation method, comprising: an output step in which the output unit outputs the translation document information created in the translation document information creation step. On the computer,
A reception step for receiving document information that is information indicating a document in the source language;
A plurality of source language examples, which are examples in the source language, and a plurality of target language examples, which are translations of source language examples in a target language different from the source language, are stored in correspondence with the parallel translation example corpus storage unit. A selection step of selecting two or more source language examples similar to the document indicated by the document information received in the receiving step from a plurality of source language examples of the parallel example corpus that is information;
Translation candidates for creating translation candidate information, which is information indicating translation candidates of the document information received in the receiving step, using two or more target language examples corresponding to the two or more source language examples selected in the selecting step Information creation step;
Selecting a translation candidate satisfying a predetermined condition from the translation candidate information created in the translation candidate creation step, and creating a translation document information that is information indicating the selected translation candidate;
An output step of outputting the translated document information created in the translated document information creating step.

Images