JP2005216127A - Machine translation method and machine translation device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a machine translation device and method capable of outputting a suitable text of a translation destination language even in case that a translation source language or the translation destination language does not have a sufficient environment such as corpus, or a text to be inputted is inaccurate. <P>SOLUTION: When a text 31 of the translation source language is inputted to the mechanical translation device 30, a characteristic word extraction part 50 extracts characteristic words 33 from the input text. A parallel translation selection part 60 selects an optimum translation 34 of each characteristic word 33. Further, a plurality of translations 34 are inputted to a text generation part 70, whereby a most natural generation text 35 containing the translated words can be obtained. A text 32 of the translation destination language is outputted from an output part 80. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a machine translation method and apparatus. In particular, this is a technique related to a processing method from inputting a text in a translation source language to outputting a text in a translation destination language.

As a conventional machine translation method, for example, a machine translation method disclosed in Patent Document 1 is known. As disclosed in the patent document, the conventional machine translation method analyzes the Japanese to derive the semantic structure of the sentence when, for example, translation from Japanese to English is performed.
This semantic structure can be expressed by a tree structure or a semantic network. Some of them analyze the generated semantic network, select an appropriate translation, and generate a target translation.

  Here, the semantic network is described by a node to which a semantic symbol is added and an arc to which relation information between the nodes is added, and a sentence generation rule necessary for generating a translated sentence is added to each node. . At the time of generation, an English word and its grammatical attribute are searched based on a sentence generation rule, and then a process for determining a word, a part of speech, and a sentence pattern that match the arc information is performed.

Japanese Patent Publication No. 8-33895

In recent years, a large amount of corpus has become available, and the corpus has also been used for the purpose of generating natural surface sentences. One typical example is a language model used for machine translation, which is disclosed in Non-Patent Document 1, for example.
In the statistical machine translation shown here, it is assumed that the text P written in the source language is S, the text written in the target language is T, and the probability P of generating T when S is given is maximized. Output text as an optimal translation.
The models used at this time include a translation model for replacing words and phrases from the original language to the target language, and a language model for rearranging the replaced words and phrases in a reasonable order on the target language side.

Brown, PF, Cocke, J., Piera, SAD, Pietra, VJD, Jelinek, F., Lalfferty, JD, Mercer, RL, and Roossin, PS (1990). "A Statistical Approach to MachineTranslation." Computational lingustics, 16 (2), 79-85

In such a conventional statistical machine translation technique, there is an assumption that a natural sentence can be generated by rearranging a given set of words. In other words, it is assumed that a set of words for generating a natural sentence is generated without excess or deficiency by a translation model.
However, a large-scale bilingual corpus is necessary for this premise, and even if the source language is a relatively well-developed language such as Japanese, the status of the bilingual corpus with the target language, Depending on the state of the corpus, there is a problem that the above-described conventional method cannot be expected to sufficiently improve the translation accuracy.

  Also, especially when the source language expression is inappropriate, such as spoken language, grammatical errors, and descriptive errors, these conventional methods do not allow correct semantic structure analysis or translation model action, and are completely incorrect. The translation result may be output.

  The present invention was created in view of the above-described problems of the prior art. The purpose of the present invention is when there is not enough environment such as a corpus for a source language or a target language, or when a text to be input is not prepared. It is an object of the present invention to provide a machine translation method and apparatus capable of outputting a text in a suitable target language even when inaccurate.

In order to solve the above problems, the present invention creates the following machine translation method.
That is, the present invention is a method for machine translation of a translation source text described in a predetermined translation source language into a translation destination text described in a translation destination language, and the translation source text is input in an input step.
A feature word extraction step for extracting a feature word having a characteristic meaning from the input translation source text, and a feature selection step for translating the feature word into a translation word expressed in the translation destination language. The post-feature translation that becomes the target word translation is obtained.

  Then, a character unit candidate generation step for generating character unit candidates from the translated words based on a predetermined generation rule, and a translation destination text for generating one or a plurality of translation destination text candidates on the assumption of the dependency between the character units A candidate generation step and an evaluation step of evaluating each translation destination text candidate when there are at least a plurality of translation destination text candidates. In the output step, at least one of the translation destination text candidates is output in relation to the evaluation result.

  The machine translation method according to claim 2 further includes a dependency extraction step for extracting dependency relationship information between the feature words out of the feature words extracted in the feature word extraction step, and has the dependency relationship information. For the character unit candidate generated from the feature word, a translation destination text candidate is generated using the dependency relationship information in the translation destination text candidate generation step.

  Alternatively, the machine translation method according to claim 3 includes a dependency extraction step for extracting dependency relationship information between the feature words out of the feature words extracted in the feature word extraction step, and in the evaluation step, The dependency relationship information is compared with the dependency relationship in the translation destination text candidate, and the evaluation of the translation destination text candidate having high dependency relationship is made high.

  Here, the generation rule used in the character unit candidate generation step is to extract a sentence / phrase including the translated word from a database including a plurality of sentences described in the translation target language, and to determine the relationship between the translated word and the sentence / phrase. It is also possible to adopt a configuration that is an automatically acquired rule.

The present invention may provide the following machine translation apparatus.
That is, the machine translation device performs machine translation of a translation source text described in a predetermined translation source language into a translation destination text described in a translation destination language.
The machine translation device according to claim 5 includes (1) an input unit for inputting a translation source text, and (2) a feature word extraction unit for extracting a feature word having a characteristic meaning from the translation source text. (3) a translation selection unit that translates the feature word into a translation expressed in the target language; (4) a character unit candidate generation unit that generates a character unit candidate from the translation based on a predetermined generation rule; ) Translation destination text candidate generation means for generating one or a plurality of translation destination text candidates on the assumption of the dependence between the character units; (6) Each translation destination text candidate when there are at least a plurality of translation destination text candidates (7) In relation to the evaluation value, there is provided an output means for outputting at least one of the translation destination text candidates.

  The machine translation device according to claim 6 is provided with dependency relationship extraction means for extracting dependency relationship information between feature words out of the feature words extracted by the feature word extraction means, and the feature words having the dependency relationship information For the character unit candidate generated from the above, the translation destination text candidate generation means can generate a translation destination text candidate using the dependency relationship information.

  Further, in the machine translation device according to claim 7, the machine translation device includes a dependency relationship extraction unit that extracts dependency relationship information between the feature words out of the feature words extracted by the feature word extraction unit, and the evaluation unit includes: The dependency relationship information is compared with the dependency relationship in the translation destination text candidate, and the evaluation value of the translation destination text candidate having high dependency relationship is increased.

  The generation rule used by the character unit candidate generating means extracts a sentence / phrase including the translated word from a database including a plurality of sentences described in the translation destination language, and automatically determines the relationship between the translated word and the sentence / phrase. The configuration may be an acquired rule.

The following effects are produced by the above invention.
That is, according to the machine translation method according to claims 1 to 4, by creating a new machine translation method, even when there is not a sufficient corpus (particularly a parallel corpus) in the source language and the target language, A highly accurate machine translation method can be easily configured.
In addition, because feature words are extracted from the text to be input, it is possible to achieve natural and easy-to-understand translation even if grammatically incorrect text is input. The translated text obtained at the same time is easy to grasp the meaning because the feature word is surely suppressed. Conventional translation techniques often output translation results that are difficult to grasp in addition to grammatical errors, but this method can provide a suitable text.

  In particular, in the machine translation method according to claims 2 and 3, since the dependency relationship in the translation source language text can be used in the generation step of the translation destination language, it can contribute to highly accurate translation.

In addition, according to the machine translation apparatus according to claims 5 to 8 of the present invention, a sufficient corpus (particularly a bilingual corpus) is not prepared for the source language and the destination language by introducing a new machine translation method. Even in this case, a highly accurate machine translation apparatus can be provided.
In addition, because feature words are extracted from the text to be input, it is possible to achieve natural and easy-to-understand translation even if grammatically incorrect text is input.

  In particular, in the machine translation device according to claims 6 and 7, the dependency relationship in the translation source language text can be used in the generation step of the translation destination language, which can contribute to highly accurate translation.

Such a machine translation apparatus is particularly suitable for translating sentences in a translation source language that is not grammatically accurate, such as an input of colloquial language or an article title. The result obtained is a translated text that is easy to understand.
Moreover, in addition to the languages with a well-developed corpus environment such as between Japanese and English as shown in this example, the language is translated into languages that are not well-developed, such as languages with strong locality and native languages of developing countries. In some cases, practical machine translation results can be generated.

DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, the best mode for carrying out the invention will be described based on examples shown in the drawings. The embodiment is not limited to the following.
FIG. 1 shows a flowchart of a machine translation method according to the first embodiment of the present invention. As shown in the figure, Japanese text (1) as a translation source language is input, and characteristic words in the text (1) are extracted (2). The extracted characteristic word is recorded as characteristic word data (6). Note that since the present invention targets any language, the present invention is not limited to those that can be classified as words, and feature words that include a wide range of words, phrases, and the like may be extracted, and subsequent processing may be performed.

Next, the optimal translation of the characteristic word is selected (3) using dictionary data (7) of the translation source language and the translation destination language. Thereby, a characteristic word in the translation destination language is generated and recorded as characteristic word translation data (8). Further, a text in the translation destination language is generated (4) using the characteristic word translation data (8). The process of performing machine translation by generating natural English text (5) from the characteristic word translation as described above is the greatest feature of the present invention that has not been created so far.

  In the text generation (4), character unit candidates are generated from characteristic word translations based on a predetermined generation rule, and one or a plurality of translation destination text candidates are generated assuming a dependency relationship between the character units. Furthermore, the English text (5) judged to be the most appropriate by evaluating the translation destination text candidate is output.

  Further, the present invention can provide the machine translation method of the second embodiment as shown in FIG. A Japanese text (10) as a source language is input, and characteristic words in the text (10) are extracted (11). The extracted characteristic words are recorded as characteristic word data (16).

At the time of extracting characteristic words, dependency relationships between the words are extracted (12). The inter-word dependency data (17) is recorded.
Then, the optimal translation of the characteristic word is selected (13) using the dictionary data (18) of the translation source language and the translation destination language. As a result, a characteristic word in the translation destination language is generated and the characteristic word translation and the dependency data (19) are recorded. Further, using the data (19), a text in the translation destination language is generated (14). To do. Thus, a natural English text (15) is generated from the characteristic word translation.

In the text generation (14), character unit candidates are generated from characteristic word translations based on a predetermined generation rule, and one or a plurality of translation destination text candidates are generated assuming a dependency relationship between the character units. When dependency relationships of characteristic words are recorded as described above, by generating translation destination text candidates based on the dependency relationships, it is possible to generate a highly accurate translation destination text with no error in dependency relationships.
Furthermore, the English text (5) judged to be the most appropriate by evaluating the translation destination text candidate is output.

As a third embodiment of the present invention, a part of the processing in the text generation (14) of the second embodiment is changed, and a predetermined generation is performed from a characteristic word translation regardless of the recorded inter-word dependency (17). Character unit candidates are generated based on the rules, and one or a plurality of translation destination text candidates are generated assuming a dependency relationship between the character units.
Then, when evaluating the translation destination text candidate, the identity with the inter-word dependency is checked, and an English text having a high value is output.

  FIG. 3 shows the configuration of a machine translation apparatus that implements the machine translation method as described above. For example, when the Japanese text (31) is input to the device (30) such as a newspaper article “A fierce urban battle is being conducted with the Dudaev administration unit in front of the official residence, etc.”, the device (30) is input to the device (30). ), And in the characteristic word extraction unit (50), “official residence”, “Dudaev”, “unit”, “violent”, “war” and “development” are extracted from the Japanese text (31).

Further, an optimal translation of each characteristic word, here “palace”, “Dudayev”, “troops”, “fierce”, “battle”, “engage” is selected by the translation selection unit (60).
The selected characteristic word translation is generated (35) while complementing appropriate words in the text generator (70), and the English text (32) is output from the output unit (80).
Next, details of each part (40) to (80) will be described.

As shown in FIG. 4, the input unit (40) includes a CPU (41) and a scanner (42) connected thereto, an external storage device (43) such as a CD drive, a hard disk drive, an MO drive, and a floppy (registered trademark) disk drive. ) Etc. In addition, a known memory can be used as necessary with the operation of the CPU (41).
When the scanner (42) is provided, the CPU (41) performs character recognition processing, converts it into text data, and records it in the external storage device (43). Even when the data of the Japanese text (31) is read directly from the external storage device (43), the data can be converted into a format that can be processed by the device (30) in the CPU (41).

Further, the present invention can receive text data from another computer server or the like via the Internet or an intranet network (44).
The input unit (40) sends the Japanese text (31) to the characteristic word extraction unit (50) shown in FIG.

The configuration of the characteristic word extraction unit (50) is shown in FIG. Again, the CPU and the memory cooperate to perform each process. The characteristic word extraction unit (50) extracts characteristic words that characteristically represent the contents of the text from the input Japanese text (31).
As such a technique, a number of techniques known as elemental techniques such as document summarization in language processing and document search are known and can be used as appropriate. The method described in 2 is used.

IPSJ Natural Language Processing Study Group 1999-NL-133, 1999 “Measure the Representativeness of Terms” Toru Hisamitsu, Yoshiki Niwa, Junichi Sakurai

According to this method, it is possible to measure topicality or field representativeness (representativeness, referred to as feature in this specification) of a word in a document in order to select a feature word, and to perform numerical evaluation. Therefore, it is suitable for the implementation of the present invention. The following is a brief description.
First, the characteristic word extraction unit (50) performs morphological analysis on the Japanese text (31) in the morpheme analysis unit (51) using a known morphological analysis technique. The analyzed morpheme is recorded as a morpheme table in a memory or an external storage device (not shown).

Then, the morphemes are sequentially read out from the morpheme table, and the characteristic is measured for each morpheme (hereinafter referred to as “target term”).
First, in the document extraction unit (52), all the documents including W for the target term W are extracted from an arbitrary document database (56). The document database (56) includes a plurality of Japanese (translation source language) documents and is stored in an external storage device or the like. A Japanese monolingual corpus or a Japanese part of a Japanese-English bilingual corpus may be used.

Next, the word distribution in the set of all the documents from which the term of interest W is extracted and the word distribution of all the documents included in the document database (56) are calculated by the word distribution calculation unit (53), Measure the degree of difference.
Specifically, the following calculation processing is performed in the degree calculation unit (54).

That is, interest term W, documents including W all set D (W), all documents in the set D _0, words in D (W) distribution P _{D (W),} as a word distribution P ₀ at D _0, W features The characteristic Rep (W) is defined based on the distance Dist {P _{D (W)} , P ₀ } of the two distributions {P _{D (W)} , P ₀ }.
In this embodiment, a log likelihood ratio is used as a distance measurement method between word distributions. In other words, all words _{{W 1, ..., W n} }, word w _i is D (W), respectively the frequency of occurrence in the D ₀ k _i, when the _{K i, P D (W)} , the P ₀ The distance Dist {P _{D (W)} , P ₀ } is defined as follows.

ここで、＃Ｄ（Ｗ）は着目タームＷについてＤ（Ｗ）の含む単語数、＃Ｄ₀は同様に全文書の含む単語数である。

Here, # D (W) is the number of words that includes the D (W) for focusing term W, # D ₀ is the number of words that includes the entire document as well.

  According to the definition of Equation 1, it is difficult to effectively compare the features of the terms of interest where #D (W) is separated. Therefore, the normalized feature Rep (W) is defined as in Equation 2. . Note that B (•) is an approximate function using an exponential function whose characteristics can be accurately obtained within a range where #D (W) is an appropriate number (for example, 1000 ≦ # D (W) ≦ 20000). .

(Equation 2)
Rep (W) = Dist {P _{D (W)} , P ₀ } / B (#D (W))

  Here, when #D (W) is remarkably large, such as “Yes”, the number of extractions of D (W) is limited so as to satisfy #D (W) ≦ 20000. The function can be used effectively and the amount of calculation can be reduced.

  The characteristic word extraction unit (50) calculates the characteristic by the above method and extracts characteristic words of the Japanese text (31) input by the characteristic word determination unit (55) according to a predetermined threshold. This result is stored in the external storage device as a characteristic word table (33).

  Next, the translation selection unit (60) shown in FIG. 6 performs a process of selecting an optimal translation and a characteristic word translation (34) for each characteristic word (33). Word translation is usually possible by referring to a dictionary database, but word ambiguity often arises between different languages. That is, it is unclear which of a plurality of translated words should be selected only by inputting a certain word, and the selection of the optimal translated word is a very important technique that affects the translation accuracy of the machine translation device (30). .

  For solving such ambiguity, many well-known excellent techniques have been proposed, and any technique can be used for carrying out the present invention. An example using the following translation selection model disclosed in Patent Document 4 will be shown.

IEICE Technical Committee on Language Understanding and Communication NLC2001-41 “Translation Selection Based on Learning Using Translation Memory and Corpus” Kiyotaka Uchimoto, Satoshi Sekine, Maki Murata, Hitoshi Isahara 2001 "Word Translation based on Machine Learning Models Using Translation Memory and Corpora" Kiyotaka Uchimoto, Satoshi Sekine, Masaki Murata, and Hitoshi Isahara. SENSEVAL-2, pp.155-158,2001

The translation selection unit (60) receives the characteristic word table (33) as input, and first attempts to determine a translation by a method based on the similarity of character strings in the similarity determination unit (61) of character strings.
That is, when the characteristic words included in the table (33) can be uniquely selected in the bilingual dictionary database (63), the translated words are output to the characteristic word translation table (34).

Also, refer to the Japanese-English bilingual corpus database (in this embodiment, one database without any clear distinction from the bilingual dictionary) (63), and similar to the characteristic word table (33) Find the degree.
For example, when “mother” and “refrain” are included as characteristic words, the similarity to “refrain from mother”, “refrain from mother”, “refrain from donation”, etc. included in the parallel corpus database (63) is obtained.

  A method for obtaining the similarity may be a method using a UNIX (registered trademark) diff command disclosed in Non-Patent Document 5, and is obtained by the following equation (3).

Information Processing Society of Japan Natural Language Processing Study 2001-NL-144, 2001 “diff and Language Processing” Masaki Murata, Hitoshi Isahara

(Equation 3)
Similarity = (number of characters matched when diff between characteristic word table and parallel corpus is taken) / (number of characters in parallel corpus)

Here, the bilingual corpus is also used to delete all the function words, verbs, adjectives, and sub-variants before the comparison, or the bilingual corpus extracts the aforementioned characteristic words, and the similarity to those results. It is desirable to compare
As a result, it is possible to avoid the comparison with a parallel corpus containing many inappropriate characters as a comparison target, and to calculate an effective similarity.

A bilingual corpus whose similarity exceeds a predetermined threshold (which can be set as appropriate depending on the contents of the bilingual corpus and the characteristic word table) is selected as the bilingual corpus including the translation of the characteristic word.
Then, the translated words in the bilingual corpus are recorded in the characteristic word translated word table (34). Since the bilingual corpus (63) is used, the bilingual relationship of each word is known in advance, and the relationship between the characteristic word and the translated word can be determined mechanically.

By the way, it is difficult to provide sentence examples similar to all the characteristic word tables input to the bilingual dictionary / parallel corpus database (63), and there may be many cases where the similarity does not exceed the threshold value.
Therefore, a similarity determination unit (62) using a machine learning model is further provided, and an optimal translation is selected using the learned data (64).

  As a machine learning model, SVM (Support Vector Machine) is used, but ME (Maximum Entropy), DL (Decision List), and SB (Simple Bayes) may be used. The certainty of each classification class is basically the probability distribution p of the event (a, b) of class a (εA) in context b (∈B) where B is the set of contexts and A is the set of classes. It is calculated as (a, b). In SVM, such a probability distribution cannot be obtained, but for the sake of convenience, the probability value is set to 1 for the optimum class and 0 for the other classes.

As the feature of the context b, for example, the following can be used.
(1) Morphological information (2) Character n-gram
(3) Information on the example that is the largest match (4) Frequency of appearance of the candidate word with the content word Machine learning is performed using the bilingual corpus by SVM and recorded as learning data (64).

Translated word candidates are extracted from the bilingual dictionary database (63) for characteristic words. The feature of the translation candidate is represented by a combination of English main word, sentence set, word, and appearance frequency. Using the learning data (64) by machine learning, a characteristic word translation is selected from the translation candidates and recorded in the characteristic word translation table (34).
A known technique can be arbitrarily used for such a translation selection model, and it is not necessary to use a machine learning model as described above.

Here, it is assumed that when a translation of a characteristic word is selected, a specific expression or the like is extracted and there is no translation in the parallel translation dictionary database. At this time, a processing unit for performing an inquiry process may be provided in the translated word selection model to inquire the user and give it to the user.
As a method for automating this, as disclosed in Non-Patent Document 6, it is possible to simply replace (transliterate) characters in the target language in accordance with certain rules according to the pronunciation of the source language, or proposed. It is also possible to use a model. What is proposed here is the name of the person or organization, first transliteration or translation with appropriate translations to create translation candidates, and frequently appearing in the specified text database from the candidates A thing is selected as a translated word. As the text database, for example, Web information on the Internet can be used.

Next, the characteristic word translation table (34) formed as described above is input to the text generation unit (70) to generate English text.
There are the following methods for inputting several words and generating text including the words. That is, the text generation method disclosed in Patent Document 2 by the applicant of the present application is applied to English as the translation destination language.

JP 2003-271582 A

  As a specific configuration example of the text generation unit (70), each unit shown in FIG. 7 is provided. The text generation unit (70) can be configured by, for example, a personal computer including a CPU and an external storage medium such as a memory and a hard disk, and the main processing is performed by the CPU, and the processing results are stored in the memory and the external storage medium as needed. Record.

In this embodiment, the input characteristic word translation (34) is used for two processes. One of them is a word string generation rule acquisition unit (71), and the other is a word string candidate generation unit (72).
At this time, the characteristic word translation (34) is defined as a content word that is the main word of the word string. A content word is a morpheme entry word whose part of speech is a verb, adjective, noun, directive, adverb, conjunction, conjunction, excitement, or undefined word, and any other morpheme entry word. Is a function word.

  When the characteristic word translation (34) is given, the word string generation rule acquisition unit (71) searches the English corpus (75) for a sentence including each and performs morphological analysis and syntax analysis (dependency analysis). . Then, a word string including the characteristic word translation (34) is extracted therefrom, and a word string generation rule (76) for generating a word string from the characteristic word translation is acquired and recorded. For example, “palace” → “at the palace”, “palace” → “in the palace”, “battle” → “battles”, “engage” → “They are engaged in”, “engage” → “I engaged”, etc. The word string generation rule (76) is acquired and recorded.

Here, the following method is used for automatic acquisition of generation rules. When the set of characteristic word translations is V and the set of rules for generating word strings from the characteristic word translation k (∈V) is Rk, the rule rk (∈Rk) is expressed in the form Define.
k → hk m *
hk is a main morpheme including a characteristic word translation, and m * is an arbitrary number of morphemes continuing to hk in the same word string. When a characteristic word translation is given, a rule satisfying this format is automatically obtained from the corpus (75) of the translation destination language.

On the other hand, the word string candidate generation unit (72) refers to the word string generation rule (76) and selects word string candidates constituting the English text (32) to be output from the input characteristic word translation (34). Generate.
For example, “palace” is unlikely to be a word string that makes up natural text, but it adds words that are very closely related to the word “palace”, such as “at the palace” or “in the palace”. In preparation for text generation by subsequent processing.

  As in this embodiment, the word string rule for the characteristic word translation (34) input from the corpus (75) is generated by the word string generation rule acquisition unit (71), thereby effectively reducing the amount of calculation. A word string generation rule can be obtained, which contributes to an improvement in processing speed.

  However, it is not always necessary to extract a phrase related to the characteristic word translation (34) from the corpus, and an arbitrary word or phrase is added before or after the inputted characteristic word translation (34) according to the calculation ability. May be. Alternatively, it is also possible to generate a word string from the information of common expressions included in the bilingual dictionary database (63) provided in the device (30). The above “palace” → “at the palace” and the like are expressions described in the bilingual dictionary database, and can be generated as word string candidates.

  When a language that omits many principals such as Japanese is input, word string candidates can be generated by supplementing the subject such as “engage” → “They are engaged in”. At this time, focusing on the fact that in many languages, such as Japanese, the main character is obvious or omitted when it is a formal subject, not only information about what the main character is in the input, but also information that there is no main character It is also possible to generate “engage” → “They are engaged in” without generating “engage” → “He is engaged in”.

Next, a text candidate is generated in the text candidate generation unit (73). Text candidates are expressed as graphs or trees. Here, the relationship of three words “palace”, “troops” and “engage” in the characteristic word translation (34) will be described as an example.
That is, as shown in FIG. 12, assuming a dependency relationship between word string candidates (34a to 34f), word strings such as text candidate 1 (35 ′) and text candidate 2 (35 ′ ′). Text candidates are generated in the form of a dependency structure tree in units of. At this time, in the case of 3 words, all dependency relationships are R! Although x2 = 12, the number of candidates can be reduced by fixing the word order according to the grammar and characteristics of the translation target language.

The generated text candidates (35 ′ and the like) are ordered using the characteristic word translation generation model (77) and language model (78) learned from the corpus by the evaluation unit (74).
Hereinafter, the morpheme model and the dependency model will be described as the characteristic word translation generation model (77) and the language model (78).

  In the characteristic word translation generation model, a model (KM1 to KM5) using the following five types of information as features is considered. In the following, it is assumed that the set V of characteristic word translations is a set of main words that have appeared in the corpus more than a certain number of times, and the word string is expressed as described above. Also, each characteristic word translation is independent, and it is assumed that the characteristic word translation ki corresponds to the word wj (1 ≦ j ≦ m) when the given text consists of the word string w1 ... wm. To do. FIG. 13 shows an explanatory diagram of the model.

[KM1]
Consider two words in front (trigram)
Suppose that ki depends only on the two preceding words wj-1 and wj-2.

[KM2]
Consider backward two words (backward trigram)
Suppose ki depends only on the back two words wj + 1 and wj + 2.

[KM3]
Consider dependency word strings (relation word strings)
If there is a word string related to a word string including ki, it is assumed that ki depends only on the two words wl and wl-1 from the end of the word string at the end of the sentence (see FIG. 13).

[KM4]
Consider received word string (received word string)
If there is a word string that receives a word string containing ki, it is assumed that ki depends only on two words ws and ws + 1 from the main word in the word string (see FIG. 13).

[KM5]
Dependency word strings are considered up to two word strings (related two word strings)
If there is a word sequence related to a word sequence that includes ki, then ki is the two words wl and wl-1 from the end of the word sequence at the end of the sentence, and the two words wh and wh- Assume that it depends only on 1 (see FIG. 13).

と表す。ここで、mi は1 からl までのいずれかの文法的属性を表わす。 Next, a morpheme model (MM) is shown. Suppose that there are l grammatical attributes that should be attached to a morpheme. When a text, that is, a character string is given, a model is used that obtains a probability value of the likelihood that the character string is a morpheme and has the j (1 ≦ j ≦ l) th grammatical attribute.
Given the text T 1, the probability that an ordered morpheme set M 1 is obtained assumes that each morpheme mi (1 ≦ i ≦ n) is independent,

It expresses. Here, mi represents any grammatical attribute from 1 to l.

と表わす。 On the other hand, in the dependency model (DM), when the text T and the ordered morpheme set M are given, the probability of obtaining the dependency ordered set D for each word string is the dependency d1. Is independent, and

It expresses.

  For example, two candidates of "They are engaged in fierce battles with troops at the palace" and "I engaged the palace's troops in fierce battles." Were generated from five characteristic word translations of "palace troops fierce battle engage." To do. Of the candidates, a candidate having a reasonable dependency structure is prioritized by the dependency model.

Using each model as described above, in the present invention, the evaluation unit (74) evaluates the text candidates (35 ′, etc.).
Then, the text candidate whose evaluation value is the maximum or exceeds the threshold or the top N evaluation values is converted into a surface sentence and output.

  As an output method in the output unit (80), in addition to display on a monitor, utterance using speech synthesis, data output to other language processing systems such as a translation system, and the like are possible. In addition, text data may be sent to another computer connected to the network.

  In the embodiment described above, a configuration in which a phrase is added before and after the characteristic word translation is mainly used. However, in the implementation of the present invention, a configuration in which the characteristic word translation (corresponding to the main word) itself is complemented. You can also. As described above, when “engage” is supplemented, a phrase is added to complement “They are engaged”, but “They” can be newly supplemented.

Specifically, the elements of FIG. 14 are added to the configuration of FIG. That is, the characteristic word translation (34) is also input to the dependency related word extraction unit (79), and the word (79) has a dependency relationship with the characteristic word translation (34) from the corpus (75). To extract.
Then, the word is added as a new characteristic word translation, and the word string candidate generation unit (72) performs processing together with the originally input characteristic word translation (34).

  For example, even if “They are engaged at the palace” itself is not in the corpus (75), if the dependency relationships “they are engaged” and “they are at the palace” are in the corpus (75), they are common to them. By adding the word “they” to be newly added as a characteristic word translation, the word string candidate generation unit (72) can generate “They are engaged at the palace”.

In this configuration, it is possible to add characteristic word translations with little metric, but in the present invention, it is not always necessary to extract words having dependency relations from the corpus, and arbitrary characteristic word translation candidates can be selected. It is possible to add and output the one that results in the highest evaluation in the evaluation unit (74).
As a result, even if an important word that determines the meaning of the text as a characteristic word translation is missing due to the expression characteristics of the translation source language, a significant translation destination language text can be output.

  In the machine translation method of the present invention, as a method for realizing a more accurate translation, it is conceivable to use the dependency relationship of the words of the source language text (Japanese text (31)) for generation. Hereinafter, this will be described as another embodiment.

  Specifically, the characteristic word extraction unit (50) in the machine translation apparatus of FIG. 3 is replaced with a characteristic word extraction unit (50 ′) that analyzes the dependency shown in FIG. 8 and can be recorded as a table. Similarly, the translation selection unit (60) can convert the dependency to the characteristic word translation dependency, and the text generation unit (70) uses the word translation dependency to convert the text into the translation selection unit (60 ') shown in FIG. The text generators (70 ′) and (70 ″) shown in FIG. 11 or FIG. Configurations other than the new functions described here are the same as those in the above embodiment, and are omitted here.

  In FIG. 8, a Japanese text (31) is input that “A fierce city battle is being developed with the Dudaev administration unit in front of the official residence”. In the dependency relationship analysis unit (57), using a known dependency relationship (dependency relationship) analysis method, as shown in FIG. 15, “In front of the official residence” (80) “With the Dudaev administration unit” (81) “Intense” It can be seen that “(82)“ A city battle ”(83)“ Developing ”(84) has a dependency (85). Here, it is generally necessary to perform morpheme analysis in advance for the parsing process in the dependency analysis unit (57), and the analysis result in the morpheme analysis unit (51) is used.

When only the dependency relationship between the characteristic words determined by the feature word determination unit (55) by the above-described processing is taken, the inter-word dependency relationship extraction unit (58) performs “government” (86) “ The dependency (91) of “unit” (87), “violent” (89), “war” (90) is extracted.
The dependency relationship between the extracted characteristic words is recorded in an external storage device or a memory as an interword dependency table (36).

Next, in the translation selection unit (60 ′) of FIG. 9, from the relationship between the characteristic word (33) and the characteristic word translation (34), the Japanese inter-word dependency relationship table (36) is converted into the English word translation between Convert to dependency table (37).
Specifically, a dependency conversion unit (65) is provided, and words that can be uniquely determined in the bilingual dictionary database (63), such as words having a clear bilingual relationship such as troops and troops, are directly used as words in the dependency table. replace. On the other hand, when a translation is selected using a machine learning model, for example, a case where ambiguity such as expansion and engagement occurs, the dependency between word translations using the determination result of the similarity determination unit (62) based on the machine learning model Record in table (37).

As shown in FIG. 17, the generated word translation dependency table (37) is “palace” (92) “troops” (93) “fierce” (94) “battle” (95) “engage” (96 ) And the dependency (97) is recorded.
Here, if the proper noun is also a characteristic word, Dudaev is extracted in FIG. 15, and “Dudayev” (98) can also be added as shown in FIG.

In the text generation unit (70 ′) of FIG. 10, the extracted word dependency relationship table (37) is input to the text candidate generation unit (73), and the dependency relationship is not assumed without performing the above-described dependency relationship assumption. decide.
Of course, it is also possible to make a dependency relationship assumption in the text candidate generation unit (73) between translations for which there is no dependency information in the inter-word translation dependency table (37).

  According to this configuration, since the text candidates can be generated with high accuracy, the evaluation unit (74) can more strictly evaluate the generation results of the word string candidate generation unit (72) and the text candidate generation unit (73).

On the other hand, the text generation unit (70 ″) in FIG. 11 can input the extracted word translation dependency table (37) into the evaluation unit (74) and use it for evaluation of text candidates.
Here, the evaluation unit (74) uses the probability related to the dependency order for each word string according to the language model (78) described above, but is recorded in the inter-word translation dependency table (37) between the word translations. For the dependency relationship, the probability value is set to the maximum and used for the evaluation of the text candidate.

  According to this configuration, even if an analysis error occurs in a part of the dependency relationship between word translations, the text candidate generation unit (73) can generate a candidate that is not limited to this, so that it is extremely unnatural. The relationship can be a relatively low evaluation in the evaluation unit (74), which contributes to the generation of natural English text (32) that is a feature of the present invention.

  The present invention is realized by the above-described configuration. However, a characteristic word extraction process, a translation selection process, and a process for generating text from several keywords are all known or will be provided in the future. Can be used. Further, as the accuracy of each process increases, further improvement in translation accuracy is expected.

  In the above, for convenience of explanation, the respective units (40), (50), (60), (70), and (80) have been described separately. However, they can be integrally provided by, for example, one personal computer. . In particular, it is desirable to share a CPU, a memory, an input / output device, a network adapter (not shown) for connecting to a network, an external storage device, and the like, which can contribute to simplification of the device.

The document database (56), the bilingual dictionary / translation corpus database (63), and the corpus (75) recorded in the external storage device can use all or part of the same database.
These are not limited to being recorded on the external storage device, but may be configured to collect those recorded on a plurality of servers on the network.

It is a flowchart of the machine translation method by this invention. 5 is a flowchart of a machine translation method (another embodiment) according to the present invention. 1 is an overall configuration diagram of a machine translation apparatus according to the present invention. It is a block diagram of the input part in this invention. It is a block diagram of the characteristic word extraction part in this invention. It is a block diagram of the translation word selection part in this invention. It is a block diagram of the text production | generation part in this invention. It is a block diagram of the characteristic word extraction part (another Example) in this invention. It is a block diagram of the translation word selection part (another Example) in this invention. It is a block diagram of the text production | generation part (another Example) in this invention. It is a block diagram of the text production | generation part (another Example) in this invention. It is explanatory drawing which shows the example of the text production | generation from a characteristic word translation. It is explanatory drawing which shows the relationship between a characteristic word translation and a word string. It is a block diagram of the dependency related word extraction part which concerns on this invention. It is a structure tree which shows the analysis result of the dependency relationship in a dependency relationship analysis part. It is the content of the inter-word dependency relationship table in the inter-word dependency extraction unit. It is the content of the inter-word dependency table converted by the dependency conversion unit. This is the content of the inter-word dependency table when the proper noun is a characteristic word.

Explanation of symbols

DESCRIPTION OF SYMBOLS 30 Machine translation apparatus 31 Japanese text 32 English text 33 Characteristic word 34 Characteristic word translation 35 Generation text 40 Input part 50 Characteristic word extraction part 60 Translation word selection part 70 Text generation part 80 Output part

Claims (8)

A method of machine-translating a source text described in a predetermined source language into a destination text described in a target language,
Input step to enter the source text,
A feature word extraction step of extracting a feature word having a characteristic meaning from the translation source text;
A translation selection step for translating the feature word into a translation expressed in the target language;
A character unit candidate generation step for generating a character unit candidate from the translated word based on a predetermined generation rule;
A translation destination text candidate generation step for generating one or a plurality of translation destination text candidates assuming the dependency between the character units;
An evaluation step of evaluating each translation destination text candidate when there are at least a plurality of translation destination text candidates;
A machine translation method comprising the steps of outputting at least one of translation destination text candidates in relation to the evaluation result. In the machine translation method,
Of the feature words extracted in the feature word extraction step, a dependency relationship extraction step for extracting dependency relationship information between the feature words is provided,
The candidate text candidate generated from the feature word having the dependency relationship information is generated by using the dependency relationship information in the translation destination text candidate generation step. The machine translation method described. In the machine translation method,
Of the feature words extracted in the feature word extraction step, a dependency relationship extraction step for extracting dependency relationship information between the feature words is provided,
In the evaluation step,
2. The machine translation according to claim 1, wherein the dependency relation information is compared with the dependency relation in the translation destination text candidate, and the evaluation of the translation destination text candidate having high dependency relation is made high. Method. The generation rule used in the character unit candidate generation step is:
4. A rule in which a sentence / phrase including the translated word is extracted from a database including a plurality of sentences described in a translation destination language, and a relationship between the translated word and the sentence / phrase is automatically acquired. Machine translation method. A machine translation device that machine translates a translation source text described in a predetermined translation source language into a translation destination text described in a translation destination language,
Input means for entering the source text,
Feature word extraction means for extracting a feature word having a characteristic meaning from the translation source text;
A translation selection means for translating the feature word into a translation expressed in the target language;
A character unit candidate generating means for generating a character unit candidate from the translated word based on a predetermined generation rule;
Translation destination text candidate generation means for generating one or a plurality of translation destination text candidates assuming the dependency between the character units;
An evaluation means for calculating an evaluation value of each translation destination text candidate when there are at least a plurality of translation destination text candidates;
A machine translation apparatus comprising output means for outputting at least one of translation destination text candidates in relation to the evaluation value. In the machine translation device,
Among the feature words extracted by the feature word extraction unit, the system includes a dependency relationship extraction unit that extracts dependency relationship information between the feature words.
6. The translation destination text candidate is generated using the dependency relation information in the translation destination text candidate generation means for the character unit candidate generated from the feature word having the dependency relation information. The machine translation device described. In the machine translation device,
Among the feature words extracted by the feature word extraction means, the apparatus includes dependency extraction means for extracting dependency relationship information between the feature words,
In the evaluation means,
6. The machine according to claim 5, wherein the dependency relationship information is compared with the dependency relationship in the translation destination text candidate, and the evaluation value of the translation destination text candidate having a high dependency relationship is increased. Translation device. The generation rule used by the character unit candidate generation means is
8. The rule according to claim 5, wherein the sentence / phrase including the translated word is extracted from a database including a plurality of sentences described in a translation destination language, and the relationship between the translated word and the sentence / phrase is automatically acquired. Machine translation device.

Images