JP2003016067A - Processing method for translation of tense, aspect and modality, and system therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately translate tense, aspect and modality of a translation target when performing a machine translation. SOLUTION: A tense/feature set extracting part 12 extracts tense, an aspect and modality (hereafter called tens and the like 4) and a set of features by case from a database 11 on tense, aspect and modality, a machine leaning part 13 learns which feature has a tendency to become which tense from the extracted set and stores a result of the leaning into a leaning result database 14, a feature extracting part 15 extracts the set when Japanese text 3 to be translated is inputted, an estimation processing part 16 for tense, aspect and modality of translation destination estimates that to which tense and the like the feature of the text 3 has the tendency to become from the set in reference to the database 14 to output the estimated tense and the like 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the field of computer-based translation systems, and more particularly, to a translation processing method and translation for translating a tense, an aspect, or a modality using a machine learning method. It is about the system.

[0002]

2. Description of the Related Art In recent years, with the development of the Internet such as WWW (World Wide Web), the need for machine translation continues to increase. In this machine translation, the tense aspect modality is known to be a difficult problem to translate.

The tense-aspect-modality is information about the tense of a verb part (tense), a phase (aspect) such as a progressive form or a complete form, or a modality that is an auxiliary verb equivalent phrase in an English sentence.

Conventionally, the expression of the tenth aspect modality has been handled by a rule created manually. However, in recent years, an approach based on corpus data such as an example-based (k-nearest neighbor) method has also come to be used. In the example-based method, an example database is prepared in which data that describes in which case which tense is used is associated with each collected example, and a tense is created based on an example that is similar to the input sentence.・ Translation of aspect modalities was performed. [Reference 1] Masaki Murata Ma Qing Kiyotaka Uchimoto Hitoshi Isahara, Japanese-English translation of tense aspect modality based on example, Journal of Artificial Intelligence, Vol.16, N
o.1, 2001 In the research described in Reference 1, when determining the tense / aspect / modality of machine translation from Japanese to English, the tense / aspect / modality of Japanese is expressed at the end of the sentence. Focusing on this, a method of determining the tense aspect modality by determining the similarity between the character string of a predetermined length at the end of the input Japanese sentence and the corpus data prepared in advance by the k-nearest neighbor method. I am using. The k-nearest neighbor method is a method in which instead of one most similar case, the most similar k cases are used, and the classification destination is obtained by the majority vote of these k cases.

[0005]

However, in the method of manually describing the rules and classifying the tense / aspect / modality based on the rules, there are problems of human resources and accuracy of manual work. There is.

Further, in the method of using an actual example that is very similar to the input sentence, it is necessary to define the similarity between the input sentence and the actual example. For example, it is easy to define the similarity like a character string at the end of a sentence. I could handle only such information. Therefore, in the method of the research described in Reference Document 1, when the classification of the tense / aspect / modality is judged only by the information of the character string at the end of the sentence, the accuracy of the judgment result may be lowered.

[0007] For example, the example data “I have already gone.”
Suppose that the sentence "I went yesterday." Was input when the tenth aspect modality of was "past completed". Even though the correct tense / aspect / modality of this input sentence is "past", the similarity of the display of the character string "Matte $ ($ = end of sentence)" indicates that "past completed" as in the example data. It may be judged.

Therefore, in addition to the character string at the end of the sentence as in the method of the research in Reference 1, information in a format different from the character string at the end of the sentence, such as "Yesterday" in this case, is combined. It is considered effective to use.

However, as information (features) for analyzing the tenth aspect modality, morpheme information (morpheme feature), semantic analysis information (word feature, etc.), syntactic analysis information (syntactic analysis feature), and other formats are used. Even if it is effective to use a combination of features, it is not possible to use multiple forms of features in the determination method that requires the definition of similarity, such as the k-nearest neighbor method used in the study of Reference 1. There was a problem.

The present invention solves the above-mentioned problems by using a machine learning method capable of handling a case of a conversion source language and a feature of a different format extracted from data related to the case. What tense in case of
It is an object of the present invention to provide a means for learning whether an aspect modality will be obtained, and using the learning result to accurately translate the tense aspect modality of the conversion destination of the input sentence.

[0011]

In order to solve the above-mentioned problems, the method according to the present invention uses a database that stores a case of a conversion source language and a tense / aspect / modality of the case, which are stored in advance. A tense consisting of an aspect modality and a set of features in multiple formats extracted from the case or data related to the case-
Using the process of extracting a feature set for each case, and learning data for determining the tense, aspect, and modality of the target language by a machine learning method using all or some of the features in the above-mentioned tense-feature set. Based on the process of creating and storing, the process of extracting the set of features of the input sentence from the input sentence of the source language, and the features of all or some of the features of the input sentence, A process of estimating the tense / aspect / modality of the input sentence with reference to learning data.

According to the present invention, as in the conventional method, a tense is generated by using only a single type of feature such as a character string at the end of a sentence.
Instead of analyzing the aspect modality, in addition to the character string, morphological features, semantic features, syntactic features of the whole sentence, the tense aspect modality of the preceding sentence, or the corresponding component data of the bilingual data, etc. This is different from the conventional method in that the analysis processing is performed by arbitrarily using the features of different formats.

Further, in the present invention, the definition of the similarity is that the analysis processing is performed by using various machine learning methods in which many types of features can be freely used and the similarity need not be set. This is different from the required determination by a method such as the conventional k-nearest neighbor method.

The processing method or system according to the present invention described above can also be realized by a program that causes a computer to execute the processing steps, means, configurations, and elements thereof. This program can be stored in an appropriate recording medium such as a computer-readable portable medium memory, a semiconductor memory, or a hard disk, provided by being recorded in these recording media, or via a communication interface. It is provided by transmitting and receiving using the communication network of.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. In the present embodiment, the case where the tense aspect modality is translated when translating a Japanese sentence into an English sentence is taken as an example.

FIG. 1 shows an example of the system configuration of the present invention.
In FIG. 1, reference numeral 1 represents a tense aspect modality translation system according to the present invention. Tens Aspect Modality Translation System 1 is composed of CPU, memory, etc.
Tens Aspect Modality Database 11,
It has a tense-feature set extraction unit 12, a machine learning unit 13, a learning result database 14, a feature extraction unit 15, and a conversion destination tense / aspect / modality estimation processing unit 16.

The tense-feature set extraction unit 12 is a tense-aspect-modality database 1 which is a previously prepared corpus for tense-aspect-modality.
It is a means for extracting a set of tense / aspect modality and a set of features of a case from 1 for each case.

The machine learning unit 13 determines, based on the set of the tense / aspect modality and the set of features extracted by the tense / feature set extracting unit 12, what kind of feature / tense / aspect / modality. It is a means for learning whether or not the result is likely to occur by a machine learning method and storing the learning result in the learning result database 14.

The feature extraction unit 15 is a means for extracting a set of features from the input Japanese sentence 3 and passing them to the conversion destination tense / aspect / modality estimation processing unit 16.

The transformation-destination tense-aspect-modality estimation processing unit 16 refers to the learning result database 14, and in the case of the set of the passed features, what tense-aspect-modality in the transformation-destination language becomes. It is a means for estimating whether it is easy and outputting the tense aspect modality 4 of the conversion destination of Japanese sentence 3.

An outline of the processing flow of the present invention will be described. FIG. 2 is a processing flowchart of the system shown in FIG.

Before starting the process shown in FIG.
The aspect modality database 11 is prepared in advance. The Tens Aspect Modality Database 11 is a Japanese-English parallel translation corpus for machine translation,
Information on tense, aspect, and modality is added to Japanese and English bilingual data.

As the classification of the tense aspect modality to be given to the bilingual data in English and Japanese, for example, the following is used. The following classifications are determined by the verb form of the English translations. (1) Each auxiliary verb equivalent phrase (be able to, be going to, c
an, have to, had better, may, must, need, ought, s
(12 kinds of hall, used to, will))
All combinations of {present tense, past tense}, {progressive, non-progressive} and {complete, not complete} (even when there are multiple auxiliary verb equivalent phrases): 2 ¹⁵ types (2) imperative (1 type) (3) Noun phrase (1 type) (4) Participle construction (1 type) (5) Verb omission (1 type) (6) Interjections, greetings, etc. (1 type) (7) Japanese and English If the verb does not correspond to (1 type) (8) work not possible (1 type) However, among the above classifications, the six categories from "(3) noun phrase" to "(8) work not possible" It is not necessary to handle it as a classification of tense / aspect / modality, or it is not necessary to translate tense / aspect / modality, so it is omitted in this embodiment.

Further, these classifications are described in two places: "verb part of main clause of English" and "verb part of English corresponding to verb of main clause of Japanese language". However, in Japanese-English translation, it is considered that the "verb part of the English main clause" may correspond to the Japanese tense / aspect / modality, so in the tense / aspect / modality translation system 1 according to the present invention, Sentence 3 is given to estimate the classification of the tense / aspect / modality of the "verb part of the English main clause" of the conversion destination, and the estimated result, tense / aspect / modality 4, is output.

Step S1: First, the tense-feature pair extraction unit 12 extracts a tense for each case from the prepared tense aspect modality database 11.
Extract a set of aspect modalities and a set of feature cases.

In the tenth-feature set extraction unit 12, as a set of features, a character string feature, a morphological feature, a word feature, a syntactic feature, information on the tense aspect modality of the preceding sentence, a verb portion of English bilingual data, and the like. A predetermined feature can be extracted from various types of features.

FIG. 3 shows a set of features to be extracted and a set of tense / aspect / modality. As shown in FIG. 3, the tense-feature pair extraction unit 12 allows the tense aspect
From the case of modality database 11 “Already registered.”, The tense aspect modality and the character string features “Already registered $”, “U registered $”, ..., “Ta $”, Word features "Already", "Registration"
A set with a set of features such as “shi”, “mashi”, “ta” is extracted. In addition, here, the extracted character string feature is suffixed with $ so that it can be distinguished from the morpheme string of the entire input sentence. In addition, punctuation has been eliminated to normalize sentence end expressions.

Step S2: Subsequently, the machine learning unit 13 makes it easy to obtain what kind of tense / aspect / modality when the set of the extracted tense / aspect / modality and the set of the characteristics is set. Machine learning is performed, and the learning result is stored in the learning result database 14.

In the machine learning, for example, the processing may be performed by using some features among a character string feature at the end of a sentence of a predetermined length, a morphological feature of all sentences of a case, and a word feature.

The machine learning method may be any machine learning method that can handle a set of features in various forms.
For example, the following decision list method, maximum entropy method, support vector machine method, etc. are used.

(1) Decision List Method In the decision list method, the feature f _j (
ΕF, 1 ≦ j ≦ k), the probability value of each classification is obtained by using only one of the features as a context, and the classification having the largest probability value is used as the classification.

The probability of outputting the classification a in a certain context b is given by the following equation.

P (a | b) = p (a | f _max ) (1) where f _max is given by the following equation.

[0034]

[Equation 1]

In addition,

[0036]

[Equation 2]

Is the classification a _i when the feature has f _j in the context
Is the rate of appearance.

Specifically, for each feature, the probability of what kind of tense / aspect / modality classification will be obtained, and using the feature classification with the maximum probability among all the features of the input sentence. Estimate the classification of tense / aspect / modality.

Although the classification based on the decision list method is simple, since the classification of the tense / aspect / modality is estimated in the context of only one feature, it is a little poor machine learning method. .

(2) Maximum entropy method When the set of features f _j (1 ≦ j ≦ k) set in advance is F, the maximum entropy method satisfies entropy while satisfying the following expression (3). Expression (4) meaning
In this method, the probability distribution p (a, b) when maximizing is calculated, and among the probabilities of the respective classifications obtained according to the probability distribution, the classification having the largest probability value is determined.

[0041]

[Equation 3]

However, A and B mean a set of classification and context, and g _j (a, b) is 1 when the context b has a feature f _j and the classification is a, and is 0 otherwise. Means a function. Also,

[0043]

[Equation 4]

Means the rate of appearance of (a, b) in the known data.

Equation (3) is to obtain the expected value of the frequency of the output and the feature set by multiplying the probability p by the function g which means the appearance of the output and the feature set. Entropy maximization (smoothing of probability distribution) is performed with the constraint that the expected value and the expected value calculated based on the probability distribution on the left side are equal, and the probability distribution of output and context is obtained. There is. See references 2 and 3 below for details of the maximum entropy method. [Reference 2] Eric Sven Ristad, Maximum Entropy Mo
deling for Natural Language, (ACL / EACL Tutorial Pro
(gram, Madrid, 1997) [Reference 3] Eric Sven Ristad, Maximum Entropy Mo
deling Toolkit, Release 1.6beta, (http: //www.mnemon
(ic.com/software/memt, 1998) In the use of this method, it is rare that a large number of features that are the same as the features of the input data to be translated originally exist in a known corpus, and the features appear in the corpus. Although the probabilities cannot be used as they are, it is considered that the expected value of feature appearance in the corpus can be used if the probability distribution is smoothed on the assumption that equivalent states have equivalent probabilities.

(3) Support Vector Machine Method The support vector machine method is a method for classifying data consisting of two classifications by dividing a space by a hyperplane. FIG. 4 shows a concept of margin maximization of the support vector machine method. In FIG. 4, white circles represent positive examples, black circles represent negative examples, solid lines represent hyperplanes that divide the space,
A broken line means a surface representing the boundary of the margin area. Figure 4
4A is a conceptual diagram when the positive example and the negative example have a narrow interval (small margin), and FIG. 4B is a conceptual diagram when the positive example and the negative example have a large interval (large margin).

At this time, if the two classifications consist of a positive example and a negative example, the larger the interval (margin) between the positive example and the negative example in the learning data, the lower the possibility of misclassification in open data. Therefore, as shown in FIG. 4 (B), a hyperplane that maximizes this margin is obtained and classification is performed using it.

Although it is basically as described above,
In the training data, the method that allows a small number of cases to be included in the marginal area and the method that makes the linear part of the hyperplane non-linear (the introduction of a kernel function) are used. This extended method is equivalent to classification using the following discriminant function, and two classifications can be discriminated depending on whether the output value of the discriminant function is positive or negative.

[0049]

[Equation 5]

However, x is the context (set of features) of the case to be identified, and x _i and y _j (i = 1, ..., 1, y _j ε)
{1, -1}) means the context of the learning data and the classification destination,
The function sgn is sgn (x) = 1 (x ≧ 0) (6) −1 (otherwise), and each α _i is expressed by the formula (7) under the constraint of the formulas (8) and (9). Is for maximizing.

[0051]

[Equation 6]

The function K is called a kernel function and various functions are used. In the present embodiment, the following polynomial function is used.

[0053]     K (x, y) = (x · y + 1)^d              (10) C and d are constants set experimentally. Specifics described below
In the example C was fixed at 1 throughout all treatments. Also,
For d, two types of 1 and 2 are tried. Where α _i> 0
X_iIs called the support vector, and is usually an expression
The part taking the sum of (5) is calculated using only this example.
Calculated. In other words, in the actual analysis, the
Only the case called the port vector is used.

For details of the extended support vector machine method, refer to References 4 and 5 below. [Reference 4] Nello Cristianini and John Shawe-Tay
lor, An Introduction to Support Vector Machines an
d other kernel-based learning methods, (Cambridge U
niversity Press, 2000) [Reference 5] Taku Kudoh, Tinysvm: Support Vector m
achines, (http://cl.aist-nara.ac.jp/taku-ku//softwa
re / Tiny SVM / index.html, 2000) The support vector machine method handles data with two classifications. Usually, by combining this with a pairwise method, the number of classifications is three or more. You will be dealing with data.

The pairwise method means that, in the case of data having N classifications, every pair of two different classification destinations (N
(N-1) / 2) is made, and which is better for each pair is obtained by a binary classifier (here, by the support vector machine method), and finally N (N-1) / 2 This is a method of obtaining a classification destination by majority voting of the classification destinations of the binary classifier.

The support vector machine as the binary classifier in this embodiment is realized by combining the support vector machine method and the pairwise method, and Tin created by Mr. Kudo according to the following reference document 6 is used.
It uses ySVM. [Reference 6] Taku Kudo Yuji Matsumoto, Support vector mac
chunk identification using hine, Natural Language Processing Research Group, 2000-N
L-140, (2000) Specifically, the hyperplane with the maximum margin is obtained for each classification of tense, aspect, and modality. Then, it is determined whether the input unknown feature belongs to the positive example region or the negative example region, for example, regarding whether the tense is the present (positive example) or the past (negative example). Then, the tense, aspect, and modality are finally estimated by the majority of the judgment results.

Step S3: The Japanese sentence 3 for which the tense aspect modality is to be translated is input.

Step S4: The feature extraction unit 15 extracts a set of features from the input Japanese sentence 3 and estimates them as the transformation destination tense / aspect / modality in the same manner as the process of the tense-feature set extraction unit 12. It is passed to the processing unit 16. FIG. 5 shows an example of the extracted feature set. From the input Japanese sentence "I already went.", The character string features "I already went $", "U went" $, ..., "Ta $" and the word features "Already" and "Go."","Better",
“Ta” and the like are extracted.

Step S5: Based on the learning result database 14, the conversion destination tense / aspect / modality estimation processing unit 16 determines what kind of tense / aspect / modality is likely to occur in the case of the passed feature set. The specified tenth aspect modality 4 is output. For example, data such as "past, completed", "past, completed, progressive", "current, with a challenge", "past; with a beable to" is output.

Using the learning results stored in the learning result database 14 for the set of features of the input Japanese sentence "I'm already gone.", It can be seen that "past completion" is sufficient. Therefore, the tenth aspect modality 4 "Past complete" is output as. In this case, when the estimation is performed simply by using only the character string at the end of the sentence as in the conventional case, the tense is estimated to be “currently completed” from the matching of the expression “Matte $”. However, in the present invention, “past completion” can be correctly estimated by referring to the result of learning using some of the word features of the whole sentence other than the character string.

When the input Japanese sentence 3 is "I went yesterday", similarly, in the conventional method, the tense is estimated to be "currently completed" from the coincidence of the expression "tas $". Will end up. However, if the learning result that the tense is "past" is stored in "I registered yesterday" in the learning result database 14, correctly estimate it as "past" based on the word feature "yesterday". You can

In the above, the case where the morphological feature and the word feature are mainly used as the set of features has been described as an example.
The feature set extraction unit 12 may extract not only the morphological feature and the word feature but also the tense aspect modality of the preceding sentence (case) as a feature. This is a tense
Uses the property of being easy to continue. That is, if the tense given to the preceding sentence is "present",
It is learned that it is better to describe the target case as “current”. This is especially useful when it is necessary to unify the tense in the description of the experiment in the paper.

Further, the tense-feature set extraction unit 12 may extract the corresponding word / phrase (verb part) of the English parallel translation data of the case as a feature. This is based on the fact that the modality to be used may change when the structure of the translated English sentence changes.

The case will be described below. Example sentence 1) Japanese sentence: He {lives} a sober and simple life. Example sentence 2) Japanese sentence: He is {lives} {lives} a sober and simple life. {Is leading} a lazy life out of habit. For example, the modality of "sending" in example sentence 1 is "present tense", but the modality of "sending" in example sentence 2 is "progressive". Although these can be thought of as sentences that have almost the same meaning and have the same modality, such a difference can be seen only by changing the verbs used for translation to “live” and “lead”.

In this case, a provisionally translated English sentence is added to the input Japanese sentence 3, and the feature extraction unit 1
In 5, the verb part of the English sentence added to the Japanese sentence 3 is extracted as a feature.

When high quality processing is desired, the structure on the English side assumed by the structure analysis unit of the machine translation system in the English parallel translation data is assumed not only in the Japanese side but in the feature extraction as described above. Extracting (or verb) is effective in improving the translation of tense, aspect, and modality.

Hereinafter, the present invention will be described in more detail with reference to an embodiment.

[Tense Aspect Modality Database] FIG. 6 shows a part of the bilingual corpus which is the tenth aspect modality database 11 used in the embodiment. This parallel translation corpus is created, for example, based on the following reference document 7. [Reference 7] Maki Murata Masao Uchiyama Kiyotaka Uchimoto Masei
Hitoshi Isahara, Modification of modality corpus for machine translation using machine learning, 7th Annual Conference of Linguistic Processing Society, (2001) In Fig. 6, English sentences have the following two tags. There is.・ Enclose the verb part of the English main clause with the tags <v> and </ v>.・ <V for the English verb part that corresponds to the Japanese main clause verb
Enclose with tags of j> and </ vj>.

Further, symbols such as "c" and "d" are added to the head of the sentence on the Japanese side, which means the tense aspect modality of this bilingual data. For example, "c" means can and "d" means past tense.

The first data of the corpus shown in FIG. 6 has ",", which is used when <vj> is used. The verb enclosed by <v> to the left of ",". The tense aspect modality for is described, and the tenth aspect modality for the verb surrounded by <vj> is described on the right. In this corpus, the present tense often appears, so in that case, no tag was added. Therefore, the left and right of "," are blank, and only "," is added to this part.

Further, when the "English verb part corresponding to Japanese" and the "English main verb part" match, only the "English main verb part" tag is added. Also, the "English verb part corresponding to Japanese" is not so carefully tagged, and the "English verb part corresponding to Japanese" and the "verb part of the English main clause" match. In some cases, the tag may not be added even when the tag is not attached.

The following (1) and (2) of the above classifications were used as the classifications of the tense aspect modality to be added to the bilingual data in Japanese and English. (1) Each auxiliary verb equivalent phrase (be able to, be going to, c
an, have to, had better, may, must, need, ought, s
(12 kinds of hall, used to, will))
All combinations of {present tense, past tense}, {progressive, non-progressive} and {complete, not complete} (even when there are multiple auxiliary verb equivalent phrases): 2 ¹⁵ types (2) imperative (1 type) The classification of the tense / aspect / modality handled by the system according to the present invention is defined based on the surface expression in English. If only Japanese sentences can be given to estimate this classification,
Japanese-English translation of modality expressions is completed. For this reason, in this example, in principle, only tags indicating the classification of tense, aspect, and modality and Japanese sentences are used. Also,
As mentioned above, these classifications are referred to in two places, "the verb part of the English main clause" and "the English verb part corresponding to the verb of the Japanese main clause". The problem is to estimate the tense / aspect / modality classification of the “verb part of the English main clause”.

In this example, the following two types of bilingual corpus are used as the tenth aspect modality database 11.・ Example sentence of K-Japanese dictionary (total number of cases is 39,660, total number of classification is 46) ・ White paper data (total number of cases is 5,805, total number of classification is 30)
Each of these corpora is tagged with checking manually, and is also created by using the corpus correction method shown in Reference 7 and Reference 8 below, which is extremely accurate. ing. [Reference 8] Maki Murata Masao Uchiyama Kiyotaka Uchimoto Masei
Hitoshi Isahara, Decision List, Corpus Error Detection / Error Correction Using Example-Based Method, Natural Language Processing Research Group, 2000-NL-
136 (2000)

[Features to be extracted] In this example, when the input of Japanese sentence 3 is given, the classification is output as the tenth aspect modality 4. Therefore, the feature is taken out from the input Japanese sentence 3. Here, we processed the following three types of feature sets. (1) Feature set F1 A character string of 1 to 10 gram at the end of a Japanese sentence and a morpheme sequence of the entire input sentence are set as a feature set. Example: “No $” “Did not do $” “Today” “Ha” “Run” In this case, the number of features is 230,134 in K company data.
The number of white paper data is 25,958. (2) Feature set F2 A character string of 1 to 10 gram at the end of a Japanese sentence is set as a feature set. Example: “None $” “Don't do $” In this case, the number of features is 199,199 in K company data.
The number of white paper data is 16,610. (3) Feature set F3 The morpheme string of the entire input sentence is used as a feature set. Example: “Today” “Ha” “Run” In this case, the number of features is 30,935 in K company data.
The number of white paper data is 9,348.

JUMAN was used to decompose the input sentence into morpheme strings. A detailed description of JUMAN can be found in Reference 9 below. [Reference 9] Sadao Kurohashi, Shin Nagao, Japanese Morphological Analysis System JUMAN Instruction Manual, version 3.6 (Kyoto University Graduate School of Engineering, 1998) The feature set F1 is a combination of the feature set F2 and the feature set F3. is there. The feature set F2 was created by referring to the research in Reference 1 above, and the expression indicating the tense, aspect, and modality in a Japanese sentence often appears in the verb at the end of the sentence. Character string is the feature. The feature set F3 is an expression in which adverbs such as "tomorrow" and "yesterday" also indicate tense, aspect, and modality, and was created with the thought that it should be used.
This is a morpheme string of the entire input sentence. [Classification of Tens, Aspect, and Modality by Machine Learning] In this example, as a method of machine learning, the decision list method,
The maximum entropy method and the support vector machine method were used. Further, processing using the k-nearest neighbor method was also performed for comparison of processing results between the machine learning method used in the system according to the present invention and the conventional method.

The k-nearest neighbor method needs to define not only the feature set but also the similarity between cases. However, in this example, since the feature set F1 and the feature set F3 include the morpheme sequence of the entire input sentence as a feature set, it is difficult to define the similarity. Therefore, only the feature set F2 is used in the k-nearest neighbor method. As the definition of similarity in the feature set F2,
When the longest matching character string between cases is x-gram,
The similarity is decided to be x.

As another machine learning method, there is a method of utilizing decision tree learning such as C4.5. In this example, the decision tree learning method is different from other methods due to various problems. The problem to be dealt with in this example is not used for two reasons: the number of types of attributes is large, and the accuracy will decrease if the number of attributes is reduced until C4.5 can be executed.

[First Example] First, a process using the example sentence data of the K Company Japanese-English dictionary was performed. The accuracy of the processing result is shown in FIG. In this example, two types of processing, closed and open, were performed. The open experiment was carried out by 10-fold cross validation. The numbers in parentheses in FIG. 7 mean the accuracy at close.

The following can be seen from the processing result. The decision list method obtains the same degree of accuracy as the k-nearest neighbor method when using the feature set F2. -The maximum entropy method has higher accuracy than the k-nearest neighbor method or the decision list method.・ The support vector machine method always gives higher accuracy than other methods. As a comparison of the feature sets, the maximum entropy method and the decision list method have the highest precision in the feature set F2, and conversely the precision decreases when the morpheme information like the feature set F1 is added. It is considered that this is because the accuracy decreases because the number of unnecessary features increases even if the number of features increases.・ In the comparison of feature sets by the support vector machine method,
The feature set F1 has the highest accuracy. This means that the addition of morphological information was effective in the support vector machine method. On the other hand, the accuracy of addition of morpheme information decreased with other methods. Therefore, it is assumed that the support vector machine method has a higher ability of feature selection to remove unnecessary features and select useful features than other methods. It

The following explanation can be given to this result from the theoretical side of the method.・ Since the decision list method is a method of finding a solution from only one feature, if there are many unnecessary features, it is easy to find a solution with only those unnecessary features as contexts. If there are many unnecessary features, the accuracy is high. descend. -The maximum entropy method always uses almost all features, so the accuracy decreases when there are many unnecessary features. -On the other hand, in the support vector machine method, there is an operation that discards cases in which only the cases that are the support vectors are used and other cases are not used, so many unnecessary features are discarded together with this case, which is unnecessary. Even if there are many such features, there is a tendency that the accuracy is not lowered so much.

As described above, the most accurate method among all the methods is the support vector machine method when d = 1 and the feature set F1.

Among the above results, in the support vector machine method, using the feature set F1 is the feature set F2.
We performed a sign test to see if the result that it was better than using, that is, the addition of morphological information was effective was significant. This was done with d = 1 because the precision was better when d = 1. All cases 39,6
Of the 60 cases, there were 648 cases where the feature set F1 was correct and the feature set F2 was erroneous, and 427 cases where the feature set F2 was correct and the feature set F1 was erroneous. 0.00000001% (8 in calculation
The actual value may be much smaller than this value because it was divided into digits. ) It was judged that there was a significant difference in the following risk factors. Therefore, it is almost certain that the support vector machine method had the effect of adding morpheme information.

Next, it was decided to investigate what kind of features actually worked effectively even though it was morpheme information.
This is done by checking the features that are biased to 648 in the case where the feature set F1 is correct and the feature set F2 is incorrect. Here, using the binomial test, it is assumed that the ones judged to be larger at the significance level of 1% than the appearance probabilities in all 39,660 cases are biasedly appearing.

FIG. 8 shows the top 20 most frequently occurring morphological features that seem to have worked effectively. In Figure 8,
Morphological features that can be useful for estimating tense aspect modality, such as "already", "recently", "would", "still", "if not", "let's", and "tomorrow" have been obtained, and actually such morphological features have been obtained. It is presumed that the accuracy was improved by the feature.

[Second Example] Next, processing was performed using white paper data. In this case, the support vector machine method, which had good accuracy, was used. Also in this example, open precision is obtained by performing 10-fold cross validation. The accuracy of the processing result is shown in FIG.

The following can be seen from this processing result.・ The accuracy of the white paper data was 64.67% at maximum.・ Feature set F2 that uses only end-of-sentence character strings even in white paper data
Feature set F that uses morpheme information of the whole sentence rather than
1 is higher accuracy. Further, in the white paper data, the feature set F3 using the morpheme information of the entire sentence has higher accuracy than the feature set F2. These results further confirm the effectiveness of the whole sentence as morphological information features.

[Third Example] Next, a process was performed in which K company data was used as learning data and white paper data was used as test data, that is, data in a different field was used as teacher data. With this example, we were able to investigate how the accuracy changes when using data from different fields. In this processing, the support vector machine method of the feature set F1 was performed on the data of d = 1 and d = 2, which had good accuracy. Even in this process, if the learning data and the test data overlap, the open precision is obtained by performing 10-division cross validation in the overlapped portion. The accuracy of the processing result is shown in FIG.

The results of this processing reveal the following. -Using data from different fields resulted in very poor accuracy. (When the white paper data is used as the learning data to analyze the K company data or the K company data is used as the learning data to analyze the white paper data, the accuracy drops to about 10% to 20%.) It turns out that it is effective to use learning data in the same field as the data.

It is difficult to make a system adapted to a different field by the method using the rules written by hand. On the other hand, with the method using machine learning as in the present invention, it is easy to create a system adapted for each field by changing learning data and re-learning.・ If both company K and the white paper are used as learning data,
The accuracy was almost unchanged or slightly lowered. From this, it can be seen that it is not necessary to have a large amount of learning data, and in the case of data of different fields, it is not so effective to use learning data mixedly.

In this example, the processing of translating the tense aspect modality was performed using various machine learning methods including the k-nearest neighbor method. Also, for verification, we investigated which of the machine learning methods is the best.

In the conventional method (reference document 1 and the like), only the character string at the end of a sentence is used as a feature when translating a tense aspect modality. In the present invention, in addition to the character string at the end of a sentence, morpheme information in the sentence is additionally used.

As a result, it was confirmed using a test that the use of morpheme information in one sentence, which has not been used conventionally, has the effect of improving processing accuracy.

Further, even if any one of the decision list method, the maximum entropy method, and the support vector machine method is used as the machine learning method, the tense / aspect / modality method can be used with higher accuracy than the conventional k-nearest neighbor method. I was able to translate.

In particular, the method using the support vector machine method has the highest accuracy, and the tense aspect modality can be translated with higher accuracy than the method using the k-nearest neighbor method of the conventional method.

Further, an example of processing using corpus of different fields (in this example, the English-Japanese dictionary data of K company and white paper data) was performed. In this process, it was confirmed that the accuracy was significantly reduced when using data from different fields, and it was confirmed that it was necessary to construct a translation system of tense, aspect, and modality for each different field. This shows the usefulness of the present invention using the machine learning method considering that it is difficult to manually create a system adapted to a different field.

Although the present invention has been described with reference to the embodiments, the present invention can be variously modified within the scope of the gist thereof. For example, in the embodiment of the present invention, the tense aspect modality translation system 1 according to the present invention is used.
Has been described as being configured independently, but can be configured as part of another machine translation system.

The machine learning method used by the machine learning unit 13 is not limited to the decision list method, the maximum entropy method, and the support vector machine method, but any method can be used as long as it can handle features of different forms in combination. The method may be used, and the features extracted by the tense-feature pair extraction unit 12 or the feature extraction unit 15 are not limited in kind as long as they are features that can be extracted from the target case or input sentence. is there.

[0098]

As described above, the present invention has the following special effects. The present invention uses a machine learning method capable of handling analysis information (features) in a plurality of formats, which is represented by the support vector machine method, due to the problem of translating tense aspect modality. As a result, the tenth aspect ratio can be calculated with higher accuracy than the conventional k-nearest neighbor method using similarity.
A tense / aspect / modality translation processing method and translation system capable of performing modality translation can be provided. In the present invention, the morpheme information in the sentence is newly added in addition to the character string at the end of the sentence when the tense aspect modality is translated. As a result, it is possible to perform highly accurate translation as compared with the conventional method in which only the character string at the end of a sentence is used as a feature. In the present invention, a tense for extracting a set of a set of tense / aspect / modality and features used in the machine learning unit 13
As an aspect modality database, corpora in various fields can be used, and learning results can be acquired based on the corpora without human intervention. This makes it possible to easily realize a tense / aspect / modality translation system that adapts to different fields.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration example of a system according to the present invention.

FIG. 2 is a processing flowchart of a system according to the present invention.

FIG. 3 is a diagram showing an example of a set of a set of a tenth aspect modality and a feature.

FIG. 4 is a diagram for explaining margin maximization in a support vector machine method.

FIG. 5 is a diagram showing an example of extraction of a feature set from an input sentence.

FIG. 6 is a diagram showing an example of a part of a tenth aspect modality database in an embodiment.

FIG. 7 is a diagram for comparing the translation accuracy of the tenth aspect modality in the first example.

FIG. 8 is a diagram showing an example of a morphological feature that seems to have worked effectively.

FIG. 9 is a diagram for comparing the accuracy of the translation processing of the tenth aspect modality in the second example.

FIG. 10 is a diagram for comparing the accuracy of the translation processing of the tenth aspect modality in the third example.

[Explanation of Codes] 1 Tence Aspect Modality Translation System 11 Tence Aspect Modality Database 12 Tence-Feature Pair Extraction Unit 13 Machine Learning Unit 14 Learning Result Database 15 Feature Extraction Unit 16 Destination Tense Aspect Modality Estimate Processing Department

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] July 29, 2002 (2002.7.2)
9)

[Procedure Amendment 1]

[Document name to be amended] Statement

[Name of item to be amended] Claims

[Correction method] Change

[Correction content]

[Claims]

[Procedure Amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0011

[Correction method] Change

[Correction content]

[0011]

In order to solve the above-mentioned problems, the method according to the present invention is a computer-based language.
When translating from a source language into another language,
Translate the Tens, Aspect, and Modality in the target language
A method, which is a database that stores a set of a case of a conversion source language and a tense aspect modality in a conversion destination language of the case, which is a database.
The process of accessing the Pect Modality database
And the tense aspect modality database
Scan for each instance of, and a set of multiple types of feature including word identity and String feature extracted from case corresponding to the Tenth Aspect modality and the Tenth Aspect Modality Tense - feature sets a process to extract the said tense - for the appearance patterns of a plurality of feature using the feature set as teacher data, for each pattern
Tens, Aspect, and Modality
Decision list method, maximum entropy method, or support vector
Learning by any one of the cuttle machine methods
And the learning result in the machine learning process,
Tense, aspect, modality of the target language of the input sentence
As the learning data for determining
And the process to save the scan, from the input sentence of the source language, and the process of extracting a set of identity of the input sentence, on the basis of the collection of the identity of the input sentence, by the machine learning method, the learning data with reference to the learning data stored in <br/> data base of the identity
Regarding the pattern of appearance of set features,
Of the input sentence, and outputting as an estimated solution of the tenth aspect modality of the input sentence.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Kiyotaka Uchimoto             4-2-1 Kanaikitamachi, Koganei City, Tokyo Independent             Communications Research Laboratory (72) Inventor Hitoshi Isahara             4-2-1 Kanaikitamachi, Koganei City, Tokyo Independent             Communications Research Institute F term (reference) 5B091 AA15 AB13 BA03 CC03 CC16                       EA01

Claims (4)

[Claims] 1. A method for translating a tense aspect modality of a source language from a source language when translating from one language to another language by a computer, the source language being provided in advance. -A feature consisting of a set of features in multiple formats extracted from the tense / aspect modality and the case or data related to the case from a database that stores the case and the tense / aspect / modality of the case. Using the process of extracting a set for each case and all or some of the features in the tense-feature set, create learning data for determining the tense / aspect / modality of the target language by a machine learning method. Then, the process of storing the features, the process of extracting the set of features of the input sentence from the input sentence in the source language, and the feature set of the input sentence. A process of estimating the tense / aspect / modality of the input sentence with reference to the learning data based on all or some of the features Method. 2. The tense aspect according to claim 1.
In the modality translation processing method, any one of a decision list method, a maximum entropy method, and a support vector machine method is used as the machine learning method, and the tenth aspect modality translation processing method is characterized. 3. A system for translating a tense aspect modality of a source language from a source language when translating from one language to another language by a computer, the source language being provided in advance. -A feature consisting of a set of features in multiple formats extracted from the tense / aspect modality and the case or data related to the case from a database that stores the case and the tense / aspect / modality of the case. To determine the tense / aspect / modality of the target language by machine learning using the tense-feature set extraction means for extracting the set for each case and all or some of the features of the tense-feature set Machine learning means for creating and storing learning data of and the set of features of the input sentence is extracted from the input sentence of the source language. Based on all or some of the features of the input sentence, and a transformation destination tense for estimating the tense aspect modality of the input sentence by referring to the learning data. A tense aspect modality translation system, comprising: an aspect modality estimation processing means. 4. The tense aspect according to claim 3,
In the modality translation system, the machine learning unit performs learning using any one of a decision list method, a maximum entropy method, and a support vector machine method, which is a tenth aspect modality translation system.