KR20080052282A - Apparatus and method for unsupervised learning translation relationships among words and phrases in the statistical machine translation system - Google Patents

Abstract

An apparatus and a method for autonomously learning translation relationships among words and phrases in a statistical machine translation system are provided to enhance correctness in arranging phrases by using unified results of arranging multiple words arrangement and a result of arranging commonly identical words. An apparatus for autonomously learning translation relationships among words and phrases includes a source language sentence preprocessor(101), a target language sentence preprocessor(102), an autonomous learning device(200), a learning terminating condition inspecting unit(210), a statistical machine translation model parameter extracting unit(300), and a decoder. The source language sentence preprocessor receives a sentence written in a source language whose morpheme and phrase are analyzed. The target language sentence preprocessor receives a sentence written in a target language whose morpheme and phrase are analyzed. The autonomous learning device receives learning sets from the source language sentence preprocessor and the target language sentence preprocessor. The learning terminating condition inspecting unit repeats word arrangement, word rearrangement, phrase arrangement, and word & phrase translation dictionary acquisition until a learning termination condition that there is no more change in the word & phrase translation dictionary is satisfied. The statistical machine translation model parameter extracting unit extracts parameters for a statistics based translation model from a word & phrase arrangement result acquired from learning the word & phrase arrangement. The decoder uses the statistical machine translation model for generating a sentence written in a target language from a sentence written in a source language inputted together with a learned language model(400).

Description

Apparatus and method for unsupervised learning translation relationships among words and phrases in the statistical machine translation system}

1 is a schematic block diagram of an exemplary machine translation system to which a learning apparatus according to the present invention is applied.

FIG. 2 is a more detailed block diagram of the autonomous learner shown in FIG. 1.

* Explanation of symbols for main parts of the drawings

101: source language sentence preprocessor

102: target language sentence preprocessor

200: self-study

201: Word Sorter

202: word rearranger

203: syntax sorter

204: Band Dictionary Generator

210: learning end condition checker

300: Statistical Machine Translation Model Parameter Extractor

400: language model

500: decoder

The present invention relates generally to machine translation systems, and more particularly, to an apparatus for autonomously learning translation relationships between words and phrases based on statistical methods, which can be applied to machine translation systems.

Machine translation system is a system that receives text composed of sentences of source language, translates it into target language and outputs the result. Learn phrase band dictionaries and acquire translation knowledge for other translations. This approach of deriving band dictionaries from a parallel set of sentence-aligned learning data measures the degree of association between words in the source language and words in the target language in the sorted sentences, and from all word pairs where the degree of association exceeds the threshold. Form a band dictionary. For example, in the conventional approach, the measure of the degree of association between words is based on how often words appear simultaneously in parallel sentences (or corresponding regions). Association scores for different word pairs are calculated, sorted in descending order of association scores, and bandword pairs are selected and added to the band dictionary according to the selected threshold. However, since this method independently calculates an association score between words, it may cause an error of incorrectly setting a band relationship between constituent words of a compound word. For example, the English bandword of Korean "file system" is "file system" and the bandword of "system file" is "system files". If these compound word pairs appear in many sentences, it is also possible that the band word pairs of the constituent words of the compound word are formed as "file / system" or "system / files".

One attempt to solve this problem is Melamed's "Automatic Construction of Clean Broad-Coverage Translation Lexicon (Second Conference of the Association for Machine Translation in the America's (AMTA, 1996)). If these higher pairs of words are derived from an array of more relevant pairs of words that contain one or both of the same words, then choose the one with the highest correlation score in translation. However, when the set of learning data is small, there is a problem that the reproducibility is lowered due to lack of data.

Another problem in learning band relationships between words is to include compound and compound words. In the prior art, it is assumed that the translation relationship of a vocabulary contains only one word, but in the case of compound words, this is obviously false. For example, "a pool of water" is a band word of "water pool", where four words correspond to one word. One attempt to solve this problem is by melamed, which includes "Automatic Discovery of Non-Compositional compounds in Parallel Data, (Conference on Empirical Methods in Natural Language Processing (EMNLP-97)". A model, namely, a test translation model including a candidate compound word and a reference translation model without it, was derived, and if the value of the objective function was higher in the test model than in the reference model, the compound word was considered valid and otherwise invalid. However, this method has a problem in that it is very complicated and requires a lot of computation since it is made through the construction of a test translation model.

On the other hand, studies using syntax information have received less attention because of the cost of acquiring knowledge for parsing and the lower accuracy of the parser, but assuming that the translation of the source language phrase is a sequence of adjacent words in a target language sentence. There has been relatively much research on fixed phrase translation techniques for translating phrases (eg, noun phrases). "Termight: Coordinating Humans and Machine in Bilingual Teminology Acquisition" by Dagan and Church (Machine Translation, 12: 89-107, 1997) is a proactive assistant tool that predicts technical terms from source language sentences and suggests them to users. For example, we propose a bandword by extracting a pair of possible bandwords from parallel corpus for those approved by users.

Robert Moore also identified in the source language in "Statistical Method and Apparatus for Learning Translation Relationships among Phrase," Patent Publication No. 2004-0044176. Generates a translation translation candidate in the target language by receiving an ordered pair of sentences containing the extracted target word extraction syntax, and internal component and source language syntax and target candidate based on the association between the target language syntax and the internal words of the source language syntax. Using the method of calculating the score for each candidate phrase including the external component based on the association between the external words of the phrase, the translation relationship between the phrases is learned and the band dictionary is derived. However, this method uses a method of setting candidate candidates for a target phrase when the source phrase to extract the bandword is already defined (captoids composed of uppercase letters). Correlation scores by words in external phrases are difficult to apply, and also for common phrase types because they generate bandword candidates for a phrase in the input statement and calculate an association score between words inside and outside the source and target phrases. There is a problem in that the influence of the increase is low accuracy of the selection of the band word.

As in the prior art described above, in Korean, in which one or more morphemes are interlaced and refracted to form words (words), words are used as basic units of translation, which is a serious problem in statistical-based approaches using learning sets. In order to solve the problem of possible data sparseness of a learning set, the present invention aims to alleviate this problem of data shortage by using information divided into morphological units.

In addition, the present invention by using parts of speech or circular information with morphemes in both the source language and the target language to be able to distinguish between homomorphic problems or syntactic use that can occur when using only surface morpheme information, Identify and use relationships statistically.

Furthermore, the present invention automatically expands by utilizing a syntax band dictionary derived through a learning process, a parallel learning set of sentences or phrases that are indispensable for constructing a statistics-based machine translation system. High quality word band dictionaries, which can be effectively used to increase the learning efficiency and accuracy of unsupervised learning, also provide a mechanism that can be automatically learned and reused through the learning process.

An apparatus for autonomously learning a translation relationship between words and phrases in a statistical machine translation system, in accordance with an aspect of the present invention, receives a form-parsed source language sentence, and the form-parsed source language sentence. Receiving a source-language sentence preprocessor for transforming the target-language sentence with the form-syntactic feature information into a tokenized source language sentence, and receiving the form-parsed target language sentence, A target language sentence preprocessor that converts the tokenized target language sentence with the target feature information into the tokenized target language sentence, and receives the tokenized source language sentence and the tokenized target language sentence, thereby receiving the tokenized source language sentence and the token. A self-learning unit performing word sorting and phrase sorting on the targeted target language sentence pairs, the word sorting and Including the learning termination condition checker for repeatedly performing the door alignment.

According to an aspect of the present invention, a method for autonomously learning a translation relationship between words and phrases in a statistical machine translation system includes: receiving a form-parsed source language sentence and a form-parsed target language sentence. And dividing each of the form-parsed source language sentences and the form-parsed target language sentences into morphemes or word units and attaching form-syntactic feature information to each morpheme or word. Tokenizing the morpheme or word with syntactic feature information as a basic unit of translation, sorting the words in the tokenized source language sentence with the words in the tokenized target language sentence, and the words in the tokenized target language sentence. Is sorted by the words of the tokenized source language sentence, and applied to the form-syntax feature information applied to the source language and the form-syntax feature information applied to the target language. Acquiring a plurality of sets of word alignments in sentence units, and performing phrase alignment by taking word alignment results common to multiple word alignment sets as initial alignment values, and performing unaligned source and target languages. Computing band scores for the content words of a phrase and reordering the words by selecting the word with the highest band score; and using phrase information, band phrases of one or more source phrases and one or more target phrases. Sorting, generating word and phrase band dictionaries from the word reordering and phrase sorting results, and then sorting words, reordering words, and phrase sorting until there are no more changes to the word and phrase band dictionaries. And repeating the word and phrase band dictionary generation step.

The present invention preferably uses Korean as the source language and English or Chinese as the target language. Hereinafter, with reference to the accompanying drawings will be described with reference to the embodiment in which Korean is the source language and English is the target.

1 is a schematic block diagram of an exemplary machine translation system to which a learning apparatus according to the present invention may be applied. Source language sentence preprocessor 101 receives the stemmed-parsed source language sentence, and target language sentence preprocessor 102 receives the stemmed-parsed target language sentence. The source language sentence preprocessor 101 divides the input Korean into morpheme units and attaches circular, part-of-speech, relative position information and syntax information (words in dependency) to each morpheme. Similarly, the target language sentence preprocessor 102 divides the input English into words, and attaches the relative position information of the words in the prototype, the part of speech, and the basic phrase for each word. Korean and English sentence pairs attached with such form-syntactic feature information are tokenized as basic units of translation of morphemes and prototypes, morphemes and parts of speech, prototypes and parts of speech, prototypes, and morphemes. Construct a separate learning set that uses units as tokens. As such, Korean morphemes and English words with circular-part-of-speech information attached are recognized as tokens that are basic units of translation, and Korean and English with form-syntactic feature information are given as inputs to the autonomous learner 200. . In this way, in Korean, where one or more morphemes are interlaced and refracted to form words (words), information divided into morphological units can be used to alleviate data problems that can be serious when using word units. In addition, by using part-of-speech or prototype information in both Korean and English together with morphemes, it is possible to discriminate between homomorphic problems and syntactic usage that may occur when using surface morphological information alone, and to statistically grasp the relationship between bandwords. It can be utilized.

Referring to FIG. 2, the components of the autonomous learner 200 are shown in more detail. The autonomous learner 200 receives a learning set from the source language sentence preprocessor 101 and the target language sentence preprocessor 102, respectively. The word sorter 201 sequentially applies and learns IBM models 1, 2, 3, and 4 to each received learning set to obtain a result of sorting words in a source language sentence into words in a target language sentence. In addition, a result of sorting the words of the target language sentence by the words of the source language sentence is obtained. As a result, a plurality of word alignment learning results are obtained first in sentence units according to the morpheme feature information applied to the source language and the morpheme feature information applied to the target language.

The word rearranger 202 receives a sentence unit word sorting set for each morpheme feature information that is primarily learned. The word rearranger 202 performs a phrase alignment by taking the word alignment result that the word alignment results for each feature information of the source language morpheme have in common as an initial alignment value, and performs the contents of the syntax of the unaligned source and target language. Band scores are calculated for word reordering for content words. In this case, the band score includes a conditional probability that a word of a source language is translated into a word of a target language, a conditional probability that a word of a target language is translated into a word of a source language, and an amount of mutual information information between words of a source language and a word of a target language (KL- divergence) is computed and weightedly combined. The conditional probability and mutual information amount are calculated and weightedly combined for each type of token (combination type of morphological feature information used), and the band score of the target word for the source language is calculated. Calculate and rearrange the words by selecting the word with the highest band score. As such, the word reordering result for each sentence pair is given as input to phrase sorter 203.

The phrase sorter 203 utilizes syntax information attached to each input sentence, and one or more source phrases and one or more target phrases may be sorted into band phrases, and a disjoint phrase and a continuous phrase or a band of discontinuous phrases may be used. Relationships are also possible.

The band dictionary generator 204 generates a word band dictionary and a phrase band dictionary including frequency information from the word rearrangement result and the phrase sort result from the word rearranger 202 and the phrase sorter 203. In this case, the frequency information included in the word band dictionary includes the word frequency of the source language, the word frequency of the target language, and the air frequency of two words. The frequency information of the phrase band dictionary is also constructed in the same manner. The generated word band dictionary and phrase band dictionary are not used as a band dictionary as it is. Through the filtering process, only band pairs having a certain reliability are present in the band dictionary, and band pairs below the threshold reliability are removed from the band dictionary. . The built-in word band dictionary is reused to improve the speed and accuracy of word sort learning, and the built-in phrase band dictionary expands the learning set and the range of word sorting to train statistics-based translation models through word sort and phrase sort. It is reused to limit alignment and improve alignment speed and accuracy.

The end-of-learning condition checker 210 performs word alignment, word reordering, phrase ordering, and word and phrase banding dictionaries for the learning set until the end-of-learning condition is no longer changed in the word and phrase banding dictionary from the learning set. By repeating the acquisition process, the quantity and quality of the word and phrase band dictionary are improved. When the word and phrase band dictionary is no longer changed, the statistical machine translation model parameter extractor 300 extracts the parameters for the statistical-based translation model from the word and phrase alignment results obtained from the learning result of the word and phrase alignment. do. In addition, the statistical machine translation model is used in the decoder 500 to generate sentences of the target language from the input language sentences input together with the separately trained language model 400.

Word and phrase rearrangement and word and phrase band dictionary generation through repetitive self-study provided by the present invention combines various form-phrase feature information to make a decision for word and phrase rearrangement, and measures reliability based on statistics. And a phrase band dictionary, thereby improving the accuracy of word and phrase alignment over the prior art. In addition, unlike the existing method, the syntax is sorted using the boundary information of the syntax, but the syntax alignment is performed based on the common word matching result by synthesizing the multiple word sorting results, thereby improving the accuracy of the syntax sorting.

According to the present invention, rather than manually forming a word and phrase band dictionary that is a clue of learning, it is automatically acquired and used through the learning process of the word and phrase band dictionary, and the band information below the reliability is removed through filtering. By minimizing the errors that occur in the self-learning process, it is possible to effectively acquire the translation knowledge necessary for increasing the efficiency of learning and automatic translation. Also, the acquired word and phrase band dictionaries are useful translation knowledge, and can be used not only in statistical-based translation systems but also in rule-based and pattern-based machine translation systems. In addition, statistical-based machine translation systems require large amounts of sentence and phrase aligned corpus, and building sentence and phrase aligned corpus is also time consuming and costly. The present invention has the effect of gradually increasing the learning set by using the high quality syntax band information obtained in the iterative self-learning process, and alleviating the problem of data shortage occurring in the self-learning process.

Claims (14)

An apparatus for autonomously learning a translation relationship between words and phrases in a statistical machine translation system, A source language sentence preprocessor for receiving a form-parsed source language sentence and for converting the form-parsed source language sentence into a tokenized source language sentence with shape-syntactic feature information; A target language sentence preprocessor for receiving a shape-parsed target language sentence and converting the shape-parsed target language sentence into a tokenized target language sentence with shape-syntactic feature information attached thereto; An autonomous learner configured to receive the tokenized source language sentence and the tokenized target language sentence and perform word alignment and phrase alignment on the tokenized source language sentence and the tokenized target language sentence pair; A learning end condition checker for repeating the word sorting and phrase sorting of the autonomous learner. Apparatus for autonomously learning a translation relationship between words and phrases, comprising. The method of claim 1, The source language is Korean and the target language is English or Chinese, and the form-syntactic feature information is translated between words and phrases, including morphemes or word prototypes, parts of speech, relative positional information within basic phrases, and syntax information. Device for learning relationships autonomously. The method of claim 2, The autonomous learner sorts the words of the tokenized source language sentence with the words of the tokenized target language sentence and the words of the tokenized target language sentence with the words of the tokenized source language sentence, To autonomously learn translation relationships between words and phrases, including a word sorter that obtains a plurality of word alignment sets in sentence units according to the form-syntax feature information applied to the source language and the form-syntax feature information applied to the target language. Device for. The method of claim 3, The self-learner performs phrase alignment by taking the word alignment result which the plurality of word alignment sets have in common as an initial alignment value, and extracts content words of syntax of the unaligned source and target languages. And a word rearranger that calculates a band score to a subject and selects the word having the highest band score to reorder the words. The method of claim 4, wherein The autonomous learner further includes a phrase sorter for sorting one or more source phrases and one or more target phrases into band phrases using the phrase information. The apparatus for autonomously learning a translation relationship between words and phrases. The method of claim 5, And the autonomous learner further comprises a band dictionary generator for generating words and phrase band dictionaries from the word reordering result and the phrase sorting result. The method of claim 6, The band dictionary generator performs a filtering process through a reliability measurement, and only band pairs having a certain reliability or more exist in the word and phrase band dictionary, and band pairs below a threshold reliability are removed from the word and phrase band dictionary. Device for autonomously learning the translation relationship of the. The method of claim 7, wherein The end-of-learning condition checker between words and phrases repeats the word sorting, word reordering, phrase sorting, word and phrase band dictionary generation process of the autonomous learner until there are no more changes to the word and phrase band dictionary. Device for autonomous learning of translation relationships. The method of claim 1, Autonomously learning a translation relationship between words and phrases, further comprising a statistical machine translation model parameter extractor, extracting a parameter for a statistical-based translation model from the word alignment result and the phrase alignment result by the autonomous learner. Device for. In a method for autonomously learning a translation relationship between words and phrases in a statistical machine translation system, (a) receiving a form-parsed source language sentence and a form-parsed target language sentence; (b) converting the form-parsed source language sentence into a tokenized source language sentence to which the form-syntactic feature information is attached and the form-parsed target language sentence to which the form-syntax characteristic information is attached. Transforming it into a tokenized target language sentence, (c) performing word and phrase alignment on the tokenized source language sentence and the tokenized target language sentence pair; (d) generating a word and phrase band dictionary from the word alignment result and the phrase alignment result And autonomously learning translation relationships between words and phrases in a statistical machine translation system. The method of claim 10, The step (b) may include dividing each of the form-parsed source language sentences and the form-parsed target language sentences into morphemes or word units and attaching shape-syntactic feature information to each morpheme or word; , Tokenizing the morpheme or word to which the form-syntactic feature information is attached as a basic unit of translation; And autonomously learning translation relationships between words and phrases in a statistical machine translation system. The method of claim 11, In step (c), the words of the tokenized source language sentence are sorted by the words of the tokenized target language sentence and the words of the tokenized target language sentence are aligned with the words of the tokenized source language sentence. Obtaining a plurality of word alignment sets in sentence units according to the form-syntax feature information applied to the source language and the form-syntax feature information applied to the target language; Performs phrase alignment by taking the word alignment result that the plurality of word alignment sets have in common as an initial alignment value, and uses a band score for content words of an unaligned source and a target language syntax. Rearranging the words by calculating the and selecting the word with the highest band score, Using syntax information to sort one or more source phrases and one or more target phrases into band phrases And autonomously learning translation relationships between words and phrases in a statistical machine translation system. The method of claim 12, The step (d) includes the steps of presenting only word pairs of predetermined reliability or higher in the word and phrase band dictionary and removing band pairs below a critical reliability from the word and phrase band dictionary. And a method for autonomously learning a translation relationship between phrases. The method of claim 13, Autonomously learning translation relationships between words and phrases in a statistical machine translation system, further comprising repeating steps (c) and (d) until there is no further change in the word and phrase band dictionary. Way.

Images