KR20000056245A - Translation example selection method using similarity reflecting discriminitive value in example-based machine translation - Google Patents

Abstract

PURPOSE: A translation example selection method is provided to select an example sentence most similar to an input sentence in an EBMT(example based machine translation) by using a similarity reflected by distinction included in the words. CONSTITUTION: A translation example selection method comprises steps of converting words with a word selection ambiguity, included in an input sentence, into a translation example expression(110), selecting a translation example sentence in relation with the words having the ambiguity from a translation example sentence database(120), calculating a similarity with the input sentence by applying the similarity calculation method to the selected example sentences(130), and selecting a target word among the most similar example sentence as a translation word(140). The translation example expression step(110) includes describing a tree structure reflecting a sentence structure as well as a word semantics expressed in the word and the context.

Description

Translation example selection method using similarity reflecting discriminitive value in example-based machine translation}

The present invention relates to English-Korean translation, and more particularly, to a method of selecting a translation example using similarity reflecting the discriminating power of a word in order to select a translation example most similar to an input sentence in example-based machine translation.

Example-Based Machine Translation (EBMT) is based on the idea of performing translations by mimicking similar translation examples. The main element of such a translation system is to build a large database of translation sentences and search the database for translation sentences that are most similar to the input sentences to be translated.

The existing research mainly evaluates the similarity by calculating the semantic distance between the words in the translated sentences and the words in the input sentences. The English word CASE can be translated into 'case', 'box', 'patient', and the like.

It was a case of bad judgment.

[Box] The porter will carry your cases up to your room.

[Patient] He is a new case of influenza.

If the input sentence is "He is a new case of cholera." It can be seen that it must be translated. However, the words that appear in the context of the word in the translation example may have different degrees of contribution to determining the band words. In the translation example for 'patient', 'HE' may also appear in translation examples for other target language words, so the contribution to selecting the target language word is very small. Therefore, it is difficult to select an appropriate translation example simply by calculating the semantic distance of words appearing in the context.

The following is a brief representation of the words and their meaning codes that appear in the context of the translation example for 'AREA'.

[Region] residential: Dd65, city: Ce80, England: Ce78, no-smoking: Ed88

Education: Gb35, mathematics: Jb30, economy: Jf141

[Area] circle: Jb44, land: Ce78, square-mile: Jc66, square-kilometer: Jc66

As can be seen in the above example, some words appearing in the translation example have a weak influence in determining the band word. Words with the meaning of Ce80 (city, land) can be used in both cases of the target language words 'region' and 'area', and words with the meaning of Jb30 (mathematics, circle) are the target language words 'field' and 'area' Can be translated as' Therefore, the words in this meaning can be said to have a weak discernment in selecting the target language word of the word 'AREA'.

On the other hand, the word meaning Gb35 (education) does not appear in the translation example of the target language other than 'field', so it is likely that the word should be translated as 'field'. As such, some words appear biased in the translation example for a particular target language word, providing a crucial clue in selecting the target language word. That is, words with high discernment. Therefore, in selecting the translation example to be used for translation, it is effective not only to calculate the semantic similarity between the input sentence and the translation example, but also to reflect the discernment of the meaning of the word appearing in the translation example.

Conventional methods generally evaluate similarity by calculating the semantic distance between words in the context of input sentences and words in translation sentences. And the words that appear in the context of the word to be translated from the input sentence _{I 1, I 2, ...,} I m, the words that appear in the translated sentence Speaking _{E ij1, E ij2, ...,} E ijm input sentence and Similarity with the translation example can be calculated as the sum of the respective semantic distances as shown in Equation (1).

Input: I = (I ₁ , I ₂ , ..., I _m )

_{Example: E ij = (E ij1} , E ij2, ..., E ijm)

_{Dist (I, E ij) =} d ((I 1, I 2, ..., I m) (E ij1, E ij2, ..., E ijm))

Therefore, the existing method focused on _finding the semantic distance between words, that is, d (I _k , E _ijk ) function. However, as described above, even if the same meaning distance is different, the discriminating power contributing to selecting the target language word is different, but the closest but not suitable band word may be selected.

Sumita and Lida used weights to determine how words in a context affect the determination of the target language words ("Experiments and Prospects of Example-based Machine Translation"; ACL91). However, by applying a fixed weight to a word appearing in the context, it does not reflect the discernment that the word has for each target language word candidate. In other words, individual distributions appearing in the context of the target language word candidates cannot be modeled.

The technical problem to be achieved by the present invention is to compare the similarity of the meaning of the word that appears in the context with weights in the semantic distance calculation method in order to select a translation example most similar to the input sentence among the translation examples that are built into the database This paper provides a method for selecting a translation example using similarity reflecting discernment in example-based machine translation that expresses translation example by calculating.

FIG. 1 is a flowchart illustrating a method for selecting a translation example using similarities in which discrimination is reflected in example-based machine translation according to the present invention.

2 is a flow chart illustrating a similarity calculation method used in the present invention.

In the example-based machine translation according to the present invention for solving the above technical problem, the method of selecting a translation example using the similarity reflecting the sensitivities includes (a) (a) translating a word having ambiguity in lexical selection among input sentences reflecting a syntax relationship An input sentence conversion step of converting an example sentence expression method; (b) selecting translation examples related to the input sentence converted from step (a) in the translation example database constructed by the translation example expression method; (c) calculating a similarity with the input sentence for each of the translation sentences selected from the step (b); And (d) selecting a translation example sentence having the largest similarity calculated from step (c) and selecting the target word TARGET_WORD of the selected translation example sentence as a target language word corresponding to a source language word. It is done.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a method for selecting a translation example using similarity reflecting discernment in example-based machine translation according to the present invention, the input sentence conversion step 110, the translation example selection step 120, and the similarity calculation step 130 ) And the target language selection step (140).

The input sentence conversion step 110 is a step of converting a word having lexical selection ambiguity among the input sentences into a translation example expression.

The translation example selection step 120 is a step of selecting translation example sentences related to a word having ambiguity in an input sentence from a previously constructed translation example database.

The similarity calculation step 130 is a step of calculating the similarity with the input sentence using a similarity calculation method for each of the translation examples selected from the translation example selection step 120.

The target language selection step 140 selects the most similar translation example sentence and selects TARGET_WORD of the selected translation example sentence as the target language word corresponding to the source language word.

The present invention will be described in detail based on the above steps.

In addition to vocabulary and semantic information, the translation example should include various knowledge to reflect the form-syntax information and tree structure. This is to exclude the selection of the target language words that are not form-syntactically suitable by describing a tree structure that reflects form-syntax information and syntax relationships.

The types of knowledge included in the translation example are as follows.

a. Tree structure (Parent, Child, Sibling): Describes the position of the sentence structure tree in the tree.

b. Constraints: Describes morphological constraints, such as article, number information, and additiveity.

c. Vocabulary: Describe the word vocabulary itself that appears in the context

d. Semantic information: describes the semantic code of the word vocabulary that appears in the context

The translation example expression method (translation example expression reflecting syntax relation) is expressed in the following form (step 110).

% SOURCE_WORD 〈Source_Language_Word〉

% TARGET_WORD 〈Target_Language_Word〉

% PARENT 〈Lexical_Item: Semantic_Code: Constraints〉

% CHILD 〈Lexical_Item: Semantic_Code: Constraints〉

% SIBLING 〈Lexical_Item: Semantic_Code: Constraints〉

The translation example "He is a new case of influenza" for "CASE" is described in the following form. At this time, 'PREP: OF' described in '[]' means form-syntax information that 'case' and 'influenza' are connected with the preposition 'of', and N1, B140, ... are the vocabulary 'be' respectively. ', represents the semantic code of' influenza '.

% SOURCE_WORD 〈CASE〉

% TARGET_WORD 〈Patient〉

% PARENT 〈BE: N1〉

% CHILD 〈INFLUENZA: B140 [PREP: OF], NEW: L241〉

% SIBLING 〈HE: G280〉

In operation 120, translation sentences related to the input sentence are selected from the translation example database constructed by the above-described translation example expression method.

The similarity with the input sentence is calculated using the similarity calculation method for each of the selected translation sentences from the translation example selection step 120 (step 130).

Similarity is calculated based on three pieces of information: semantic distance, constraint, and discernment. Similarity calculation method for this is as shown in Equation 2.

Input: I = (I ₁ , I ₂ , ..., I _m )

_{Example: E ij = (E ij1} , E ij2, ..., E ijm)

_{Dist (I, E ij) =} d ((I 1, I 2, ..., I m) (E ij1, E ij2, ..., E ijm))

Semantic distances are calculated according to the degree of separation of the meaning of two words on a hierarchical thesaurus. A general method of calculating the mean distance is expressed as in Equation 3.

d ((I _k , E _ijk ) = 0 if (L exical _ I tem (I _k ) == L exical _ I tem (E _ijk ))

= 0.125 if (Semantic_Code (I _k ) and Semantic_Code (E _ijk )

Present at the same level of thesaurus)

= 0.25 if (Semantic_Code (I _k ) and Semantic_Code (E _ijk )

1 level difference in the thesaurus)

= 0.5 if (Semantic_Code (I _k ) and Semantic_Code (E _ijk )

2 levels difference in the thesaurus)

= 1.0 others

The closer you are to the hierarchy, the smaller the meaning of the two words. In other words, the small meaning of the two words means that the similarity of the two words is high.

Constraints are for form-syntax information and are expressed as in Equation 4.

Is a _{Penalty (I k, E ijk)} = 1 if Constraint (E ijk) I k satisfies

= 2 otherwise

If the corresponding word in the input sentence does not satisfy the form-phrase condition described in the translation example, it has the effect of deducting the similarity calculation by multiplying the above-described semantic distance by two. Multiplying by two lowers the matching level by one in the semantic distance calculation.

Discrimination is high when it appears biased in translation examples for a particular target language. Equation 5 is an equation for calculating the discrimination power for selecting the target language word. This value is calculated to have a lower value for higher discernment. Discernment is an item that is multiplied by the distance calculation value, which means that the closer the distance is, the higher the similarity is.

W _k = disc_value (semantic (E _ijk ))

Here, freq (SM) in E _ijk for all i, j means all frequencies in which a specific semantic code (SM) appears in a translation example for an ambiguous source language word, and freq (SM) in E _ijk for all j Denotes the frequency of a specific semantic code (SM) appearing in the translation example for the specific target language word i among the translation examples for the source language word. Therefore, the greater the specific gravity of freq (SM) in E _ijk for all j value, the higher the discernment.

2 is a flowchart illustrating a similarity calculation process between the above-described input sentence and the translated example sentence.

First, a maximum value "0" is assigned to the variable Distance, and a value "0" is input to a variable K representing the number of translation examples to be processed (step 210). Check whether there is a translation example to be processed, and if there is no more translation example to process (step 220). The semantic distance between the word of the translation example sentence and the word of the input sentence is calculated (step 230). In operation 240, the penalty is calculated by checking whether a word in the input sentence satisfies a constraint (step 240). The classification power of the word meaning of the translation example sentence is calculated, and the similarity is calculated by multiplying the value calculated in step 240 with the classification power (step 250). The similarity calculated in step 250 is summed with the variable distance (step 260), and the similarity between the input sentences and the translated sentences is calculated by repeating steps 220 to 260 for the translated example to be processed (step 130).

From the above similarity calculation step 130, a translation example sentence having the largest similarity is selected, and the target word TARGET_WORD of the selected translation example sentence is selected as the target language word corresponding to the source language word (step 140).

word Test sentences Model A Model B noun 42 27 (64.3%) 27 (64.3%) verb 46 33 (71.7%) 30 (65.2%) adjective 15 11 (73.3%) 9 (60.0%) adverb 6 4 (66.6%) 2 (33.3%) Sum 109 75 (68.8%) 68 (62.4%)

Table 1 compares and evaluates the model (A) and the model (B) using discriminant power by constructing about 720 translation examples for 33 words according to an embodiment of the present invention. The test sentences were used by randomly extracting 109 sentences. As a result, the performance improved from 62.4% to 68.8%. These results show that discernment can be useful in vocabulary selection.

According to the present invention, the possibility of selecting a proper translation example compared to the conventional method is improved by reflecting the discrimination power of the word appearing in the translation example as the weight in the similarity calculation. In other words, the accuracy of selecting the target language word suitable for the source language word in machine translation increases.

In addition to the meaning of the words appearing in the translation example, it is easier to build the translation example by reflecting the discernment.

Claims (6)

(a) an input sentence conversion step of converting a word having ambiguity of lexical selection among input sentences into a translation example expression method reflecting a syntax relationship; (b) selecting translation examples related to the input sentence converted from the step (a) in the translation example database constructed by the translation example expression method; (c) calculating a similarity with the input sentence for each of the translation sentences selected from the step (b); And (d) selecting a translation example sentence having the largest similarity calculated from step (c) and selecting the target word TARGET_WORD of the selected translation example sentence as a target language word corresponding to a source language word. Translation Example Selection Method using Similarity Reflecting Discrimination in Example-based Machine Translation. The method of claim 1, wherein the translation example expression method in step (a) is A method of selecting translation examples using similarity reflecting discrimination in example-based machine translation, which describes a tree structure that reflects form-syntax information and syntax relationship, in addition to word vocabulary in context and semantic code of word in context. The method of claim 1, wherein step (c) (c1) calculating a meaning distance between the word of the translation example sentence and the word of the input sentence; (c2) calculating a penalty by checking whether the input sentence word satisfies a constraint; (c3) calculating the discernment of the word meaning of the translation example; And and (c4) multiplying the meaning distance by the penalty and the discriminant power. The method of claim 3, wherein step (c1) A method of selecting a translation example using similarity reflecting discernment in example-based machine translation, characterized in that the word meaning of the translation example and the word meaning of the input sentence are calculated according to the distance from the hierarchical thesaurus. The method of claim 3, wherein step (c2) If the word corresponding to the input sentence word satisfies the form-phrase condition described in the translation example, and if it is not satisfied, it calculates by deducting the similarity by lowering the matching level by one step in the semantic distance calculation. Translation Example Selection Method using Similarity Reflecting Discrimination in Example-based Machine Translation. The method of claim 3, wherein step (c3) In the translation example of the ambiguity of a primitive word, A method for selecting translation examples using similarity reflecting discrimination in example-based machine translation, wherein all frequencies in which a specific semantic code appears are divided by the frequency of a specific semantic code for a specific target language word.