KR100792204B1 - Apparatus for automatic translation customized for restrictive domain documents, and method thereof - Google Patents

Abstract

The present invention relates to a method and apparatus for extracting translation knowledge customized for a patent domain using a patent document as a target of translation and automatically translating the patent document using the extracted specialized knowledge. Extract terminology from the mass and assign a band word to the extracted terminology, extract expressions used in high frequency in patent documents, not documents in general domain, and build band expressions for the expressions Using the translated translation knowledge, and applying sentence pattern pattern, parallel structure recognition and sentence division by syntax cues to the excessively long sentences in the patent document, and extracting a parseable translation unit to perform structural analysis An apparatus and method for automatically translating a patent document are provided.

Machine Translation, Automatic Translation, Knowledge Extraction, Patent Translation

Description

Apparatus for automatic translation customized for restrictive domain documents, and method

1 is a view showing the overall configuration of an automatic translation apparatus specialized for patent documents according to the present invention

2 is a flowchart illustrating an automatic translation method specialized for patent documents according to the present invention.

3 is a flowchart illustrating a method of constructing a terminology among specialized automatic translation methods for a patent document according to the present invention.

4 is a flowchart showing a method for constructing and refining a band word among automatic translation methods specialized for patent documents according to the present invention.

5 is a flowchart illustrating a method for constructing and refining a bandword of a sentence or phrase pattern unique to a patent among an automatic translation method specialized for a patent document according to the present invention.

6 is a flowchart illustrating a method of generating a final parsing result among automatic translation methods specialized for patent documents according to the present invention.

7 is a flowchart illustrating a method of converting a structure analysis result of a native language into a structure of a target language among automatic translation methods specialized for patent documents according to the present invention.

* Explanation of symbols for main parts of the drawings

10: sentence division rule 20: stemming dictionary

30: Vocabulary context context probability information 40: Vocabulary probability information

50: target language patent corpus 60: conversion pattern

70: dictionary of single words and compound nouns 100: knowledge extraction unit

110: terminology construction unit 120: band word construction and refinement unit

130: sentence / phrase pattern construction unit 200: translation unit

210: preprocessing unit 220: morphological analysis and tagging unit

230: structure analysis unit 240: structure and vocabulary conversion unit

250 generator

TECHNICAL FIELD The present invention relates to an automatic translation system, and more particularly, to an automatic translation apparatus and method specialized for documents in a limited domain.

The machine translation system or automatic translation system has been a field of research that has been started with the invention of the computer. In the field of computer, its history is not short.

However, despite such a long history of development, considering the current market situation, it can be said that there is almost no automatic translation system that provides a level of translation quality satisfactory to the user in the general domain.

In fact, the reason for this is that the conventional automatic translation system has been the subject of a document having various words and various expressions such as a web document translation system with the development of the web environment. For this reason, the dictionary vocabulary or the rules or patterns for conversion, which are the basics of automatic translation, have been difficult to build completely due to the characteristics of the language.

For this reason, serious errors such as the problem of non-registered words, the problem of out of coverage of analysis rules, and the problem of no conversion information existed, and the quality of translation is not satisfactorily close to the level of commercialization. This problem eventually became a major obstacle to the commercialization of the automatic translation system.

Various problems arising from these unrestricted domains have naturally led to attempts to reduce the scope of automatic translation into restricted domains. In fact, the limited domains have been aimed at commercialization. It was a very realistic target.

Especially in the case of representative patent domains of limited domains, the volume of patent documents filed and registered worldwide is increasing rapidly every year, and in the global era, there is a great interest in not only domestic patents but also foreign patents. have. Currently, most patent translations are done by professional patent translation experts, and for each individual who does not belong to a company, the difficulty of searching and writing patents due to linguistic differences remains a big problem. In addition, in the case of companies, the cost (cost) and time required for patent translation, etc. are increasing day by day.

On the other hand, a description will be given of the problems appearing when translating a document of a limited domain such as patent translation using the existing general domain knowledge.

First, in general, the most important knowledge in automatic translation may include lexical dictionaries, analysis rules / patterns, transformation rules / patterns, and the like. If this existing knowledge is used to translate a document corresponding to a patent domain, the first problem that arises is the unregistered language problem. In other words, domains such as patents have extreme terminology used in various fields such as electricity, electronics, chemistry, physics, computers, etc. Also, general vocabularies used in patent documents are different from general domains in patent documents. It is often used in different meanings.

Secondly, in the case of patent documents, there are expressions that are used at extremely high frequencies only in the patent domain, and such expressions are rarely used in the general domain. Therefore, coverage problems arise with existing general syntax rules or patterns.

Thirdly, in automatic translation, the longer the sentence length, the more the structural ambiguity explodes, and the analysis time is significantly increased and the structure analysis performance is deteriorated. In the patent document, long sentences exceeding several hundred words frequently occur. do. Therefore, it is never easy to analyze and translate these long sentences without special processing.

Therefore, the present invention has been made to solve the above problems, in the translation of the document corresponding to the limited domain to extract the specialized knowledge of the domain and the automatic translation device using the extracted knowledge effectively performs the automatic translation And to provide a method.

Another object of the present invention is to provide an automatic translation apparatus and method for performing structural analysis by extracting a parseable analysis range for long sentences that cannot be handled by general analysis using specialized knowledge extracted in a corresponding domain.

Another object of the present invention is to construct a reasonable number of not lowering the quality of the translation results, for the high frequency expressions frequently used in the domain, such as analysis rules or translation patterns can be built in advance to produce a natural translation results An automatic translation apparatus and method are provided.

Looking at one aspect of the present invention for achieving the above object, by building a corpus according to the documents of the restricted domain through morphological analysis and tagging, the terminology construction unit for extracting and building the terminology, and by applying a weight The band word construction and refinement unit extracts the longest-priority high frequency expression to refine the sentence / phrase pattern, and constructs the band word for the constructed terminology, and the high frequency redundant lexical string and usage based on the corpus. And a knowledge extracting unit for extracting translation knowledge necessary for translation of sentences constituting a document of a restricted domain, including a sentence / phrase pattern building unit for constructing a phrase translation pattern and a sentence translation pattern of the corresponding domain according to the present invention. It includes a translation unit for generating a translation for the input sentence based on the dictionary and the conversion pattern applied It can provide automatic translation device.

Preferably, the knowledge extracting unit constructs a corpus according to a document of a restricted domain through morphological analysis and tagging, and extracts and constructs a terminology, and applies a weight to extract a high frequency expression of the longest priority method. A band word construction and refining unit for refining a sentence / phrase pattern and constructing a band word with respect to the constructed terminology, and a phrase translation pattern and sentence of a corresponding domain according to a high frequency redundant lexical string and usage based on the corpus And a sentence / phrase pattern construction unit for constructing a translation pattern.

Preferably, the weight is applied corresponding to the frequency of each vocabulary according to the restricted domain and the closeness to the air vocabulary.

Preferably, the translator separates the input sentences, and separates the vocabulary appearing in the separated sentences into tokens (tokens), and divides them into symbols, formulas, and words, and analyzes the morphemes of the tokens. Morphological analysis and tagging unit for statistical part-of-speech tagging using the lexical transformed HMM (Lexicalized Hidden Markov Model), and sentence division by sentence pattern and syntax pattern for the morphological analysis and tagged sentences for each node of the pattern A structure analysis unit for parsing and generating a final syntax analysis result, and a structure conversion for the generated syntax analysis result using a conversion pattern according to the extracted translation knowledge of the knowledge extraction unit to convert into a structure of a target language, Structure and vocabulary conversion unit for performing each individual vocabulary conversion using a dictionary, and output from the structure and vocabulary conversion unit It is characterized in that it comprises a generator which generates the final target language sentence using a transformed structure and vocabulary.

In order to achieve the above object, a feature of the automatic translation method specialized for a limited domain document according to the present invention is (a) specific corpus according to the restricted domain through morphological analysis and tagging process from documents written in the native language. Extracting the terminology by constructing the terminology, and (b) extracting the high frequency expression of the longest priority method by applying weights according to the restricted domain to refine the sentence / phrase pattern, and construct a band term for the terminology constructed. (C) constructing a phrase translation pattern and a sentence translation pattern based on the specific corpus constructed in step (a), and (d) sentence division by sentence pattern for morphological analysis and tagged sentences Generating a parsing result by parsing each node of the sentence pattern, and (e) translating the phrase and sentence constructed in step (c) Performing a structure conversion of each target language through a structure conversion on the generated parse result using each turn and each individual vocabulary conversion, and (f) generating a target language sentence through the converted structure and vocabulary. It is to include.

Preferably, the step (a) comprises: (a1) separating a large amount of document corpus constructed from documents written in a source language into each sentence, and separating the vocabulary of the separated sentence into tokens; (a2) analyzing morphemes and attaching all possible parts of speech to each token; and (a3) statistical part-of-speech tagging that assigns specific parts of speech to each word using pre-defined lexicon context probability information and lexical probability information. And performing a specific corpus to which a specific part-of-speech is assigned and extracting a terminology from the constructed specific corpus.

Preferably, the step (a1) is characterized in that it comprises the step of separating the long sentence into a few sentences by the long sentence separation rule when the input document is a long sentence.

Preferably, the token of step (a1) is characterized in that it is defined by any one of a symbol, a formula, a word.

Preferably, step (a4) is characterized by satisfying at least one of the following conditions.

Condition 1) unknown word: A word that is not in the regular domain dictionary.

Condition 2) The word w _i that satisfies <Condition 1>:

<Equation 1>

Where f (w _i ); The total frequency of the word w _i in the general domain,

f (w _i , t _ij ): how often the word w _i appears as part of speech t _ij in the general domain,

f '(w _i ): the total frequency of the word w _i in that domain,

f '(w _i , t _ij ): how often the word w _i appears as part of speech t _ij in that domain,

α: total threshold value (0.15 in the present invention)

β: maximum threshold value (in the present invention, used as 0.1).

Preferably, step (b) comprises: (b1) calculating a frequency of occurrence of each band word of each entry in advance from the constructed specific corpus; and (b2) co-occurring vocabularies with each band word. Extracting and calculating mutual information between each vocabulary, (b3) calculating a usage weight of each band word using values obtained through the vocabulary frequency and air vocabulary, and (b4) the band word Refining each bandword of a dictionary entry according to the importance of usage in a specific domain by applying a usage weight; and (b3) applying a bandword to the constructed terminology based on the refined sentence / phrase pattern and a defined weight. Characterized in that it comprises the step of building.

Preferably, the weight of step (b) is characterized in that the value corresponding to the frequency of each vocabulary according to the restricted domain and the closeness to the air vocabulary.

Preferably, the step (c) comprises: (c1) extracting a string having the highest frequency among the specific corpus constructed in the step (a), and generating a high frequency redundant vocabulary string and usage through the extracted string; (c2) determining a possibility of a syntax pattern candidate or sentence pattern candidate with respect to the generated high frequency redundant lexical string and usage; and (c3) if the determination result is a syntax pattern candidate, whether the phrase is a start / end node of a phrase, Or checking whether it is a part-of-speech node and constructing a specific phrase translation pattern; and (c4) if it is determined that the sentence pattern candidate is a result of the determination, constructing a specific sentence translation pattern for the entire sentence.

Preferably, the step (d) comprises: (d1) determining a pattern for the morphological analysis and tagged sentences; and (d2) if the determination result is a syntax pattern, recognizing a parallel structure and parsing a sentence through parallel node parsing. Performing syntax node parsing by performing partitioning; (d3) if it is determined that the sentence pattern is a sentence pattern, performing syntax node parsing on each node of the sentence pattern; and (d4) parsing the syntax node. It is characterized in that it comprises the step of generating a final structural analysis results by parsing the entire sentence by treating as a single chart.

Preferably, the parallel structure recognition includes generating a parallel structure candidate based on a syntax pattern when the size of the partial sentence to be parsed by the syntax node is greater than or equal to a specific length as a result of the determination, and parallel node recognition means for each candidate of the history structure. And selecting a parallel structure through syntax node constraints.

Preferably, the step (e) comprises: (e1) performing a structural conversion of the input source language document by using the syntax and sentence translation pattern constructed in the step (c) and converting it into a sentence structure of the target language; e2) converting at the lexical level by selecting an optimal band word for each source vocabulary according to the source language document using a single word and a compound noun dictionary, and (e3) based on the results of the above structure conversion and vocabulary conversion. And generating a conversion data structure.

Preferably, the structural transformation of the step (e1) is performed in a sentence unit, a clause unit, or a phrase unit.

Other objects, features and advantages of the invention will become apparent from the following detailed description of embodiments taken in conjunction with the accompanying drawings.

A preferred embodiment of an automatic translation apparatus and method specialized for a document in a limited domain according to the present invention will be described with reference to the accompanying drawings. Prior to the description, in the present specification, for convenience of description, the document of the limited domain is limited to the document of the patent domain and described as an embodiment. However, the document of the patent domain is only one preferred embodiment, and the limited domain is not limited thereto.

1 is a view showing the overall configuration of an automatic translation apparatus specialized for patent documents according to the present invention.

Referring to FIG. 1, the automatic translation apparatus extracts translation knowledge necessary for translation of sentences constituting a document of a corresponding domain based on a document of a limited domain, and the extracted translation knowledge. It is composed of a translation unit 200 for generating a translation for the sentence input by applying to the dictionary and the conversion pattern.

At this time, the knowledge extraction unit 100 is a terminology construction unit 110 for extracting and building a terminology by building a patent corpus according to the patent domain through a morphological analysis and tagging process by inputting documents written in the source language; By applying a weight corresponding to the frequency of each vocabulary according to the patent domain and the closeness to the air vocabulary, the high frequency expression of the longest priority method is extracted to refine the sentence / phrase pattern, and the bandword that constructs the bandword for the constructed terminology. And a sentence / phrase pattern construction unit 130 for constructing a phrase translation pattern and a sentence translation pattern of a corresponding domain according to a high frequency redundant vocabulary string and usage based on the patent corpus.

In addition, the translator 200 separates the inputted original text into sentences using sentence division rules, separates the vocabulary represented by the separated sentences into tokens, and then divides the tokens into symbols, equations, words, and the like. Morphological analysis using the preprocessing unit 210 and the preprocessed token to analyze the morphemes using a morphological analysis dictionary, and statistical part-of-speech tagging using the lexical transformed HMM (Lexicalized Hidden Markov Model) through the knowledge extraction unit. The tagging unit 220 and the structural analysis unit 230 for parsing each node of the sentence pattern by applying sentence division by sentence pattern and syntax pattern to the morphological analysis and tagged sentences to generate a final syntax analysis result. And, using the conversion pattern according to the sentence / phrase pattern constructed in the sentence / phrase pattern construction unit 130 of the knowledge extraction unit 100 to perform a structural transformation for the generated syntax analysis result After converting to the structure of the target language by using a dictionary, the structure and vocabulary conversion unit 240 for performing each individual vocabulary conversion using a dictionary, and the converted structure and vocabulary output from the structure and vocabulary conversion unit 240 It consists of a generator 250 for generating a final target language sentence through.

The automatic translation method for the patent document according to the present invention configured as described above will be described in detail with reference to the accompanying drawings.

2 is a flowchart illustrating an automatic translation method specialized for a patent document according to the present invention.

Referring to FIG. 2, first, a patent corpus is constructed through morphological analysis and tagging by inputting documents written in a native language as a first step, and a terminology is constructed from the constructed patent corpus (S100).

3 is a flowchart illustrating a method of constructing a terminology among specialized automatic translation methods for a patent document according to the present invention, and a method of constructing a terminology will be described in detail with reference to this.

First, a large amount of patent document corpus written in a native language is received as an input (S110), and each sentence is separated using a sentence division rule 10 of the input patent document corpus, and a vocabulary appearing in each sentence is token. Perform the pretreatment to separate into (S120).

The token is divided into symbols, formulas, words, and the like. In addition, the sentence dividing rule 10 applies a long sentence dividing rule for dividing a long sentence into several sentences when the input sentence is a long sentence.

That is, the long sentence separation rule is a regular expression that uses a vocabulary, a sentence symbol, a sentence symbol, and a sentence division symbol as a token. The sentence separation regular expression is composed of a <condition> and a <statement separator>. In this case, the <conditional part> is composed of an input vocabulary and a list of sentence / end symbols, and the <sentence separator> describes a sentence separation type including a sentence division symbol.

Therefore, in the sentence separation rule, when the input sentence satisfies the conditional sentence, the sentence is divided into the expression of the sentence separation unit.

For example, if <conditional> is "including:" <sentence separator> is "including as follow: \ n", and the input sentence encounters the word "ncluding:" instead of "including:" instead of "including:" Replace and separate sentences.

After such preprocessing, the morphological analysis dictionary 20 is searched to analyze the morphemes and attach all possible parts of speech to the respective tokens (S130). At this time, a word that does not appear in the morphological analysis dictionary is treated as an unknown word.

And a statistical part-of-speech for allocating an optimal part-of-speech to each word using the predefined lexical part-of-speech probability information 30 and the lexical probability information 40 to determine the parts of speech used correctly in the sentence among the possible parts of the word. Tagging is performed (S140). In this case, it is preferable to tag using the lexical converted HMM (Lexicalized Hidden Markov Model).

Through this serial operation, an automatically tagged patent corpus in which an optimal part of speech is assigned to each word of the input patent document corpus is constructed (S150).

After extracting the terminology from the constructed patent corpus (S160), the terminology thus extracted is constructed as a DB (S170). In this case, the terminology for the domain is very different from the parts of speech used in the domain even though words that do not appear in the general domain appear. Therefore, if any one of the following conditions is satisfied, the terminology of the domain is extracted.

At this time, the constructed automatic tagged patent corpus is used to find one satisfying the following conditions.

Condition 1) unknown word: A word that is not in the stemming dictionary.

Condition 2) The word w _i that satisfies Equation 1 below:

Where f (w _i ): the total frequency of the word w _{i in} the general domain,

f (w _i , t _ij ): how often the word w _i appears as part of speech t _ij in the general domain,

f '(w _i ): the total frequency of the word w _{i in} the automatically tagged patent corpus 209,

f '(w _i , t _ij ): frequency of word w _i appearing as part-of-speech t _ij in the automatically tagged patent corpus 209,

α: total threshold value (0.15 in the present invention)

β: maximum threshold value (0.1 in the present invention)

to be.

In Equation 1, f (w _i ) and f (w _i , t _ij ) values obtained in the general domain are already included in the lexical probability information 208 previously constructed from the tagged corpus of the general domain for statistical part-of-speech tagging. Use the value contained.

The second step is to extract the high frequency expression of the longest priority method by applying weights corresponding to the frequency of each vocabulary according to the patent domain and the closeness to the air vocabulary, to refine the sentence / phrase pattern, and to construct the terminology. Build a band word (S300).

That is, in the case of a dictionary or terminology dictionary used in the existing general domain, the use frequency of the band words may be changed in a specific domain such as a patent. In other words, when translating documents of a specific domain such as a patent using an existing general domain dictionary, because the use weights of the band words are different, the band words frequently used in the specific domain are not used, and thus, the structural transformation is Even if done properly, problems arise in conveying the overall meaning.

Therefore, a method of refining the bandword of each entry of the existing general domain dictionary based on the monolingual corpus written in the target language is introduced.

4 is a flowchart illustrating a method for constructing and refining a band word among automatic translation methods specialized for a patent document according to the present invention. Referring to this, a method for constructing and refining a band word for the constructed terminology will be described in detail.

First, each entry band word of the general domain dictionary is extracted (S210), and a frequency of occurrence of each band word is calculated for the extracted band word (S220).

In addition, the frequency of each vocabulary generated on the corpus 50 of the patent domain written in the target language is calculated (S220), and the relatedness is calculated based on mutual information of the vocabulary words and the vocabulary of each band. (S230).

Subsequently, the use weight of each band word is calculated using a weight function obtained in the calculation of the relationship between the frequency of each vocabulary and the air vocabulary (S250).

The sentence / phrase pattern is refined by determining the degree of importance of the corresponding bandword in a specific domain based on the weight function of the frequency of each vocabulary and the closeness of the air vocabulary (S260).

In this process, each of the band words of the existing general dictionary and the terminology dictionary constructed to fit the general domain builds the band word for the terminology constructed by the weight function newly defined to be specific to the new specific domain (S270). ).

In the next third step, a phrase translation pattern and a sentence translation pattern of the corresponding domain are constructed based on the high frequency redundant vocabulary strings and examples based on the patent corpus constructed in the first step (S100) (S300).

FIG. 5 is a flowchart illustrating a method for constructing and refining a word or phrase of a patent-specific sentence or phrase pattern among automatic translation methods specialized for patent documents according to the present invention, and constructs a phrase translation pattern and a sentence translation pattern of a patent domain with reference thereto How to do this in detail.

First, the boundary conditions are checked by inputting a large amount of automatically tagged patent corpus constructed in the first step to extract the most frequent lexical strings and usages (S320), and the high frequency overlaps through the extracted lexical strings and usages. A lexical string and an example are generated (S330).

Subsequently, a possibility of a syntax pattern candidate or a sentence pattern candidate is determined based on the generated high frequency redundant lexical string and usage (S340).

If the determination result is a syntax pattern candidate (S340), whether the words at both ends of the proposed syntax pattern candidate correspond to the parts of speech of the start / end node of the phrase of the structural analysis rule, or the words next to the syntax pattern candidate are between the old nodes in the rule. Checking whether the part-of-speech node corresponds to constructs a patent syntax pattern (S350). Then, using the constructed syntax phrase pattern for patent construct a phrase translation pattern DB (S360).

In addition, when it is determined that the sentence pattern candidate is determined as a result of the determination (S340), a sentence sentence pattern for the whole sentence is constructed (S370), and a sentence translation pattern DB for patent is constructed (S380).

The following shows an embodiment in which the patent syntax pattern S350 and the patent sentence pattern S370 are constructed according to an application in the patent document described above.

1) Example of constructing a phrase translation pattern for a patent (S350)

Extracted High Frequency Duplicate Vocabulary Strings and Usages:

in_accordance_with 20063 The present invention relates to a DC transformer / reactor in accordance with the introductory part of claim 1.

Constructed patent translation patterns: in accordance! with-> according to!

2) Example of constructing a sentence translation pattern for patent (S370)

Extracted High Frequency Duplicate Vocabulary Strings and Usages:

relates_to 20063 The present invention relates to a DC transformer / reactor in accordance with the introductory part of claim 1.

Constructed patent translation sentences: NP1 relate to NP2-> NP1: [] is about NP2: [!

In the fourth step, the sentence analysis by sentence pattern and syntax pattern is applied to the morphological analysis and tagged sentences, and the result is parsed for each node of the sentence pattern (S400).

FIG. 6 is a flowchart illustrating a method of generating a final parsing result of an automatic translation method specialized for a patent document according to the present invention. Referring to this, a method of generating a final parsing result is described in detail.

First, when the morphological analysis and tagged sentence is input (S410), the sentence pattern is recognized for the sentence (S420), and when the sentence pattern is applied, syntax node parsing is performed for each node of the sentence pattern (S430). .

In this case, the sentence pattern means a pattern in which the range of the pattern is the entire sentence, and is composed of vocabulary and syntax nodes. The phrase node mainly corresponds to a noun phrase (NP), a verb phrase (VP), a sentence (S), and the like. In addition, the sentence pattern recognition method uses the sentence pattern as a chart parser rule, and if the sentence is found, the word is searched from the current position for the tagging result, and the word following the syntax node in the sentence pattern is matched. Recognize ranges as syntax nodes until At this time, simply check the condition for the recognized syntax, and create the syntax node only if the condition is satisfied and add it to the inactive chart.

In this case, when the syntax node parsing unit is greater than or equal to a certain length in parsing the syntax node, the parallel structure is recognized (S440), and the sentence is divided into parallel nodes through sentence division based on a syntax pattern (S450).

In addition, the parallel node parsed as the parsing unit is attempted as the parsing unit (S460).

In this case, the parallel structure recognition first recognizes a range of possible parallel structures using syntax information. That is, in English, the parallel syntax takes the form X-> X, X, .., and X, so all of these possible ranges are recognized as candidates for parallelism. The parallel node is recognized using the tagging result for the recognized parallel structure. Parallel nodes are divided into NP (noun phrase), VP (verb phrase), VPG (verb noun), S (statement), and SG (independent injection phrase). The distinction is not to parse, but the following heuristics are used.

1) If there is more than one main verb and there are nouns / pronouns / investigators in front of the main verb, S, VP

2) NP / SG or VPG if there is a participle of the verb and a noun / pronoun / investigator before the part

3) or NP

And two constraints are checked for the above syntax recognition result.

1) Syntax nodes of all parallel nodes must be identical.

2) NP / VPG is only possible if a verb or preposition precedes the parallel structure.

The longest parallel structure is selected among the parallel structures that satisfy the above conditions.

Even when such a parallel node parsing is attempted, if the parsing unit is greater than or equal to a certain length, sentence division is performed according to a predetermined syntax pattern, and partition sentence parsing is performed on the result of the division. In this case, sentence division is performed unconditionally by commas.

And finally, treating the partial parsed results so far as one chart to parse the entire sentence again to generate the final structural analysis results (S470).

If the fourth step (S400), the method for generating the final syntax analysis results are shown in the examples through the English patent example as follows.

Example

[Input Sentence]:? Construction of fixing a flexible sheet for use in an electronic device comprising a case being formed with a plurality of through holes, a chassis being accommodated in an interior of the case, a flexible sheet being disposed on a surface of the chassis and having a plurality of flexible switches arranged thereon, a circuit board being provided below the chassis and having a connector fixed thereon, and a multiple of manual buttons being provided above each flexible switch and being exposed from the through holes of the case to the outside of the case, the construction of fixing the flexible sheet wherein the flexible sheet comprises a flat plate portion being in close contact with the chassis and a flat cable portion which projects on an edge of the flat plate portion and with which a connecting terminal portion is provided on its end, the flat cable portion is folded back to the chassis to have the connecting terminal portion connected to the connector, and the cha ssis is provided with a lift-up prevention piece to prevent a part of the flat plate portion of the flexible sheet from being lifted up from a surface of the chassis. "

[Pattern]: S-> S: [vg], NP where S, S

[Pattern Recognition Result]:

(S: [vg] Construction of fixing a flexible sheet for use in an electronic device comprising a case being formed with a plurality of through holes, a chassis being accommodated in an interior of the case, a flexible sheet being disposed on a surface of the chassis and having a plurality of flexible switches arranged thereon, a circuit board being provided below the chassis and having a connector fixed thereon, and a multiple of manual buttons being provided above each flexible switch and being exposed from the through holes of the case to the outside of the case), (NP the construction of fixing the flexible sheet) wherein (S the flexible sheet comprises a flat plate portion being in close contact with the chassis and a flat cable portion which projects on an edge of the flat plate portion and with which a connecting terminal portion is provided on its end), (S the flat cable portion is folded back to the chassis to have the connecting terminal portion connected to the connector, and the chassis is provided with a lift-up prevention piece to prevent a part of the flat plate portion of the flexible sheet from being lifted up from a surface of the chassis.)

[Perform parsing of each recognized syntax node]

Parse S [vg], NP, S

[Parallel Structure Recognition]

(S: [vg] Construction of fixing a flexible sheet for use in an electronic device comprising (NP a case being formed with a plurality of through holes), (NP a chassis being accommodated in an interior of the case), (NP a flexible sheet being disposed on a surface of the chassis and having a multiple of flexible switches arranged thereon, (NP a circuit board being provided below the chassis and having a connector fixed thereon), and a multiple of manual buttons being provided above each flexible switch and being exposed from the through holes of the case to the outside of the case)

The last node above is not bound to a syntax node. This is because we do not know the end range of the last node.

Parsing of each parallel node

Parse tree for each recognized NP and construct tree for parallel structure

Above, if the sentence pattern does not exist, first perform parallel structure recognition, parse each syntax node, and then parse the whole sentence again. When the sentence length exceeds a certain length, The division is made by.

(Construction of fixing a flexible sheet for use in an electronic device comprising (NP a case being formed with a plurality of through holes), (NP a chassis being accommodated in an interior of the case), (NP a flexible sheet being disposed on a surface of the chassis and having a plurality of flexible switches arranged thereon), (NP a circuit board being provided below the chassis and having a connector fixed thereon), and a multiple of manual buttons being provided above each flexible switch and being exposed from the through holes of the case to the outside of the case), (the construction of fixing the flexible sheet wherein the flexible sheet comprises a flat plate portion being in close contact with the chassis and a flat cable portion which projects on an edge of the flat plate portion and with which a connecting terminal portion is provided on its end), (the flat cable portion is folded back to the chassis to have the connecting terminal portion connected to the connector, and the chassis is provided with a lift-up prevention piece to prevent a part of the flat plate portion of the flexible sheet from being lifted up from a surface of the chassis.)

In the fifth step, a structure conversion is performed on the generated parsing result using the conversion pattern 605 according to the constructed sentence / phrase pattern, and then converted into a structure of a target language. Each individual vocabulary conversion is performed using (S600).

7 is a flowchart illustrating a method of converting a structure analysis result of a native language into a structure of a target language among automatic translation methods specialized for patent documents according to the present invention.

First, a structural transformation of a structural analysis result of an input source language sentence is performed using the transformation pattern 60 constructed in the third step S300 (S510). In this case, the structural transformation is performed in units of sentences, clauses, and phrases. The transformation pattern 60 is used to select a transformation pattern that is most optimally matched with the structural analysis result of the fourth step S400.

When the conversion to the target language sentence structure is completed through this, the conversion of each individual vocabulary is performed using the single word and compound noun dictionary 70 (S520). In this case, a bandword selection ambiguity, that is, in the case of a primitive vocabulary having an intermediate meaning, is performed in the lexical conversion process to perform the bandword selection function for eliminating the ambiguity and selecting an optimal bandword. In the case of a patent document, due to the feature that includes a lot of terminology, the complex band word selection as in the general domain is not implemented, and the band word selection technique is mainly applied according to the frequency of use.

When the structure conversion and the lexical conversion are completed as described above, a conversion data structure for generating the corresponding results is generated (S530).

Then, as a sixth step, which is the last step, the final target language sentence is generated through the converted structure and vocabulary that is output after the structure and the vocabulary are converted (S600).

As described above, the optimum embodiment of the present invention has been disclosed through the detailed description and the drawings. The terms are used only for the purpose of describing the present invention and are not used to limit the scope of the present invention as defined in the meaning or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

As described above, the automatic translation apparatus and method specialized for the document of the limited domain according to the present invention have the following effects.

First, as a narrower specific translation domain, the patent domain is set up to build a specialized translation knowledge for the patent domain, and the specialized translation knowledge thus constructed is used, and the long document is divided into long sentences of the patent document to automatically generate the patent document. By translating, the quality of automatic translation can be improved to the extent that it can directly help the actual patent field.

Secondly, with the development of wired / wireless communication technology, it is possible to refer to patents of other countries more easily and economically at the time of sharing of patents around the world, thereby lowering the possibility of future patent disputes. The economic costs that have been spent on searching / referencing foreign patents can also be significantly reduced.

Claims (16)

A terminology construction unit that extracts and constructs terminology by establishing a corpus according to a document of a limited domain through morphological analysis and tagging; A band word construction and refining unit for extracting a high frequency expression of a longest priority method by applying weights to refine a sentence / phrase pattern, and constructing a band word for the constructed terminology; Translation knowledge required for translation of sentences constituting a document of a limited domain, including a sentence / phrase pattern construction unit for constructing a phrase translation pattern and a sentence translation pattern of the domain according to the high frequency redundant lexical string and usage based on the corpus; A knowledge extraction unit for extracting And a translation unit configured to generate a translation for an input sentence based on the dictionary and the conversion pattern to which the extracted translation knowledge is applied. delete The method of claim 1, And the weight is applied corresponding to the frequency of each vocabulary according to the limited domain and the closeness to the air vocabulary. The method of claim 1, wherein the translation unit A pre-processing unit that separates the input sentences, separates the vocabulary appearing in the separated sentences into tokens, and separates them into symbols, formulas, and words; A morpheme analysis and tagging unit for analyzing the morphemes of the token and tagging the statistical parts of speech using a lexical transformed HMM (Lexicalized Hidden Markov Model) through the knowledge extraction unit; A structural analysis unit for parsing each node of the pattern by applying sentence division by sentence pattern and syntax pattern to the morphological analysis and tagged sentences, and generating a final syntax analysis result; A structure and vocabulary that converts the generated syntax analysis result into a structure of a target language by using a transformation pattern according to the extracted translation knowledge of the knowledge extractor, and converts each vocabulary by using a dictionary. With a conversion unit, And a generator for generating a final target language sentence through the converted structure and vocabulary output from the structure and the vocabulary converter. (a) extracting a terminology by constructing a specific corpus according to a restricted domain through morphological analysis and tagging from documents written in a source language; (b) extracting a high frequency representation of the longest priority method by applying weights according to restricted domains, refining sentence / phrase patterns, and constructing a band word for the constructed terminology; (c) constructing a phrase translation pattern and a sentence translation pattern based on the specific corpus constructed in step (a); (d) applying sentence segmentation by sentence pattern to the morphological analysis and tagged sentences, parsing each node of the sentence pattern pattern to generate a parsing result; (e) performing a structure conversion of the target language and each individual vocabulary conversion through structure conversion on the generated parse result using the phrase and sentence translation pattern constructed in step (c); (f) generating a target language sentence through the converted structure and vocabulary. The method of claim 5, wherein step (a) (a1) separating a large amount of document corpus constructed from documents written in the source language into each sentence, and separating the vocabulary of the separated sentence into tokens; (a2) analyzing morphemes and attaching all possible parts of speech to each of the tokens, (a3) constructing a specific corpus to which a specific part of speech is assigned by performing statistical part-of-speech tagging that assigns a specific part-of-speech to each word using the predefined lexical part-of-speech probability information and the lexical probability information; (a4) Automatic translation method comprising the step of extracting the terminology from the constructed specific corpus. The method of claim 6, Wherein (a1) is the automatic translation method comprising the step of separating the long sentence into a few sentences by the long sentence separation rule, if the input document is a long sentence. The method of claim 6, The token of step (a1) is an automatic translation method characterized in that it is defined by any one of a symbol, a formula, a word. 7. The automatic translation method of claim 6, wherein step (a4) satisfies at least one of the following conditions. Condition 1) unknown word: A word that is not in the regular domain dictionary. Condition 2) The word w _i that satisfies <Condition 1>: <Equation 1>

Where f (w _i ); The total frequency of the word w _i in the general domain, f (w _i , t _ij ): how often the word w _i appears as part of speech t _ij in the general domain, f '(w _i ): the total frequency of the word w _i in that domain, f '(w _i , t _ij ): how often the word w _i appears as part of speech t _ij in that domain, α: total threshold value (0.15 in the present invention) β: maximum threshold value (in the present invention, used as 0.1). The method of claim 5, wherein step (b) (b1) calculating frequency of occurrence of each band word of each entry in advance from the constructed specific corpus; (b2) extracting co-occurring vocabulary with each band word and calculating mutual information between each vocabulary; (b3) calculating a use weight of each band word using the values obtained through the lexical frequency and the air vocabulary, (b4) applying each of the bandword usage weights to refine each bandword of a dictionary entry according to the importance of usage in a specific domain; (b3) an automatic translation method comprising constructing a band word for the constructed terminology based on the refined sentence / phrase pattern and a defined weight. The method of claim 5, The weighting of step (b) is a value corresponding to the vocabulary for each vocabulary according to the restricted domain and the closeness to the air vocabulary. The method of claim 5, wherein step (c) (c1) extracting a string having the highest frequency among the specific corpus constructed in the step (a), and generating a high frequency redundant lexical string and usage through the extracted string; (c2) determining a possibility of a syntax pattern candidate or a sentence pattern candidate for the generated high frequency redundant lexical string and usage; (c3) if it is determined that the candidate is a phrase pattern candidate, constructing a specific phrase translation pattern by checking whether the phrase is a start / end node or a part-of-speech node; (c4) if it is determined that the sentence pattern candidate is a result, the automatic translation method comprising the step of constructing a specific sentence translation pattern for the entire sentence. The method of claim 5, wherein step (d) (d1) determining a pattern with respect to the morphological analysis and the tagged sentence, (d2) if it is determined that the syntax pattern is a result, recognizing the parallel structure and performing syntax parsing by parsing the sentences through parallel node parsing; (d3) if it is determined that the sentence pattern is determined, parsing a node of each sentence of the sentence pattern; and (d4) generating the final structural analysis result by re-parsing the entire sentence by treating the parsed node parsed result as one chart. The method of claim 13, wherein the parallel structure recognition Generating a parallel structure candidate based on a syntax pattern when the size of the partial sentence to be parsed as the result of the determination is greater than or equal to a certain length; Selecting a parallel structure through parallel node recognition means and syntax node constraints for each candidate of the history structure. The method of claim 5, wherein step (e) (e1) performing a structural conversion of the input source language document by using the syntax and sentence translation pattern constructed in step (c) and converting it into a sentence structure of the target language; (e2) performing conversion at the lexical level by selecting an optimal band word for each source vocabulary according to the source language document using a single word and a compound noun dictionary; (e3) generating a transformation data structure based on the structure transformation and lexical transformation results. The method of claim 15, The structural transformation of the step (e1) is automatic translation method, characterized in that performed in units of sentences, clauses, phrases.

Images