KR20100072384A - Method for generating korean connectives in chinese-korean machine translation and its apparatus - Google Patents

Abstract

PURPOSE: A method for generating Korean connectives for Chinese-Korean machine translation and a device thereof are provide to generate clear vision of connectives in Korean point of view and generate Korean sentence, thereby generating high quality of Korean. CONSTITUTION: A Chinese morpheme analyzer(101) selects optimum morpheme part about Chinese word in a Chinese input sentence. A Chinese construction analyzer generates a Chinese construction tree. A Chinese-Korean converter(103) converts the Chinese construction tree into Korean construction tree. If an inter-short sentence logical connection mark is inexplicit, a connective ending determining unit(104) generates inter-short connection ending through connective knowledge DB(106). A Korean generator(105) generates natural Korean.

Description

TECHNICAL FOR GENERATING KOREAN CONNECTIVES IN CHINESE-KOREAN MACHINE TRANSLATION AND ITS APPARATUS

The present invention relates to a method and apparatus for generating Korean connected endings for automatic Chinese translation. In more detail, when a Chinese input sentence is a complex sentence, the present invention does not express a logical connection between a short sentence and a short sentence. The present invention relates to a method and apparatus for enabling a connection ending generation for a connection to solve an unnatural problem.

The present invention is derived from the research conducted as part of the IT growth engine technology development project of the Ministry of Knowledge Economy and the Ministry of Information and Communication Research and Development. [Task management number: 2006-S-037-03, Task name: Development of application-specific Korean-Chinese automatic translation technology] ].

As is well known, in the case of a compound sentence consisting of two or more short sentences in Chinese, the logical connection between the sentences forming the compound sentence is sometimes expressly expressed as an conjunction, but in more compound sentences in general, There is a characteristic that the connection is not represented at all.

On the basis of this feature, in the case of Chinese-Chinese automatic translation method, if there is a Chinese vocabulary expressing a logical connection between short sentences, the logical connection between the short sentences can be expressed by translating the vocabulary into Korean. In the case of non-explicit logical connections that do not explicitly express the logical connection between them, a suitable Korean connection ending must be created. In addition, logical connection between short sentences is made by using “~ go”, which is the most commonly used connection meaning because it cannot grasp the characteristics of logical connection between short sentences.

In other words, Korean connection ending “~ Go” can express various logical connections, and Chinese compound sentences which have non-explicit logical connection mostly have parallel or sequential logical connections that can be represented by “~ high”. In the automatic translation method, we did not grasp the logical connection between short sentences, but used the link between short sentences using “~ high” in a batch.

However, since Chinese did not analyze the logical connection between short sentences, it is obvious that when Korean short texts are created with only "~" in Korean translation, it is not natural in Korean and logic that "~" cannot express. You can see the case of connection. In the following examples 1, 2, and 3, when the connection ending “~ high” is used, the logical connection of the front and back sentences is abnormal. In other words, in the case of sentences 1,2 and 3, it is clear that the semantic logical connection is obvious from the Korean point of view.

(Example 1)

(Translation 1) (1-1) Good European stock expressions on that day, (1-2) All major stocks closed high.

(Example 2) (2-1)

(2-1) The stock has recently fallen significantly, and (2-2) may rebound quickly in the short term.

(Example 3)

(3-1) Affected by rising stocks at the same time, (3-2) Funds no longer flow into US Treasury bonds.

As described above, in order to translate the Chinese compound sentence into Korean in the Chinese-Chinese automatic translation device, it can be seen that the generation of natural connection endings in accordance with the exact meaning of the logical connection between the short sentences and the natural meaning of the Korean language has a significant effect on the overall meaning transmission. . In order to generate Chinese compound sentences with non-explicit logical connections into Korean compound sentences, a Chinese automatic translation apparatus must know exactly the meanings of each short sentence and figure out the logical state of the connections between the meanings.

However, in the prior art as described above, it is very difficult to grasp the meaning represented by the sentence and the logical connection between the meanings. To solve this problem, very complicated language knowledge and an engine are required. There is a problem that I did not try.

Accordingly, the technical problem of the present invention was devised to solve the above-mentioned problems, and in translation of a Chinese sentence having a logical connection between short sentences not expressly expressed in Chinese in Chinese automatic translation into Korean, the original Chinese text to be translated The present invention provides a method and apparatus for generating a Korean linking ending for automatic Korean-to-Korean translation, which can generate an explicit linking ending from a Korean point of view, in the case of having an explicit Chinese short logical linking.

In accordance with an aspect of the present invention, a method of generating Korean connected endings for automatic Korean-Chinese translation comprises: separating Chinese input sentences into Chinese words, selecting morpheme parts of speech for the separated words, and a Chinese syntax tree based on morpheme parts of speech. Generating a step, converting a Chinese syntax tree into a Korean syntax tree, converting a word in the Chinese syntax tree into a Korean band word, and connecting the extracted extracted from the Korean native corpus when the Korean syntax tree is a plurality of short syntax trees. Generating the linking ending between each short sentence using the mother knowledge DB, rearranging the Korean words according to the Korean word order using the Korean syntax tree in which the linking end is generated, and generating functional words according to the aspects of Korean grammar and verbs. To generate Korean language, and to add knowledge of connected endings from Korean native corpus. It is characterized in that it comprises a step of building in the connection mother knowledge DB.

In addition, the Korean connected ending generation apparatus for automatic Korean-Chinese translation according to another aspect of the present invention, the Chinese morpheme analysis unit for separating the Chinese input sentence into Chinese words, and analyzes the morpheme to select the morpheme parts of speech for the separated words; A Chinese parsing unit for parsing a Chinese syntax tree based on morpheme parts of speech, a Chinese / Korean translator for converting a Chinese syntax tree to a Korean syntax tree, and converting a word in the Chinese syntax tree into a Korean band word. And, if the Korean syntax tree is a plurality of short phrase tree, using a connection ending determination unit for generating a connection ending between each sentence using the connection ending knowledge DB extracted from the Korean source corpus, and using the Korean syntax tree generated linking ending Rearrange Korean words according to Korean word order, and follow the pattern of Korean grammar and verbs. Through the function words generation it characterized in that it comprises a generator for generating a Korean Korean.

The present invention solves the problem of not expressing an explicit logical connection between the translated Korean clauses when translating a Chinese sentence that does not express the logical connection between clauses with an explicit vocabulary in Korean / Chinese translation. By expressing the logical connection between clauses with Korean connection endings, the translation performance of Chinese / Korean automatic translation can be improved by automatically translating the meaning of the original Chinese text into natural Korean sentences.

In addition, the present invention, if the original Chinese text to be translated has an explicit Chinese short-term logical connection by generating an explicit connection ending from the Korean perspective, by generating a Korean sentence, while maintaining the original meaning of the original Chinese text, high quality Korean There is an advantage that can be generated.

Hereinafter, with reference to the accompanying drawings will be described in detail the operating principle of the present invention. In the following description of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. The following terms are defined in consideration of the functions of the present invention, and may be changed according to the intentions or customs of the user, the operator, and the like. Therefore, the definition should be made based on the contents throughout the specification.

1 is a block diagram of an apparatus for generating Korean connected endings for automatic Korean-Chinese translation according to an embodiment of the present invention, wherein a Chinese input sentence is divided into Chinese words using a Chinese morpheme dictionary and an input sentence for the separated words Based on the morpheme parts of Chinese words separated by the Chinese morpheme analysis unit 101 and the Chinese morpheme analysis unit 101 which select the optimal morpheme parts of speech from the context, the structure of words combined with each other to form a sentence is included in the current sentence. The Chinese syntax tree analyzed by the Chinese syntax parser 102 and the Chinese structure parser 102, which suitably resolves and analyzes the grammatical structural ambiguity to generate the sentence structure of the sentence, is adapted to Korean grammar. Translating into a syntax tree and translating from Korean to Korean bandwords for the words in the Chinese syntax tree When the Korean syntax tree generated by the unit 103 and the Chinese / Chinese translator 103 is composed of two or more short phrase trees, the logical linkage mark between the short sentences is unspecified from the Korean raw corpus. Korean syntax provided by the connection ending determination unit 104, the Chinese / Korean conversion unit 103, and the connection ending determination unit 104 which generate the connection endings between the respective short sentences using the extracted connection ending knowledge DB 106. Using the tree to rearrange Korean words according to the Korean word order and connect them to the Korean language generation unit 105 that generates natural Korean by generating functional words such as surveys and endings according to Korean grammar and verbal modality. Linked-edge knowledge database (DataBase, DB hereinafter) 106 that stores the linked-end knowledge extracted from the Korean native corpus constructed by the mother-based knowledge construction unit 107 and Korean native The connection mother knowledge construction unit 107 extracts the connection mother knowledge from the corpus S1 and stores the connection mother knowledge in the connection mother knowledge DB 106.

Therefore, when the original Chinese text to be translated has an unexplicit Chinese short logical link, the present invention generates an explicit linking ending from the Korean point of view, thereby generating a Korean sentence, while maintaining the original meaning of the original Chinese text while maintaining the high quality Korean. Can be generated.

Hereinafter, a method of generating a Korean linking ending for the Chinese-Korean automatic translation in which the Chinese non-explicit linking in the Korean-Korean automatic translation according to the present invention is generated in natural Korean will be described in more detail.

The Chinese morphological analysis unit 101 arbitrarily divides three possible word separations of all possible combinations using the Chinese morphological analysis dictionary by using a Chinese morphological analysis dictionary because the Chinese sentence is input and the Chinese sentence is not spaced. The total length of the three separated words is divided into the largest word combinations, and when there are several word combinations having the same total length for the separated words, a word combination having the maximum product of each word length of each combination is selected. If the product is also the same, the word combination is repeated by selecting a word combination having a word having a high frequency of one or two Chinese words, and word separation is performed on the input sentence. Each word in the context of the current input sentence after assigning the morpheme parts of speech in An optimal morpheme part-of-speech for the morpheme part-of-speech tagging may be selected and provided to the Chinese parser 102. Here, the morpheme part-of-speech tagging may tag the morpheme parts-of-speech for the separated Chinese using a morpheme part-of-speech tagging method based on the HMM (hidden markov model), which is widely used in all languages such as Korean and English.

The Chinese parsing unit 102 generates a syntax tree for the syntactic structure according to the grammatical role between words in the input Chinese sentence by using the optimal morpheme parts of words separated by the Chinese morphological analysis unit 101. First, based on the result of word separation and morpheme part-of-speech tagging by the Chinese morphological analysis unit (101), the segmentation sentences (sections) are separated and separated using section delimiters such as periods, commas, and colons, and other sectioning pattern rules. By analyzing the syntax that shows the syntactic relationship between words separately for each section, a section parsing tree is generated by the chart parsing method using syntax rules in the extended context-free grammar format. It may be provided to the Chinese / Korean conversion unit 103. Here, a method of selecting a syntax tree that is most suitable for the context of the current sentence among the generated syntax trees is a heuristic structural rule created by considering morpheme parts of speech, syntax relationship information between words, phrase size, and syntax attributes of words. Is selected by the scores and scores, and does not display the syntax relations between the sections separated by the Chinese parser 102, but merely generate independent syntax trees in units of sections.

The Chinese / Korean translator 103 converts the section unit syntax trees generated by the Chinese parser 102 into syntactic and semantically corresponding Korean truncated phrase trees using translation knowledge information called a translation pattern. The conversion pattern is expressed as a Chinese syntax relationship and a corresponding Korean syntax relationship. For example, to express each syntax relationship, predicates, objects, subjects, bores, etc., the syntax information and the central word of each syntactic node ( head word) can be expressed as morpheme part-of-speech information, vocabulary information, semantic information, root information, and syntax attribute information.

In addition, the Chinese / Korean translator 103 visits the section-specific Chinese syntax tree generated by the Chinese parser 102 from the root and prioritizes the head node among the subtrees of each syntax node. When converting a Chinese phrase node into a Korean phrase node through conversion pattern matching while visiting the network, when the phrase node is a terminal node, the phrase information of the Korean phrase node is passed through a band word selection process of converting a Chinese word into a Korean word. The change may be provided to the connection ending determination unit 104. Here, when the Chinese word of the conversion pattern is matched in the band word selection process, it is translated into the corresponding Korean vocabulary in the corresponding conversion pattern, and when the meaning or attributes match, the Korean band word having the meaning or property in the heavy band dictionary If a Chinese word does not match all of the conditions, it can be translated into a representative Korean band word described in the Chinese Band Dictionary.

The linking end determining unit 104 generates a linking ending suitable for the term of the central predicate of each clause for logical connection between clauses when two or more truncated Korean syntax trees converted by the Chinese / Korean conversion unit 103 are used. In the case of expressing a logical connection between clauses in a Chinese input sentence explicitly in a Chinese vocabulary, the information is included in a Korean syntax node converted from a corresponding Chinese syntax node so that it can be expressed as a connection or connection adverb. In the case where the logical connection between the clauses is not expressed in an explicit Chinese vocabulary, the connection ending knowledge DB 106 extracted from the Korean primitive corpus is used to determine whether the attributes between the current clauses match the knowledge of the connection ending extracted from the Korean primitive corpus. We first calculate the probability of generation of the linking end that can represent the logical link between the two clauses. Korean generation unit 105 generates a connection ending between all clauses by grasping the most natural and clear meaning when the connection endings of each clause have a clear combination in terms of the whole sentence, using the probability of generating each connection ending. Can be provided to

In other words, the connection ending determiner 104 inputs a Korean syntax tree in the form of a dependent syntax tree divided by a truncation point obtained by converting the Chinese syntax tree by the Chinese / Korean conversion unit 103. If it is a short sentence and there is only one Korean dependent syntax tree converted by the Sino-Korean translator 103, nothing is done. On the other hand, if the input Chinese sentence is a compound sentence and there is a connection between clauses expressed as non-explicit logical connection, Calculate the probability of generation for each linking end in the sections, and find the linking combination that best fits the whole sentence by using the probability of generating the linking ends of the two cutting points. Instead, a linking ending based on the entire sentence can be generated.

Here, the method of calculating the generation probability for each connection ending between the two clauses is to connect the two clauses of the non-explicit logical concatenation with the dependent clause tree of the first clause as the dependent clause and the syntax clause of the later clause as the governing clause. After generating a verb pattern y in the same form as the verb pattern of the mother knowledge DB 106, the generated verb pattern y is searched in the connection mother knowledge DB 106 to be used for each connection ending e. Probability of generation for Equation 1

Probability of generation of linking ending (e) = 0.8 * Probability of generation of connecting ending (e) by lexical verb pattern + 0.2 * Probability of generation of connection ending (e) by semantic verb pattern

(Here, the description of the lexical verb pattern and semantic verb pattern can be replaced with the description of Table 8 and 9 and the description of each feature value, and it should be indicated as NULL if there is no value of the feature.)

It can be calculated using the formula of.

Here, referring to Equation 1, the generation probability of the connection ending e by the vocabulary verb pattern in the calculation of the generation probability of the connection ending e may be calculated by the following steps.

That is, as a first step, when the generated verb pattern y is in the linking ending knowledge DB 106, the generation probability value of the linking end e in the linking ending information of the found verb pattern is calculated,

As a second step, when the probability of generating the connected ending e by the lexical verb pattern is not determined by the calculation, assuming that the number of subject or object center vocabulary features other than NULL in the verb pattern y is m, i From 1 to m, the following third step may be repeated until the probability of generation of the connecting ending (e) is greater than zero.

As a third step, if any one of the lexical verb patterns created by setting i to NULL among the subject or object center term lexical qualities that is not NULL in the verb pattern y is searched in the linking knowledge DB 106 It is to calculate the average of the probability of generation of the connection ending (e) in the connection ending information of the verb pattern.

Next, referring to Equation 1, the generation probability of the connection ending e by the semantic verb pattern in the calculation of the generation probability of the connection ending e may be calculated by the following step process.

That is, as a first step, in the generated verb pattern y, the semantic verb pattern ys which changes all the non-null subject or object center word vocabulary qualities to the meaning for the corresponding vocabulary is generated in the connected ending knowledge DB 106. If found, calculates the probability of generation of the linking ending (e) in the linking ending information of the found verb pattern,

As a second step, if the calculation probability of generating the linking ending e by the ending verb pattern is not determined by calculation, assuming that the number of subject or object center vocabulary features other than NULL in the semantic verb pattern ys is m, With i from 1 to m, the third step may be repeated until the probability of generation of the connecting end (e) is greater than zero.

As a third step, if any one of the semantic verb patterns created with i NULL in the verb pattern ys or the object core word lexical feature is i NULL, the retrieved semantic verb pattern is retrieved. It is to calculate the average of the probability of generation of the linking ending (e) in the linking ending information of.

Next, since the verb pattern y does not exist in the connected ending knowledge DB 106 as a vocabulary / proverbs pattern, the probability of generating each of the connection endings e of the verb pattern y calculated as described above is all zero. In this case, the default generation probability of each connection ending e may be the generation probability of the connection ending e of the verb pattern y. The default probability of generating each linking end (e) is the sum of the frequency of the linking end (e) in all patterns of the linking end knowledge DB 106 divided by the sum of the frequencies of all linking endes in all the patterns. That is, it may mean the probability that each connection ending e will appear in the connection ending knowledge DB 106.

In other words, when the probability of generating the connection ending between the two clauses is calculated using Equation 1, the connection ending determination unit 104 optimally generates the connection endings to be generated between the non-explicit logical connection clauses in terms of the whole sentence. In order to make the linking end combinations between all two clauses so that they do not appear immediately next to each other when they are not parallel linking ends, the probability of generating the linking end combinations for the whole sentence is generated by multiplying the generation probabilities of the connection endings in each combination. The linking endings in the combination of linking endings with the highest probability values can be selected as Korean linking endings that express the non-explicit logical linking between clauses.

The Korean language generation unit 105 inputs a Korean syntax tree in which a logical connection is explicitly created by the Chinese / Korean conversion unit 103 and the connection ending determining unit 104, and then starts from the root node of the Korean syntax tree. Rearranges the terminal syntax nodes in order of subject, object, and predicate, which is the order of Korean sentence generation, and generates natural words by generating functional words such as surveys and endings according to Korean grammar and verbal modality. can do. Here, when rearranging the terminal syntax node, processing of duplicate verbs used for adjusting and emphasizing adverbs, correcting and deleting the two excessively used words in Chinese, generating surveys according to violent changes, tense, manual / Each terminal syntax node can be generated as a natural Korean word by processing for generating a parent for a passive / mineral and generating a verb for an irregular / regular verb.

The connection ending knowledge DB 106 may store the connection ending knowledge extracted from the Korean native corpus constructed by the connection ending knowledge construction unit 107.

As shown in FIG. 2, the connection ending knowledge construction unit 107 is a block for extracting the connection ending knowledge from the Korean primitive corpus S1 and storing the connection ending knowledge in the connection ending knowledge DB 106. 1071, the Korean syntax analyzing unit 1072, the feature pattern extracting unit 1073 for each connection ending, the connection ending feature pattern DB 1074, and the feature pattern expansion unit 1075.

The Korean morpheme analysis unit 1071 separates the phrases belonging to the sentences into morpheme units for each sentence of the Korean primitive corpus (S1), and tags Korean morpheme parts of speech appropriate to the context, that is, the morpheme vocabulary and the corresponding vocabulary. Corresponding morpheme parts of speech can be assigned to each word by using the morpheme left and right connection table indicating whether access is possible within a word between morphemes according to the Korean morpheme dictionary and access information. . In addition, using the lexical probability and the transition probability learned in the part-of-speech tagged corpus by the Korean morpheme part-of-speech tagging method based on the HMM model, The morphological analysis result tagged with the appropriate morpheme parts of speech may be provided to the Korean syntax parser 1072.

For example, if the Korean sentence extracted from the Korean primitive corpus (S1) says, "The Korean won inflation has risen rapidly this week, domestic consumer prices have risen and income has decreased." The result of morphological analysis tagging the morpheme parts of speech most appropriately: “Don't (noun) + (adjective investigation), won dollar (noun), exchange rate (noun) + (early investigation), rapid (verb) + crab (adverb) Mother (or verb) + (consolidated mother) domestic (noun) consumer (noun) (price) noun + price (main investigation) rising (verb) + (in the past tense first mother) + high (consolidated) income (Noun) + degrees (subordinate) decreased (verb) + (previous preliminary ending ending) + multi (ending ending) +. (Ending sign) ”can be provided to the Korean parsing unit 1072.

The Korean syntax analyzing unit 1072 generates a syntax tree for the syntactic structure according to the grammatical role between the words in the Korean sentence based on the morphological analysis result analyzed by the Korean morphological analysis unit 1071. In Korean, since the central word trailing and the arrangement of words are relatively free, the Korean language has the characteristics of freedom of word order. Therefore, to reflect this feature, the syntax tree is created by using dependency grammar to characterize patterns of connected endings. The extraction unit 1073 may be provided. Here, the syntax tree may generate a dependency tree connecting the dominant node and the dependent node between two words.

For example, FIG. 3 is a diagram illustrating a dependency tree generated by the Korean syntax analyzing unit 1072. Referring to FIG. 3, the node (1) "Nonju (noun) + (subject investigation)" is (5). ) Node is the subordinate node of "or (verb) + sub (connection ending)", on the contrary, node (5) is the dominant node of node (1), and the relationship between two syntax nodes is "adverb". Indicates a relationship between adverbs.

Based on the Korean syntax tree generated by the Korean parsing unit 1072, the feature pattern extracting unit 1073 for the linking end finds the syntax node whose syntactic phrase node is the linking end and finds the control node node of the found syntax node. After finding a verb phrase node that is (dependent node), use the syntax node associated with both verb phrase nodes.

Connected endings: [subordinate clause = verbal vocabulary, negation, subject-centered vocabulary, object-centered vocabulary]
[Domination clause = verbal vocabulary, whether negative, subject-centered vocabulary, object-centered vocabulary]
Verb vocabulary: Verb morpheme vocabulary of the verb phrase node
· Negation: In the sentence for each verb indefinite morphemes
Display 1 if used, 0 otherwise
Main keyword vocabulary: Actual except the functional words of the syntactic node that is the subject relation for the verb node.
Morpheme vocabulary
Object-centered vocabulary: Actual except the functional words of the syntax nodes that are object relations to the verb node.
Morpheme vocabulary

As shown in FIG. 2, feature patterns can be extracted.

That is, in the case of extracting the feature pattern as shown in Table 1, the syntax node is searched in the order of sentence input to find the dependent verb phrase node having the linking end, and the verb phrase node which is the dominant syntax node of the verb phrase node is found. A feature pattern may be generated by extracting a feature related to syntax nodes. If the dominant syntax node contains equivalent linking endings such as “high”, “or”, or “or”, then the current dependent dependency is searched for the dominant syntax node by searching for the dominant syntax node of the dominant syntax node in the pattern of the linking ending of the current dependent syntax node. Extracts the feature pattern related to the dominant syntax node of the syntax node and the dominant syntax node, extracts the feature pattern, and checks whether it is connected to the equal linking end again. If not, you can exit.

In addition, if the linking end of the current dependent syntax node is a linking end that is not a parallel linking end, and there is a syntactic node having the same linking end among the dependent nodes of the current dependent syntax node, A new feature pattern can be extracted using the dependent verb phrase node and the dominant verb phrase node, and iterated over and over again to find out if the connection term of the last dependent phrase node is an equal connection, and if so, repeat the above operation and terminate otherwise. have. The process can be repeated for syntax nodes according to the input word order to extract all feature patterns.

For example, referring to FIG. 3, when searching for a verb including a connection ending in the order of input sentences, the node is searched for node (5), and the search for node (8) is the dominant node of node (5). Search for syntax nodes with node (5) as the control and fill in the subclause qualities above. Then, it searches for the syntax nodes with node (8) as the dominant and fills the governing qualities. In the case of the above-described example sentence, Table 2

Come on: [subordinate = rising, 0, exchange rate, NULL] [controlling = rising, 0, inflation, NULL]

You can extract qualities as shown in

Next, since node (8) contains the equal connection ending “high”, find node (10), which is the dominant node of node (8), make node (5) the subordinate clause, and node (10). As the rule

Come on: [subordinate = rising, 0, exchange rate, NULL] [controlling = decreasing, 0, income, NULL]

As shown in FIG. 6, feature patterns can be extracted.

Next, since feature pattern extraction with node (5) as a subordinate clause has been completed, find the syntax node (8) with the next connection ending and extract the feature pattern for the corresponding governing node (10).

High: [subordinate = rising, 0, inflation, NULL] [controlling = decreasing, 0, income, NULL]

It can be expressed as shown.

As such, the feature pattern extracting unit 1073 for each linking mother extracts the feature pattern as described above for all sentences of the Korean native corpus S1, records the frequency, and provides the feature pattern to the linking mother feature pattern DB 1074. Can be built.

The connection mother feature pattern DB 1074 may store the feature pattern extracted by the feature pattern extracting unit 1073 for each connection mother as a key and use the frequency as a value.

The feature pattern expansion unit 1075 extracts and stores the connection mother feature pattern DB extracted by the feature pattern extracting unit 1073 for each connection mother so as to be able to process the features that do not appear in the Korean native corpus S1. Using the feature patterns of 1074, the new feature pattern is expanded with the combination of the corresponding features according to the connection ending, and the extended feature patterns are converted into the format of the connection ending knowledge DB 106 used by the connection ending determining unit 104. Can be built in the connection mother knowledge DB (106).

Here, the method of extending the new feature pattern can be extended by the following four methods.

In other words, the first feature pattern expansion method is a lexical feature reduction expansion method in the feature pattern extracted by the feature pattern extracting unit 1073 for each ending mother, excluding the verb and the negation feature from the feature of the dependent clause and the control clause. To extend NULL by replacing them with NULLs one by one. At this time, the frequency for the extended feature pattern is

New feature pattern frequency = original feature pattern frequency × 0.5 × (non-NULL subject and object core vocabulary features in extended feature pattern + 1) / (number of subject and object core word vocabulary features in original feature pattern)

It can be calculated using the formula of.

For example, if the feature expansion of the feature pattern of the above example is performed, Table 5

Kinds Feature pattern Frequency original Come on: [subordinate = rising, 0, exchange rate, NULL] [controlling = rising, 0, inflation, NULL] 14 expansion Come on: [subordinate = rising, 0, exchange rate, NULL] [controlling = rising, 0, NULL, NULL] 14 × 0.5 × (1 + 1) / 2 = 7 expansion Come on: [subordinate = rising, 0, NULL, NULL] [controlling = rising, 0, prices, NULL] 14 × 0.5 × (1 + 1) / 2 = 7 expansion Come on: [subordinate = rising, 0, NULL, NULL] [controlling = rising, 0, NULL, NULL] 14 × 0.5 × (0 + 1) / 2 = 3.5

It may have a feature pattern and a frequency as shown. In other words, it is a pattern in which the feature pattern whose type is extension is a newly increased pattern.

As described above, the extended feature patterns are restored to the connection mother feature pattern DB 1074. In this case, if the newly extended feature pattern is already stored in the connection mother feature pattern DB 1074, the frequency can be increased by the frequency that is obtained while the current feature is expanded.

Next, the second feature pattern extension method is a feature pattern transition method, and the control clause of an arbitrary feature pattern according to the connection ending in the connection end feature pattern DB 1074 extended by the first feature pattern extension method. A new feature pattern can be generated by combining subordinate clauses and dominant clauses that do not match in the feature patterns in which the dependent clauses of other feature patterns match. This method is applied to all connected endings, and the frequency for the newly generated feature pattern is calculated by multiplying the average of the frequency of the original feature patterns by a certain weight.

New feature pattern frequency = mean frequency of original feature patterns × 0.4

It can be calculated using the formula of.

For example, when performing feature expansion on feature patterns, Table 6

Kinds Feature pattern Frequency original Come on: [subordinate = rising, 0, exchange rate, NULL] [controlling = rising, 0, inflation, NULL] 14 original Come on: [subordinate = rising, 0, inflation, NULL] [controlling = declining, 0, consumption, NULL] 6 original Come on: [subordinate = ord, 0, exchange rate, NULL] [dominant clause = shrunken, zero, consumed, NULL] ((14 + 6) / 2) × 0.4 = 4

As shown in FIG. 1, the feature pattern and the frequency may be used.

As described above, the extended feature patterns are restored to the connection mother feature pattern DB 1074. In this case, if the newly expanded feature pattern is already stored in the connection feature pattern DB 1074, the frequency is determined. It can be increased by the frequency of the current expansion.

Next, the third feature pattern expansion method is a symmetric expansion method, and the subordinate clauses such as “high”, “or”, “or” in the connection mother feature pattern DB 1074 that have passed through the first and second feature pattern expansion. Even if the order of clauses is changed without distinguishing the dominant clauses, the qualities patterns in the case where the meanings are consistent are expanded by changing the order of the subordinate clauses and the dominant clauses, and the frequency of the extended feature patterns is expressed in Equation 4.

New feature pattern frequency = original feature pattern frequency × 0.2

It can be calculated using the formula of, and in order to calculate the frequency, the order of clauses can not be changed unconditionally due to the diversity of meanings in the case of the linking endings such as “high”, “or”, or “or”. It can be recalculated with low weight for frequency.

For example, if you perform feature expansion on the original feature pattern below,

Kinds Feature pattern Frequency original High: [subordinate = rising, 0, inflation, NULL] [controlling = decreasing, 0, income, NULL] 8 original High: [subordinate = declining, 0, income, NULL] [controlling = rising, 0, inflation, NULL] 8 × 0.2 = 1.6

As shown in FIG. 1, the feature pattern and the frequency may be used.

As described above, like the first and second methods, the extended feature patterns are restored to the connection mother feature pattern DB 1074, and the newly expanded feature patterns are already stored in the connection feature pattern DB 1074. If so, you can increase that frequency by the number of frequencies you have as you expand it.

Finally, the fourth feature pattern extension method uses a semantic dictionary that uses subordinate and dominant clause subject and object-centered word vocabulary in heavy translation in the linkage feature pattern DB 1074 that has undergone the feature pattern extension methods as a semantic generalization extension method. To expand the vocabulary to its meaning. In this case, when generalizing and expanding the vocabulary, a value obtained by multiplying the frequency of the original feature pattern by 0.1 may be used as the frequency of the extended feature feature pattern. The newly created semantic feature pattern and its frequency may also be stored in the connection mother feature pattern DB 1074. If the feature pattern is already stored, the semantic feature pattern and the frequency may be re-stored by adding the current frequency.

In addition, the feature pattern extension unit 1075 changes the connection end feature pattern DB 1074 including the extended feature patterns into the connection end knowledge DB 106 used by the connection end determination unit 104. , Connection ending knowledge DB 106 is shown in Table 8

Pattern Connected Mother Information [Subordinate clause = verbal vocabulary, whether negative, subject-centered vocabulary, object-core vocabulary] [dominant clause = verbal vocabulary, negation, subject-core vocabulary, object-core vocabulary] (Connection probability 1 generation probability 1), (connection probability 2 generation probability 2),... , (Probability of concatenation p)

It may have a form as shown in.

That is, in the feature pattern of the connection ending feature pattern DB 1074, the pattern of the subordinate clause and the dominant clause except the connection ending is changed to the verb pattern of the connection ending knowledge DB 106, and the connection is made using the connection ending of the feature pattern and the corresponding frequency. It can be calculated with the mother information, and the probability of generating the linked mother in the calculated linked mother information can be calculated by dividing the frequency of the verb pattern for each linked mother by the total verb pattern frequency. Therefore, the sum of the generation probabilities in the linking end information for each term pattern becomes 1.0, and the linking end information is stored as a value in the linking end knowledge DB 106 and the linking end information is stored as a value. .

For example, the verb pattern and the connection end information stored in the connection mother knowledge DB 106, Table 9

Key (Proverbs Pattern) Value (connection ending information)
[Subordinate = rising, 0, exchange rate, NULL] [controlling = rising, 0, inflation, NULL] (0.5622 [= 14 / 24.9]),
(High 0.1687 [= 4.2 / 24.9]),
(0.1406 [= 3.5 / 24.9])
(Page 0.1285 [= 3.2 / 24.9])

As shown in, the value between [] in the linking mother information shows the process of calculating the generation probability.

Therefore, in the case of translating a Chinese sentence that does not express a logical connection between clauses into an explicit vocabulary in Chinese / Chinese translation, the present invention does not express the problem of not expressing an explicit logical connection between translated Korean clauses. By expressing the logical connection between clauses with Korean connection ending, the translation performance of Chinese / Korean automatic translation can be improved by automatically translating the meaning of Chinese original text into natural Korean sentences.

Meanwhile, in the detailed description of the present invention, specific embodiments have been described, but various modifications are possible without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the scope of the following claims, but also by those equivalent to the scope of the claims.

1 is a block diagram of an apparatus for generating Korean connected endings for automatic Korean-Chinese translation according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating a connection mother knowledge construction unit for extracting connection mother knowledge from a Korean primitive corpus and constructing the connection mother knowledge DB according to an embodiment of the present invention; FIG.

3 is a view showing a dependency tree generated by the Korean syntax analysis unit according to an embodiment of the present invention.

<Description of the symbols for the main parts of the drawings>

101: Chinese morphological analysis unit 102: Chinese syntax analysis unit

103: middle and Korean conversion unit 104: connection ending determination unit

105: Korean generation unit 106: Connected mother knowledge DB

107: Connected mother knowledge building

Claims (15)

Dividing a Chinese input sentence into Chinese words and selecting a morpheme part of speech for the separated words; Generating a Chinese syntax tree based on the morpheme parts of speech; Converting the Chinese phrase tree into a Korean phrase tree, and converting a word in the Chinese phrase tree into a Korean band word; Generating a connection ending between each short sentence using a connection ending knowledge DB extracted from a Korean raw corpus when the Korean syntax tree is a plurality of short syntax trees; Rearranging Korean words according to Korean word order using the Korean syntax tree in which the connected endings are generated, and generating Korean through a function of generating a functional word according to Korean grammar and verbal modality; Extracting connection mother knowledge from the Korean primitive corpus and constructing the connection mother knowledge in the DB; Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 1, Generating the connection mother, Obtaining a probability of generating a Korean connection ending using the connection ending knowledge DB automatically extracted from the Korean source corpus by inputting the Korean syntax tree; Selecting a linkage combination between clauses suitable for the whole sentence based on the generation probability of the linkage ending; Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 2, Probability of generation of the connecting mother, Equation Probability of generation of linking ending (e) = 0.8 * Probability of generation of connecting ending (e) by lexical verb pattern + 0.2 * Probability of generation of connection ending (e) by semantic verb pattern A method of generating Korean concatenated endings for Chinese-Korean automatic translation, which is calculated using the formula of. The method of claim 3, wherein The probability of generating the connection ending e by the lexical verb pattern is When the verb pattern is in the connection ending knowledge DB, a calculation probability value of the connection ending e in the connection ending information of the found verb pattern is calculated. When the probability of generating the connected ending e by the lexical verb pattern is not obtained by the above calculation, i is 1 to m if the number of subject or object center vocabulary features other than NULL in the verb pattern y is m. Repeating until the connection probability (e) is greater than zero; If any one of the lexical verb patterns made by setting i to NULL from the subject or object center term lexical features that are not NULL in the verb pattern y is searched in the linking knowledge database, the linking mother information of the found verb pattern Calculating the average of the probability of generation of the connected endings (e) Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 3, wherein The probability of generating the connection ending e by the semantic verb pattern is In the verb pattern, if the subject or object central word vocabulary feature that is not NULL is generated and changed in the meaning for the corresponding vocabulary, the semantic verb pattern (ys) is generated and searched in the connection term knowledge DB. Calculating the probability of generating a connection ending (e) in When the probability of generating the connected ending e by the ending verb pattern is not determined by the calculation, if the number of subject or object central vocabulary features other than NULL in the semantic verb pattern ys is m, i is 1 from 1. repeating until the generation probability of the connecting end (e) is greater than 0 up to m, and If any one of the semantic verb patterns created with i or NULL in the verb pattern ys is set to i NULL as the object center object vocabulary feature, it is found in the linked ending information of the found semantic pattern. Calculating an average of the probability of generating the connection ending (e) Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 2, The step of selecting the inter-connection linkage combination, If it is not a parallel connection, avoiding neighbors, Calculating the probability of generating a combined ending combination for the whole sentence by multiplying the probability of generating the associated endings in the combination; Selecting a combination having the highest calculated probability of generation; Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 1, The step of building in the connection mother knowledge DB, Separating morpheme units by inputting sentences one by one from the Korean primitive corpus, and selecting morpheme parts of speech for the separated morphemes; Generating a Korean phrase tree by analyzing syntactic relationships between words using the selected morpheme parts of speech; Extracting a feature pattern according to a linking ending using a generated syntactic syntax node with a linking ending and a syntax node corresponding to a dominant node of the verbal syntax node using the generated Korean syntax tree; Combining the connection patterns which are not extracted from the Korean native corpus from the connection ending feature pattern DB and combining the connection patterns according to the connection ending to expand the new feature pattern to build in the connection ending knowledge DB. Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 7, wherein Extracting and patterning the qualities may include: Extracting a feature pattern using a syntax node related to a verb phrase node based on the Korean syntax tree; Repeating the step of extracting the feature pattern when the syntax node is connected with an equal connection ending; If the syntax node is a linking end and there is a verb syntax node having a parallel linking end in the dependent syntax node, a new feature pattern is used for the linking end of the dependent syntax node by using the dependent syntax node and the governing syntax node of the dependent syntax node. Extracting the Repeating the step of extracting the new feature pattern when the connection ending of the last dependent phrase node is an equal connection ending Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 7, wherein Expanding the new feature pattern to build in the connection mother knowledge DB, Expanding the feature by changing the subject and object central word vocabulary, except for the verb and the negation feature, from the features of the dependent clause and the governing clause to NULL by the lexical feature reduction expansion method in the extracted feature pattern; A new feature pattern is generated by combining the inconsistent subordinate clause and the dominant clause in the feature patterns in which the control clause of an arbitrary feature pattern and the subordinate clause of another feature pattern coincide with the feature pattern transition method for the extended feature. Steps, For the features expanded by the lexical feature reduction method and the feature pattern transfer method, even if the order of the clauses is changed by the symmetric expansion method without the distinction of the subordinate clause or the dominant clause, Expanding, Extending the vocabulary to the corresponding meaning by using the semantic dictionary used in the Chinese-to-Korean translation of the subject and object of the subordinate clause and the governing clause with the feature pattern extension method Method of generating Korean connected endings for automatic Korean-Chinese translation. The method of claim 9, The frequency for the feature pattern expanded by the lexical feature reduction extension method is Equation New feature pattern frequency = original feature pattern frequency × 0.5 × (non-NULL subject and object core vocabulary features in extended feature pattern + 1) / (number of subject and object core word vocabulary features in original feature pattern) A method of generating Korean concatenated endings for Chinese-Korean automatic translation, which is calculated using the formula of. The method of claim 9, Frequency for the feature pattern by the feature pattern transfer method, Equation New feature pattern frequency = mean frequency of original feature patterns × 0.4 A method of generating Korean concatenated endings for Chinese-Korean automatic translation, which is calculated using the formula of. The method of claim 9, Frequency for the feature pattern by the symmetric expansion method, Equation New feature pattern frequency = original feature pattern frequency × 0.2 A method of generating Korean concatenated endings for Chinese-Korean automatic translation, which is calculated using the formula of. A Chinese morphological analysis unit that separates Chinese input sentences into Chinese words and analyzes morphemes to select morpheme parts of speech for the separated words; A Chinese parsing unit for parsing the syntax to generate a Chinese syntax tree based on the morpheme parts of speech; A Chinese / Korean conversion unit for converting the Chinese syntax tree into a Korean syntax tree and converting a word of the Chinese syntax tree into a Korean band word; A connection ending determination unit for generating a connection ending between the short sentences using the connection ending knowledge DB extracted from the Korean source corpus when the Korean syntax tree is a plurality of short syntax trees; Korean generation unit for rearranging Korean words according to Korean word order using the Korean syntax tree in which the connected endings are generated, and generating Korean through a function word generation function according to aspects of Korean grammar and verbs Korean connected endings generating device for automatic Chinese-to-Korean translation. The method of claim 13, The Korean connected mother generation device, Connection mother knowledge construction unit for extracting the connection mother knowledge from the Korean native corpus to build in the connection mother knowledge DB Korean-connected ending generation device for the Chinese-Korean automatic translation further comprising. The method of claim 14, The connecting mother knowledge building unit, A Korean morphological analysis unit for separating morpheme units by inputting sentences one by one from the Korean primitive corpus and selecting morpheme parts of speech from the separated morphemes; A Korean syntax analysis unit for generating a Korean syntax tree by analyzing syntactic relations between words using the selected morpheme parts of speech; A feature pattern extracting unit for each connection ending extracting a feature pattern according to a connection ending around the verb phrase node having a connection ending and the syntax node corresponding to the dominant node of the verb phrase node using the generated Korean syntax tree; , A linked mother feature pattern DB for storing the extracted feature pattern as a key and storing the frequency as a value; A feature pattern extension unit for constructing the connection pattern knowledge DB to expand the new feature pattern by combining the connection patterns which are not extracted from the Korean native corpus from the connection mother feature pattern DB. Korean connected endings generating device for automatic Chinese-to-Korean translation.

Images