KR100512541B1 - Machine translation machine and system, method - Google Patents

Abstract

The present invention includes an input step of inputting an input sentence written in a source language; A lexical analysis step of morphologically analyzing the input sentence; A syntax analysis step of revealing the syntactic structure of the input sentence, and a meaning analysis step of analyzing the meaning of the input sentence; A target sentence step of generating an output sentence translated into a target language based on the syntax analysis and the semantic analysis; A machine translation method having an output step of outputting the output sentence, wherein after the lexical analysis step, a rule applying step of applying a sentence division rule to the input sentence morphologically analyzed in the lexical analysis step, and applying the rule A machine translation method further comprising: a position selection step of selecting a position where a sentence can be divided, an analysis step of obtaining a probability value through the probability distribution, and a decision step of determining sentence division based on the probability value, and a machine translation using the same Provide devices and systems.

Description

Machine translation machine and system, method

The present invention relates to a machine translation apparatus, system, and method for translating an input sentence described in a primitive language (first language) and outputting an output sentence described in a target language (second language). A machine translation apparatus, system, and method for reducing the temporal and spatial complexity of syntax and semantic analysis by segmentation.

Referring to FIG. 1, the conventional machine translation method includes an input step 10 for inputting an input sentence written in a source language, a lexical analysis step 20 for morphological analysis of the input sentence, and a syntactic structure of the input sentence. A syntax analysis step 40, a semantic analysis step 50 for analyzing the meaning of the input sentence, a target sentence step 60 for generating an output sentence translated into a target language based on syntax analysis and semantic analysis, And an output step 90 for outputting the output sentence.

This conventional machine translation (English-Korean machine translation) targets translation between two languages (Korean, English) belonging to different cultures. Since these languages show considerable differences in their structure or word order, idiom-based analysis methods have been used in the past to obtain accurate analysis and natural translation. However, the idiom-based analysis method recognizes idioms before analysis and treats them as a single unit of analysis. The ambiguity of idioms can generate many units of idioms, which increases the complexity of sentence analysis. There is this.

In order to reduce the complexity of sentence analysis, various sentence splitting methods have been proposed. One of them is by partial parsing. This method breaks up the statements into chunks and analyzes them by chunk. chunks correspond to the language performance structures presented by Gee and Grosjean [3]. Chunks correspond to phrase structures such as NP (noun phrase), VP (verb phrase), and PP (prepositional phrase). They consist of a combination of one central word and several functional words. These chunks can be analyzed by the context-free grammar which is commonly used for natural language analysis. However, since the language application structure corresponds to a part of the sentence that is spoken at a time when a person speaks, a prosodic pattern, it is difficult to analyze the relationship between chunks based on this based on context-free grammar.

Another method using pattern rules has been proposed to segment long English sentences. The division pattern rule was composed by a person and the sentences corresponding to the rule were divided to analyze each segment (segment) independently, and the analysis results were synthesized to generate the entire sentence structure. This method can reduce the complexity of parsing the sentence to which the pattern rule can be applied, but there is a problem that is not practical because it is impossible for a person to regularize a pattern of a long sentence.

 In another method, a multi-layered pattern matching method has been proposed to segment long Japanese sentences. However, likewise, there was a problem in that a long sentence pattern had to be constructed, and only the sentences matching the pattern could be divided. In addition, if the divided short sentence does not have a subject, there is a problem that an additional algorithm for finding the subject is required.

 Another way of analyzing English is to divide sentences using the fact that English declarative sentences almost always consist of a combination of three consecutive parts (front part + part + part). It became. This method attempts to reduce the complexity of syntax analysis by dividing a sentence into three parts using the pattern matching capability of neural networks. This method has a problem mainly applicable to simple sentences. In addition, this method has a problem in that it is difficult to apply to a coordinate sentence or a complex sentence having a plurality of subjects and descriptions.

 As another method, a sentence division method using a sentence pattern has been proposed. This method analyzes the divided segments individually and combines each analysis result as instructed in the pattern to create a sentence structure. However, this method also had a problem in that a person constructs a long sentence pattern, that is, a sentence of a Chinese sentence or a compound sentence, and divides the sentence using the sentence.

In order to solve such a conventional problem, it is an object of the present invention to provide a machine translation apparatus, system, method for reducing human effort through the generation of segmentable position classification rules by learning.

It is also an object of the present invention to provide a machine translation apparatus, system, and method capable of safe division of accuracy that can be practically applied through division position determination by a maximum entropy probability model.

It is also an object of the present invention to provide a machine translation apparatus, system, and method for maintaining the application rate and accuracy of a certain level or more even in the sentence segmentation of learning data and other areas.

It is also an object of the present invention to provide a machine translation apparatus, system, and method for performing machine translation in real time by improving the efficiency of divisional analysis by sentence division.

In order to achieve the above object, the present invention includes an input unit for inputting an input sentence written in a source language; A lexical analysis module for stemming the input sentence, a syntax analysis module for revealing the syntactic structure of the input sentence, a semantic analysis module for analyzing the meaning of the input sentence, a target based on the syntax analysis and the semantic analysis A central processing means equipped with a target sentence generation module for generating an output sentence translated into a language; A machine translation apparatus having an output unit for outputting the output sentence, the machine translation apparatus comprising: a classification rule DB storing a classification rule of a position at which a sentence can be divided; The apparatus further includes a probability distribution DB in which a probability distribution of sentence division possible positions is stored, and the central processing unit selects a segmentation position by applying a classification rule stored in the classification rule DB to the input sentence, and selects the selected division. The machine translation apparatus further comprises a sentence division module for determining a sentence division of the input sentence by obtaining a probability value using a probability distribution stored in the probability distribution DB.

In another aspect, the present invention is a machine translation system having a terminal, a server, and an open network connecting the terminal and the server to translate an input sentence written in a source language into a target language and output the output sentence as an output sentence. An input unit for inputting an input sentence written in a language and an output unit for outputting an output sentence translated into a target language; The server may include a classification rule DB in which a classification rule of a sentence dividable position is stored, a probability distribution DB in which a probability table of a sentence dividable position is stored, a lexical analysis module for stemming the input sentence, and a syntactic structure of the input sentence. A syntax parsing module for identifying a language, a semantic analysis module for analyzing the meaning of the input sentence, a target sentence module for generating an output sentence translated into a target language based on the syntax analysis and the semantic analysis, and the input sentence. Segmentation is determined by applying a classification rule stored in the classification rule DB, and a sentence value is determined using a probability distribution stored in the probability distribution DB. Provided is a machine translation system having a central processing unit equipped with a module.

In addition, the present invention, the classification rule DB constructs a plurality of corpus marked with the split position to generate the learning data, and using the learning data to learn the concept of the splittable position to generate a classification rule of the segmentable position Machine translation apparatus and system characterized in that the stored.

In addition, the present invention, the probability distribution DB to build a plurality of corpus marked with the split position to generate the training data, to generate a rule and / or characteristics using the training data to create a probability model, the probability model It provides a machine translation apparatus and system, characterized in that by generating a probability distribution using the stored.

In another aspect, the present invention provides a machine translation apparatus and system, characterized in that the sentence segmentation module adds and stores the output sentence in the classification rule DB as a segmentable position, modifies a probability model and stores it in a probability distribution DB. .

In another aspect, the present invention, the input step of inputting an input sentence written in the source language; A lexical analysis step of morphologically analyzing the input sentence; A syntax analysis step of revealing the syntactic structure of the input sentence, and a meaning analysis step of analyzing the meaning of the input sentence; A target sentence step of generating an output sentence translated into a target language based on the syntax analysis and the semantic analysis; A machine translation method having an output step of outputting the output sentence, wherein after the lexical analysis step, a rule applying step of applying a sentence division rule to the input sentence morphologically analyzed in the lexical analysis step, and applying the rule There is provided a machine translation method further comprising: a position selection step of selecting a position where a sentence can be divided, an analysis step of obtaining a probability value through the probability distribution, and a decision step of determining sentence division based on the probability value.

The present invention also provides a machine translation method, wherein the syntax analysis step divides the input sentence into segments by sentence division according to the determination step, and synthesizes the segments to reveal the syntactic structure of the input sentence. to provide.

In another aspect, the present invention, the sentence division rule applied in the rule applying step is generated by building a plurality of corpus marked with the split position to generate the learning data, the learning data is generated by learning the concept of the segmentable position It provides a machine translation method characterized in that.

In addition, the present invention, the probability distribution to be applied in the analysis step to construct a plurality of corpus marked with the split position to generate the training data, and to generate a rule and / or characteristics using the training data to create a probability model The present invention provides a machine translation method, which is generated using the probability model.

In addition, the present invention, after the analysis step, further comprises a storage step of storing the probability value obtained in the analysis step, wherein the determining step is characterized in that the sentence division to determine the position having the maximum probability value of the stored probability value in the storage step It provides a machine translation method.

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

Referring to FIG. 2, the machine translation apparatus according to the present invention includes an input unit 1, an output unit 2, a central processing unit 3, a classification rule DB 4, and a probability distribution DB 5; Have

The input unit 1 is an input device such as a keyboard, a mouse, or a scanner for inputting an input sentence written in a source language. The output unit 2 is an output device such as a monitor or a printer that outputs an output sentence output in a target language.

The central processing unit 3 includes a lexical analysis module 6A for morphological analysis of the input sentence of the input unit 1, a sentence division module 6B for dividing the input sentence into segments, and a syntactic structure of the input sentence. It includes a parsing module 6C, a semantic analysis module 6D for analyzing the meaning of the input sentence, and a target sentence module 6E for generating an output sentence translated into the target language based on the parsing and semantic analysis.

The classification rule DB 4 generates a learning data by constructing a plurality of corpus in which the division positions are displayed, and learns the concept of the dividable position using the learning data to generate and store a classification rule of the dividable position.

Probability distribution DB (5) generates a learning data by building a plurality of corpus marked with the split position, create a probability model by generating rules and / or characteristics using the training data, and generates a probability distribution using the probability model Are stored.

The sentence dividing module 6B of the central processing unit 3 selects the dividable position by applying the classification rule stored in the classification rule DB 4 to the input sentence inputted by the input unit 1. The sentence division module 6B determines the sentence division of the input sentence by obtaining a probability value using the probability distribution stored in the probability distribution DB 5 at the selected segmentable position.

In addition, the sentence segmentation module 6B further stores the output sentence in the classification rule DB 4 as a segmentable position, modifies the probability model, and stores it in the probability distribution DB 5. By further storing the partitionable position and modifying the probability model, a more accurate and rich classification rule DB 4 and a probability distribution DB 5 are prepared.

The machine translation apparatus according to the present invention has been described above, but the present invention is not limited thereto.

The present invention provides a client PC having an input unit 1 and an output unit 2, a server having a classification rule DB 4, a probability distribution DB 5, and a central processing unit 3 equipped with the above modules; It may be a machine translation system having an internet network connecting a client PC and a server. Such a machine translation system can be used to translate the results retrieved on the Internet into a desired language in real time.

Referring to FIG. 3, the machine translation method according to the present invention includes an input step 10 for inputting an input sentence written in a source language, a lexical analysis step 20 for morphological analysis of the input sentence, and a syntactic structure of the input sentence. Parsing step (40) for revealing, a semantic analysis step (50) for analyzing the meaning of the input sentence, the target sentence generation step (60) for generating an output sentence translated into the target language based on the parsing and semantic analysis And having an output step 90 for outputting an output sentence is the same as the conventional machine translation method.

The machine translation method according to the present invention further includes a sentence division step 30 after the lexical analysis step 20.

FIG. 4 is a detailed flowchart of the sentence segmentation step 30 of the machine translation method of FIG. 3. Referring to FIGS. 3 and 4, the sentence segmentation step 30 first applies a sentence segmentation rule to an input sentence whose morpheme has been analyzed. Apply (31).

FIG. 5 is a flowchart illustrating a classification rule generating method used in the sentence division rule applying step 31 of FIG. 4. Referring to FIG. 5, a process of generating a sentence division rule will be described. Build passively (102). A corpus is a set of sentences that show a split position, where each sentence is constructed with the appropriate split positions marked by the human reading it. The person who builds the corpus is preferably someone who has some knowledge of the source grammar. The marked splitting positions in these corpus have the characteristics that people consider when splitting sentences.

Next, learning data (lexical context) is generated from this corpus and stored (104). The learning data (lexical context) is divided into two parts of the sentence: whether or not w _i can be divided, the word (w _i ), the left two words of w _i and the two right words (four), and the left two of w _i . part of speech and the right second part of speech of the words of the word (4) and a sub-categorization (subcategorization) information (w _i is the binary value representing the authorized company to take the section in direct object) of the two words the left side of the w _i, etc. There are 12 attributes in all (see Table 1).

s_pos word _i w _i-2 , w _i-1 , w _{i + 1} , w _{i + 2} p _i-2 , p _i-1 , p _{i + 1} , p _{i + 2} s_cat _{i_2} s_cat _{i_1}

Next, a concept of defining a segmentable position through learning of the training data is obtained (106).

Next, an aggregated representation of the active vocabulary context represented by the version graph is generated by learning the segmentable position concept using the learning data to obtain a rule represented by the path between the generalization boundary and the specialization boundary on the version graph (108). . The classification rule thus obtained is stored in the classification rule DB 4 of FIG.

3 and 4, the sentence division step selects a segmentable position of an input sentence by applying classification rules stored in the classification rule DB 4 (32). The splittable position calculates a probability value of the selected splittable position using the probability distribution stored in the probability distribution DB 5 (33). This segmentation probability value represents the degree of belief that it will be a segmentable position.

FIG. 6 is a flowchart of a probability distribution generating method used in the probability analysis step 33 of positioning. Referring to FIG. 6, the probability distribution is generated based on the principle of maximum entropy. In the same manner as in the process of storing the classification rule of FIG. 5, the probability distribution constructs a corpus in which the split position is displayed (112), generates learning data (lexical context) (114), and generates a rule (116).

Next, various factors and candidate features are taken into consideration in order to achieve a large efficiency improvement in determining the split position and generate a safe segment (118). The candidate characteristics are (1) the lexical contextual characteristics of the word, (2) the local characteristics (location information) of the words, and (3) the characteristics of the initial splitting position (whether or not there are other splitting positions in front). (1) Vocabulary contextual features are extracted from rule 110 for segmentable location classification. The classification rules express the attributes of the lexical context of segmentable practices, and the frequently occurring attributes provide strong evidence for segmentation positioning. The classification rule is defined as a path between the generalization boundary and the specialization boundary on the version graph, and features are extracted while tracking this path. 7 shows an algorithm for lexical contextual feature extraction. (2) The regional characteristics of the word are used to consider the same word because the preference that can be selected as the split position is different depending on the position in the sentence. In a sentence of n words, the region value (position value) of the i th word is

Where R means the number of regions of the sentence.

(3) The characteristics of the initial dividing position, i.e. whether another dividing position is present in front, are considered for safe dividing. For example, the first splitting position in a sentence is a position where a person splits a sentence for the first time, so it is relatively safe compared to other splitting positions.

The probability distribution is generated based on the maximum entropy principle in consideration of given characteristics (120). The probability distribution p * is represented by the following equation 2 by the principle of maximum entropy.

For generating the probability distribution, the probability variables X and Y are distributed exponentially, and the probability rule distribution is expressed by a probability model of a series of conditional exponential functions as shown in Equation 3 below.

Generalized Iterative Scaling (GIS) algorithm is used to calculate the weight of this probability model (see Fig. 8).

As a final step in the probability analysis 33 of partition position selection, a probability distribution is generated based on the principle of maximum similarity (122). The principle of maximum similarity indicates that finding a probability model with maximum similarity is the same as finding a probability model with maximum entropy. Maximum Likelihood Principle is represented by the following equation (4).

Considering only one of the candidate features 118 useful for splitting position determination in the probability model maintains the accuracy of splitting position determination using the generated model while reducing computation time for model generation. In order to consider only the useful features among the candidate features, the incremental feature selection (IFS) algorithm of FIG. 9 is used. However, the frequency-based feature selection (FFS) can be used. It may be. The probability distribution generated by the above process is expressed by the following equation (5).

Here, x represents a context situation (information) to be considered in determining the split position, and y represents a split position as a value of 0 or 1.

Referring to FIG. 4 again, when the split probability value is determined based on the characteristic of the context of the splittable position, the split position is determined by calculating the split probability of the splittable position (34). An algorithm for division position determination is shown in FIG. At this time, the set A is a set of divisible positions whose split probability is larger than the threshold of each word, and B means a set of other divisible positions. This threshold is expressed by the following equation (6).

This threshold means an expected value at which a particular word will be a split position, and if there is a probability greater than this value, the word is considered appropriate as the split position. The split probability above the threshold means that the split position is appropriate as the split position, among which the questionable position having the largest probability is determined as the split position. If the split probability of all the splittable positions is less than the threshold, it means that the reliability as the split position is low. In this case, the size of the segment resulting from the segmentation is also considered. In this case, the split probability has a value from 0 to 1, and the size of the segment has an integer value greater than 1. In order to equalize the effects of the two factors on partition position determination, the segment size values are normalized from 0 to 1. The sum of these two values determines the score of the divided position, and the position having the largest score is determined as the divided position. Since the determination of the split position should be divided into segments of easy parsing length, the segmentation continues until no segment longer than a certain length exists. Here, among the factors considered in determining the split position, whether the split information is different from the position information is different in each sentence division, so the split probability value is different. 11 shows an algorithm for dividing into segments of a certain length or less.

Referring again to FIG. 3, after the sentence is divided, the segment is analyzed in the syntax analysis step 40, and the syntax is analyzed by synthesizing the segments. After the syntax analysis step 40, a target sentence is generated through semantic analysis 50 (60). After checking whether the target sentence is satisfied, if it is not satisfied, the parsed step 40 is repeated by manually inputting the modified sentence segmentation 70. If satisfied, the sentence division classification rule is added to the classification rule DB (4), and the probability distribution stored in the probability distribution DB is modified (80). This last step may be omitted.

Example

In order to generate a rule for classifying segmentable locations, 3,000 sentences were extracted from 10,300 sentences of Wall Street Journal. A sentence without commas was translated into sentence sentences by extracting 300 sentences from 2,640 high school English textbooks, 1,000 byte magazines in the computer field, and 1,200 sentences in the Washington Post politics. .

5,375 active vocabulary contexts and 40,236 inactive vocabulary contexts were generated from 3,000 sentences of the training data. As a result of learning the segmentable position concept, 360 version graphs having 9,002 nodes were generated, and 5,851 segmentable position classification rules were generated from this. We generated 6,596 candidate characteristics from the generated rule and used the probabilistic model. Among the feature selection methods, FFS showed 2,866 features at 12 minutes of creation time and 1,853 features at 1,115 minutes of creation time. IFS requires considerably more time than FFS, but it is advantageous in calculating probability because it provides split probability considering only a few useful features.

According to the division rule that the application rate, accuracy, and segmentation error sentences (partition contribution value using the number of sentences without division errors and the number of sentences to be divided) of the machine translation results according to the machine translation method according to the present invention are made by man Compared to the rule-based method of partitioning. This rule-based method builds a rule for defining segmentable positions through observation of the context-free grammar used for long sentence analysis. Dividable locations are categorized by type and partition priority is assigned to each type to determine a split position. The split position is determined in consideration of the split priority of the splittable position and the size of the segment created by the split.

Table 2 compares the partitioning performance of sentence division according to the present invention.

Application rate accuracy Split error sentences Standard method 100 77.6 0.776 Rule-based method 85.2 86.5 0.703 FFS 98.3 88.2 0.865 IFS 98.3 91.2 0.895

In terms of application rate, accuracy, and segmentation performance by IFS, segmentation error sentences were 0.913 for test sentences and 0.906 for high school textbooks. It was 0.883 for the byte magazine in the computer domain and 0.87 for the political post in the Washington Post. The overall average value is 0.895, which indicates that about 90% of sentences generate the correct syntax structure with low analysis complexity through sentence division of the machine translation method according to the present invention.

In addition, the parsing efficiency is that the sentence is more than 20 words when the sentence is not divided, the analysis is not finished in many cases. Table 3 shows the time, space and efficiency improvement of the analysis using segmentation and the analysis without segmentation.

Analysis using segmentation Analysis Without Segmentation Improve effect (%) Wall Street Journal Time in seconds 4.8 22.1 77.4 MB space 1.1 3.9 71.8 High School English Textbook Time in seconds 4.6 19.6 76.5 MB space 0.9 3.4 73.5 Byte Magazine Time in seconds 5.4 25.1 78.5 MB space 1.1 3.7 70.3 Washington Post Time in seconds 5.1 29 82.4 MB space 1.1 4.3 74.4

In case of sentence segmentation according to the invention, efficiency improvement of 30.9% in time and 57.8% in space was obtained compared to analysis without sentence segmentation. In addition, it could be applied regardless of the type or length of sentences, and showed about 98% application rate and about 90% segmentation accuracy.

Machine translation apparatus, system and method according to the present invention can reduce the human effort through the generation of a segmentable position classification rule by learning, the accuracy that can be practically applied through the determination of the segmentation position by the maximum entropy probability model It has the effect of safe partitioning.

In addition, the machine translation apparatus, system and method according to the present invention can maintain the application rate and accuracy over a certain level even in the sentence segmentation of the learning data and other areas, and can perform the practical machine translation in real time by improving the efficiency of the segmentation analysis by the sentence segmentation. It can be effective.

1 is a flowchart of a conventional machine translation method.

2 is a conceptual diagram of a machine translation apparatus according to the present invention.

3 is a flowchart of a machine translation method according to the present invention;

4 is a detailed flowchart of sentence division of the machine translation method of FIG.

5 is a flowchart of a classification rule generating method used for applying the sentence division rule of FIG. 4; FIG.

FIG. 6 is a flowchart of a probability distribution generating method used for probability analysis of position selection in FIG. 4; FIG.

7 is an algorithm for lexical contextual feature extraction in the probability distribution generation method.

8 is a GIS algorithm in a probability distribution generation method.

9 is a gradual feature selection algorithm in the probability distribution generation method.

10 is a partition position determination algorithm in the probability distribution generation method.

11 is a sentence segmentation algorithm in the probability distribution generation method.

Claims (13)

An input unit for inputting an input sentence written in a source language; A lexical analysis module for stemming the input sentence, a syntax analysis module for revealing the syntactic structure of the input sentence, a semantic analysis module for analyzing the meaning of the input sentence, a target based on the syntax analysis and the semantic analysis A central processing means equipped with a target sentence generation module for generating an output sentence translated into a language; A machine translation apparatus having an output unit for outputting the output sentence, A classification rule DB storing a classification rule of a sentence splittable position; Build a plurality of corpus marked with the split position to generate training data, generate rules and / or properties using the training data to create a probability model, and generate probability of sentence segmentation using the probability model. It has more probability distribution DB where the index is stored. The central processing unit selects a segmentable position by applying a classification rule stored in the classification rule DB to the input sentence, and obtains a probability value using the probability distribution stored in the probability distribution DB using the selected segmentable position. And a sentence dividing module for determining a sentence dividing of the input sentence. The method of claim 1, The classification rule DB may be configured to generate a plurality of corpus in which the division positions are displayed to generate learning data, and use the learning data to learn a concept of a dividable position to generate and store a classification rule of a dividable position. Machine translation device. delete The method of claim 2, And the sentence splitting module adds and stores the output sentence in the classification rule DB as a segmentable position, and modifies and stores a probability model in the probability distribution DB. A machine translation system having a terminal, a server, and an open network connecting the terminal and the server to translate input sentences written in a source language into target sentences and output them as output sentences. The terminal has an input unit for inputting an input sentence written in a source language and an output unit for outputting an output sentence translated into a target language; The server may include a classification rule DB in which a classification rule of a sentence division position is stored; Build a plurality of corpus marked with the split position to generate training data, generate rules and / or properties using the training data to create a probability model, and generate probability of sentence segmentation using the probability model. A probability distribution DB in which the index is stored, A lexical analysis module for stemming the input sentence, a syntax analysis module for revealing the syntactic structure of the input sentence, a semantic analysis module for analyzing the meaning of the input sentence, a target based on the syntax analysis and the semantic analysis A target sentence module for generating an output sentence translated into a language, and applying a classification rule stored in the classification rule DB to the input sentence to select a segmentable position, and storing the selected segmentable position in the probability distribution DB. And a central processing unit equipped with a sentence division module for determining a sentence division of the input sentence by obtaining a probability value using a probability distribution. The method of claim 5, The classification rule DB may generate a plurality of corpus in which the division positions are displayed and store learning data. The classification rule DB may be configured to learn a concept of a dividable position by using the learning data and generate and store a classification rule of a dividable position. Machine translation system. delete The method of claim 7, wherein The sentence division module further stores the output sentence in the classification rule DB as a segmentable position, modifies a probability model, and stores the output sentence in a probability distribution DB. An input step of inputting an input sentence written in a source language; A lexical analysis step of morphologically analyzing the input sentence; A syntax analysis step of revealing the syntactic structure of the input sentence, and a meaning analysis step of analyzing the meaning of the input sentence; A target sentence step of generating an output sentence translated into a target language based on the syntax analysis and the semantic analysis; In the machine translation method having an output step of outputting the output sentence, After the lexical analysis step, a rule applying step of applying a sentence dividing rule to the input sentence morphologically analyzed in the lexical analysis step, a position selection step of selecting a possible position to divide the sentence by applying the rule, and the position selection The method further includes an analysis step of obtaining a probability value through a probability distribution of and a decision step of determining sentence division based on the probability value. The probability distribution may be generated by constructing a plurality of corpus marked with the split position to generate training data, generating a rule and / or characteristic using the training data to create a probability model, and using the probability model. Machine translation method characterized in that. The method of claim 9, In the parsing step, the sentence is divided into segments by the sentence division according to the determining step, and the segment is synthesized, thereby synthesizing the syntactic structure of the input sentence. The method of claim 10, The sentence division rule applied in the rule applying step may be generated by constructing a plurality of corpus in which the division positions are displayed to generate learning data, and using the learning data to learn a concept of a segmentable position. Machine translation method. delete The method according to any one of claims 9 to 12, After the analysis step, further has a storage step of storing the probability value obtained in the analysis step, The determining step is a machine translation method, characterized in that for determining the sentence division to the position having the maximum probability value of the stored probability value.