KR20000019194A - K-best part-of-speech tagging apparatus and method based on statistics - Google Patents

Abstract

PURPOSE: A K-best part-of-speech tagging apparatus and method based on a statistics is provided, which performs a part-of-speech tagging using a state-based part-of-speech tagging. CONSTITUTION: The K-best part-of-speech tagging apparatus comprises: a state-based part-of-speech tag (100) for performing a part-of-speech tagging using an instant optimum reference to each character in a phrase; a path-based part-of-speech tag (102) for performing a part-of-speech tagging using a Viterbi algorithm according to a global optimum reference to each character in a phrase; and a post-processing unit (104) for combining the part-of-speech tagging result of the state-based part-of-speech tag (100) and the part-of-speech tagging result of the path-based part-of-speech tag (102). Thereby, it is possible to decrease the cost and effort required to the construction of a part-of-speech tagging rule.

Description

Statistics-based K-Best POS tagging device and method

The present invention relates to part-of-speech tagging, and more particularly, to an apparatus and method for statistical-based K-best part-of-speech tagging.

In language processing, part-of-speech tagging, which solves the lexical significance of each word used in a sentence, is a process of determining the correct part-of-speech for each word by inputting morphological analysis results. Essential for language processing such as document summaries, machine translation, information retrieval, and dictionary compilation.

Recently, in order to improve the accuracy of the part-of-speech tagging, a part-of-speech tag with over 97% to 99% has been proposed in English, and a part-of-speech tag with 95-97% in Korean has been proposed. The part-of-speech tagging system's role in natural language processing systems is to reduce unnecessary parsing of each word, reducing the load of parsing that requires a lot of processing time, and supporting correct sentence analysis. Therefore, the part-of-speech tagging system with a high sentence-level accuracy is needed for correct sentence analysis. Most parts-of-speech tagging systems have been evaluated in terms of words, which is not good when the part-of-speech tagging system is used for actual syntax analysis. Does not show performance. For example, suppose you are analyzing a 100-word sentence with 20 words using a part-of-speech tagging system with a high accuracy rate of 99%. In this case, it can be considered that an error about 20 words occurs after tag-of-speech tagging. In the worst case, if an error about 20 words is distributed once in 20 sentences, sentence analysis for 20 sentences fails and 80% ( Only the accuracy of 80 sentences / 100 sentences * 100) can be obtained.

In natural language, there are words with lexical significance that cannot be solved only by information at the lexical level. However, part-of-speech tagging systems use parts of speech information such as statistical information of words or word form or part-of-speech, and assign parts of speech to each word by force. This causes errors in part-of-speech tagging. Therefore, for accurate sentence analysis, rather than assigning a part-of-speech to words that are difficult to resolve lexicality, it is desirable to suspend them to the parsing stage to obtain information that can resolve the importance. In this case, after the part-of-speech tagging, the average weight of a word may increase, thereby increasing the load of parsing. Therefore, all words in a sentence should include the correct part-of-speech while minimizing this.

If parts of speech tagging are difficult to resolve completely, the part-of-speech tag that assigns two or more candidate parts of speech is called K-best part-of-speech tagging. In this case, K refers to the average number of parts of speech assigned to a word after part-of-speech tagging. K-best part-of-speech tagging, which performs part-of-speech tagging with an average weight of 1 or more per word, resolves the problem of sentence analysis by enabling the use of information necessary for resolving neutrality, such as parsing. The accuracy rate can be improved. In addition, when used as a preprocessor for constructing a part-of-speech tagged corpus, high accuracy can be guaranteed with a small amount of manual work. If you presume tagging a 10-word word corpus with a part-of-speech tagger with 95% accuracy, and then postprocess the result to construct a part-tagged tagged corpus, you get a half-word error equal to 5%. In order to revise, 10 million words must be examined and corrected. However, if K-best part-of-speech tagging is used to obtain an accuracy rate of 99% or more, it can postprocess only words with one or more parts-of-speech to construct a part-of-speech tagged corpus with an accuracy rate of 99% or more.

K-best part-of-speech tagging requires a way to remove parts of speech that a word can have that are inappropriate for the current context or to select only the parts of speech that are appropriate. Rule-based part-of-speech tagging mainly uses negation rules or linuistic constraints to eliminate inappropriate parts of speech. On the other hand, the statistical-based part-of-speech tagging method mainly performs the K-best part-of-speech tagging by calculating the probability value of the part-of-speech sequence and selecting only the parts of speech having a probability value above a certain threshold. Therefore, the average neutrality and the accuracy rate per word change according to the threshold value used, and it is very important to determine the threshold value that can have the minimum neutrality per word while showing the maximum accuracy rate.

Representative researches on K-best part-of-speech tagging include the work of Marcken, Weishedel, Brill, and Voutilainen. Marcken is a variation of DeRose's algorithm. The sentence is read from left to right, with the kth part of speech of the current word, w ⁱ w _k ⁱ Determine the best route to. Next, each path is expanded to each part of w ^{i + 1} to calculate a new probability value. At this time, if the maximum probability value of P d, and P values and determine the path already part of speech of w ⁱ included in up to w ^{i + 1} with a probability value of less than a given threshold as part of speech of w ^i. According to the given threshold, one part of speech or more parts of speech is assigned to each word. If the P value and the threshold are the same, only one part of speech is assigned to each word. Marcken experimented with an 80,000 word size corpus with an average weight per word of 2.2. The average weight per word was 1.03, which was about 97.6% accurate, and 1.27 was 99.93%.

Weishedel uses a forward-backward algorithm to observe the input sentence and to determine the specific part of speech of the current word w ⁱ . w _k ⁱ We propose a K-best part-of-speech tagger that calculates the probability value to be located at and selects the part-of-speech having a probability value below a certain threshold by comparing with the highest probability value. The Weishedel method considers the whole sentence differently from Marcken's method, which uses only the information of the previous word of the current word when calculating the probability for part-of-speech selection. Weishedel's K-best part-of-speech tagger, when tested in a corpus of 10,000 words, had an accuracy of 99.3% when the mean weight per word was 3, which was somewhat higher when evaluated based on the same accuracy as Marcken's method. To produce a result with

Brill modifies a variation rule such as 'change a part-of-speech to a different part-of-speech in a specific context', such as 'add another part-of-speech to a specific part-of-speech (or word) in a specific context', and uses it to tag K-best parts-of-speech. Suggested. When we assigned 1.04 parts-of-speech per word using 100 transformation rules, it showed 97.4% accuracy, and when we assigned an average of 1.50 parts-of-speech per word using 250 transformation rules, it showed 99.1% accuracy. Brill's method has the advantage of being able to perform part-of-speech tagging with initial taggers and automatically extract the transformation rules to be applied for error correction from the corpus, but with the same generality as hand-built rules and compatibility with other corpus than the learning corpus. The problem with is pointed out.

Voutilainen's ENGCG2 adds and refines neutralization rules to ENGCG. ENGCG2 reduces the importance of words by eliminating inappropriate parts of speech for each word using rules that describe the context in which a particular part of speech can be removed, thereby remaining neutrality for words that can no longer be applied. The word has an average gravity of one or more parts of speech. ENGCG2 has an accuracy of 99.7% when 4,000 rules are assigned on an average of 1.04 to 1.08 parts of speech per word. ENGCG2 reported the best performance compared to other systems described above. In addition, the rule-based part-of-speech tagging method overcomes the limitations of the statistics-based part-of-speech tagging method and claims that the rule-based approach can have good performance for tagging parts like other natural language processing steps. However, although ENGCG2 has a high accuracy rate, the problem of rule acquisition bottlenecks for neutrality resolution, rule conflict as the number of rules increases, and all the problems of rule-based approach in system expansion and change Have

The technical problem to be achieved by the present invention is to overcome the problems of the above-described ENGCG2, by using a part-part tagging method using both the state-based part-of-speech tagging method and the base part-of-speech tagging method, it has a high performance, such as rule-based parts-of-speech The present invention provides an apparatus and method for statistical-based K-best part-of-speech tagging that is easier to implement, extend, and manage than a tagger.

1 is a block diagram of a statistics-based K-best part-of-speech tagging apparatus according to the present invention.

2 is a flowchart illustrating a statistical-based K-best part-of-speech tagging method according to the present invention.

In order to achieve the above object, the statistics-based K-best part-of-speech tagging device according to the present invention,

State-based part-of-speech tagging that performs part-of-speech tagging for each word in a stemmed sentence from a primitive corpus using a predetermined state-based part-of-speech tagging method, and a predetermined path-based part-of-speech tagging method for each word in a stemmed input sentence. Path-based part-of-speech tagging and state-based part-of-speech tagging results and path-based part-of-speech tagged results that perform part-of-speech tagging, removes inappropriate part-of-speech of specific words to which two parts of speech are assigned from the merged result And a post-processing unit for correcting an error tagging result to obtain a part-of-speech tagged corpus even if one part-of-speech is assigned.

In order to achieve the above another object, a method for statistical-based k-best part-of-speech tagging according to the present invention, which obtains a part-of-speech tagged corpus from a raw corpus using a statistical-based approach,

Merging step of merging two parts-of-speech tagging results, each part tagging step that performs tagging according to the state-based part-of-speech tagging method and the path-based part-of-speech tagging method belonging to the statistic-based approach in the input sentence stemmed from the raw corpus And a post-processing step of removing an inappropriate part of speech of a specific word assigned two parts of speech from the merged result or assigning an appropriate part of speech, correcting an error tagging result even if one part of speech is assigned. It is characterized by.

Hereinafter, with reference to the accompanying drawings, the statistical K-best part-of-speech tagging apparatus and method according to the present invention will be described.

The present invention proposes a statistics-based K-best part-of-speech tagging system that can have a high sentence-precision rate while minimizing the mean weight per word. In the proposed part-of-speech tagging method, the part-of-speech tagging is performed using instantaneous optimality criterior and global optimality criterior, respectively. We use the method of assigning both parts of speech to the word in which the conflict occurs. In addition, the results based on the statistic-based model eliminate heuristics remaining in the words by heuristic rules or are assigned the correct parts of speech.

First, to aid understanding of the present invention, a state-based part-of-speech tagging method according to the instantaneous optimal criterion applied to the present invention and a statistical-based part-of-speech tagging method according to the global optimal criterion will be described.

The present invention has been initiated for the development of a K-best part-of-speech tagger that can overcome the problems of ENGCG2 presented above and uses a high performance statistical based approach such as ENGCG2. Statistical-based part-of-speech tagger has the advantages of being relatively easy to implement, robust to various language phenomena, and easy to expand and change the system, compared to rule-based part-of-speech tagger. Therefore, the development of statistics-based K-best part-of-speech tagger with superior performance, such as ENGCG2, is of great significance.

The study of the present invention began by analyzing the errors and problems of the previously proposed K-best statistics-based part-of-speech tagging. The existing statistical K-best part-of-speech tagging method generates parts-of-speech or part-of-speech sequences with probabilities above a certain threshold calculated by instantaneous or global optimal criteria. The following describes a part-of-speech tagging method based on instantaneous optimality criteria or global optimality criteria.

The part-of-speech tagging model is defined as a problem of finding a part-of-speech sequence t ^{1, n} corresponding to each word by receiving a sentence w ^{1, n} consisting of n words. State-based part of speech tagging method is how to assign a part of speech of the k-th having the maximum probability value of calculating the probability value, γ _i (k) for generating a sentence through the respective parts of speech in the current word w ⁱ to have a tagging results As shown in Equation 1 below.

In Equation 1, α _i (k) is observed from the first word of the sentence to the i-th word, it means a probability value when w ⁱ has a k-th part-of-speech, β _i (k) is w ⁱ is a k-th part of speech , Denotes a probability value at which words from i + 1 to the end of a sentence are observed, and Pr (Ο | λ) represents a language model. That is, it means the _{α i (k) β i (} k) is the probability to generate a sentence w ^{i, n} through the k-th part of speech of w ⁱ used in the numerator of equation (1). Part-of-speech tagging using the instantaneous optimal criterion assigns the most likely parts of speech individually in each word, but disallowed transitions may occur because the parts of speech that are most likely to be not considered are not considered. For example, in "the fly have wings", "fly" can have parts of nouns and verbs, although "fly" is a verb even though "the" is used after the determiner. Part of speech may be tagged.

A method for solving this error of state based part-of-speech tagging is path-based part-of-speech tagging using surrounding context. Path-based part-of-speech tagging is defined as in Equation 2 below.

In Equation 2, T (w ^{i, n} ) represents the average tagging probability used most statistically in the combination of parts of speech t ^{i, n} that a sentence can have. The calculation of probability values in Equation 2 requires a large amount of statistical information, and in fact, it is almost impossible to obtain these information. Therefore, the following equations (3) and (4) are often used to approximate only using lexical information or short-range contextual information.

Equations 3 and 4 approximate Equation 2 using the Markov assumption that the occurrence of the current word depends only on the current part of speech and the part of the current word depends only on the part of speech of the previous word (or the first two words). will be. For the calculation of probability values in Equations 3 and 4, a Viterbi algorithm using dynamic programming is used, and the probability values of parts of speech parts can be calculated in a linear time using the Viterbi algorithm.

As described above, state-based part-of-speech tagging assigns the most likely parts of speech individually, but has the disadvantage that unacceptable state transitions can occur, and path-based part-of-speech tagging can be used as a way to overcome this disadvantage. have. However, the existing statistics-based K-best part-of-speech tagging method uses only the method of modifying path-based part-of-speech tagging or state-based part-of-speech tagging appropriately.

The state-based part-of-speech tagging method and the path-based part-of-speech tagging method produce different parts-of-speech tagging results, which are very interesting when comparing the results. Most of the part-of-speech tagging results are concentrated on the words that caused the part-of-speech tagging errors of each method, and there are many cases where the correct part-of-speech tagging results exist. This is because when the two methods assign the same part-of-speech, this is a result of both the instantaneous optimal criterion and the global optimal criterion, so that the accuracy is reliable, but otherwise the error of either result is likely. In fact, the part-of-speech tagging results of words in which both methods assign the same part-of-speech are correct results. In the case of words that produce different results, the part of speech included the correct part-of-speech. This means that the state-based part-of-speech tagger and the path-based part-of-speech tagger have complementary features to correct each other's part-of-speech tagging errors.

The statistically-based K-best part-of-speech tagging apparatus and method proposed by the present invention utilizes these complementary features to tag the input sentences with both state-based part-of-speech tagging and path-based part-of-speech tagging, and differs only in words to which different parts of speech are assigned. Post-processing, such as allocating parts of speech.

1 is a block diagram of a statistics-based K-best part-of-speech tagging apparatus according to the present invention, which includes a state-based part-of-speech tagger 100, a path-based part-of-speech tagger 102, and a post-processing unit 104.

In FIG. 1, the state-based part-of-speech tagger 100 performs part-of-speech tagging using instantaneous optimal criteria for each word in a sentence morphologically analyzed from a raw corpus. Meanwhile, the path-based part-of-speech tagger 102 similarly performs the part-of-speech tagging using the Viterbi algorithm for each word in the sentence stemmed from the raw corpus. Post-processing unit 104 merges these two parts-of-speech tagged results and post-processes the merged results to obtain a part-of-speech tagged corpus.

The role of the post-processing unit 104 is to remove non-suitable parts of speech that are assigned as two parts of speech from the result of the state-based parts of speech tagger 100 and the path-based parts of speech tagger 102 using heuristic rules, or add the appropriate parts of speech. This reduces the average neutrality per word or improves the accuracy. In the present invention, the heuristic rule used in the post-processing unit 104 preferably analyzes the results of the state-based part-of-speech tagger 100 and the path-based part-of-speech tagger 102 and analyzes the rule manually constructed and the part-of-speech tagged corpus. It consists of the rules automatically extracted syntagma (syntagma). The New Tagma Rule is a form that specifies the individual parts of speech tagging results in the New Tagma as follows.

Syntagma rule forms:

Word1 word2, ... word n => {word1 / part of speech word2 / part of speech, ... word n / part of speech}

For example, if a predetermined number of phrases having two meanings such as "because of" are defined as a new tagma rule, both 'because' and 'of' are previously specified as prepositions.

2 is a flowchart illustrating a statistical-based K-best part-of-speech tagging method according to the present invention.

Referring to FIG. 2, the operation principle of the present invention will be described in detail, and a state-based part-of-speech tag and a path-based part-of-speech tagger perform a part-of-speech tagging process and a post-processing process by a heuristic rule. Java was called Pralape in the old book. "

First, prepare a raw corpus (step 200). The input sentences are assigned the following possible parts of speech by the morpheme analyzer (which may be embedded in the state-based part-of-speech tagger and the path-based part-of-speech tagger respectively) before analyzing the part-of-speech tagging (step 202).

An input sentence Stemming results OneislandsthatwerenearJavawascalledPralapeintheoldbookPERIOD noun det (determiner) pron (pronoun) nounconj (conjection) det pronverbd (verb: past) behad (past beha) verbd (verb: present) adj (adject) adv (adverb) prep (preposition) nounverbd behadverbd verbn (verb: past perfect) nounadv prepdetadjnoun verbp

The parts of speech in bold in Table 1 represent the correct parts of speech in the current sentence, and from the morpheme results, the mean weight per word is 1.86 (26/14).

Next, state-based and path-based part-of-speech tagging are performed for each word that has been stemmed (step 204). The results of path-based part-of-speech tagging and state-based part-of-speech tagging are shown in Table 2 below. In this case, a * indicates that an incorrect part-of-speech is assigned, and a 1 after the part-of-speech result indicates that the part-of-speech is assigned by path-based part-of-speech tagging. Also, 2 after the part-of-speech result is an indication indicating that it is a state-based part-of-speech tagging result.

An input sentence Path-based part-of-speech tagging results State-based part-of-speech tagging results OneislandsthatwerenearJavawascalledPralapeintheoldbookPERIOD noun1noun1 * conj1verbd1prep1noun1 * verbd1verbn1noun1prep1det1adj1noun1.1 noun2noun2pron2verbd2 * adj2noun2 * verbd2verbn2noun2prep2det2adj2noun2.2

Looking at Table 1, the path-based part-of-speech tagging results show that 'that' and 'was', which should be part-tagged with pron and behad, were incorrectly tagged as conj and verbd, respectively. The state-based part-of-speech tagging result shows that the part-of-speech tag for the 'that' that caused the error in the path-based part-of-speech tagging result is correctly tagged with pron, whereas the part-of-speech for 'near' is incorrectly assigned as part of adj. In addition, the error for 'was' which caused an error in path-based part-of-speech tagging still exists in state-based part-of-speech tagging.

After step 204, the two parts-of-speech tagging results are merged (step 206). Part of the path-based part-of-speech tagging and state-based part-of-speech tagging results in a part-of-speech tag that is assigned the same part-of-speech. Merge. Next, post processing is performed by applying a predetermined heuristic rule to the merged result (step 208). Table 3 below shows the result of merging the two parts-of-speech tagging results and the final K-best part-of-speech tagging results processed.

An input sentence Route-based + state-based part-of-speech tagging results Final K-best part-tagging tagging results OneislandsthatwerenearJavawascalledPralapeintheoldbookPERIOD nounnounconj1 pron2verbdprep1 adj2noun * verbdverbnnounprepdetadjnoun. nounnounconj1 pron2verbdprep1 adj2nounbehadverbnnounprepdetadjnoun.

Postprocessing removes inappropriate parts of speech from the part of speech or assigns the correct parts of speech. In this example, "Cword: verb * && Next word in 2 position: verbn => change Cword: beha * Heuristic rules, such as ", change the part of 'was' to behad. This rule states that if the part of speech of the current word is verbp (verbd, verbn), and the part of speech after or after the current word is tagged with part of speech, the part of speech of the current word is modified to behap (behad, behan). It is a rule. In other words, these rules are built by hand. In Table 3, the final result after the heuristic rule is applied allocates an average of 1.14 parts-of-speech per word, with an accuracy of 100%.

Now, the statistics-based K-best part-of-speech tagging apparatus and method according to the present invention and the conventional part-of-speech tagging will be compared through examples.

The context transfer probability calculation used for the path-based part-of-speech tagging uses a bigram model that considers only the part of speech of the word before the current word and a trigram model that considers the part of speech of the previous two words. In general, part-of-speech tagging using trigrams shows high accuracy. However, since the characteristics of each method are slightly different, it is not known whether the results of using the trigram model and the state-based part-of-speech tagging results show better performance than the integration with the bigram results.

Accordingly, the present invention implements a path-based part-of-speech tagger using a Baigram model and a trigram model, respectively, and performs an experiment integrating the results of two parts-of-speech tagging and state-based part-of-speech tagging. For the raw corpus used in the experiment, 500 English sentences were used that were not used to extract probability information used for part-of-speech tagging. The total number of words in the primitive corpus used in the experiment was 6,137, with an average weight of 2.53 per word.

Table 1 below shows the experimental results of the state-based part-of-speech tagging method, the path-based part-of-speech tagging, and the part-of-speech tagging method proposed by the present invention. 'State' refers to the result-based part-of-speech tagging result, and 'Path 2' and 'Path 3' refer to the path-based part-of-speech tagging result using a bigram model and a trigram model, respectively. 'Post-processing' means applying the heuristic rule used in the present invention.

Part of Speech Tagging Method Word unit Sentence Chinese word per word condition 97.07% 70.40% One Route 2 96.43% 65.80% One Route 3 96.35% 69.80% One Route 2+ Status 98.21% 80.80% 1.06 Route 3+ Status 99.02% 88.80% 1.10 Path 2 + Status + Post Processing 98.57% 84.20% 1.06 Path 3 + Status + Post Processing 99.22% 91.40% 1.09

Experimental results show that the part-of-speech tagging method that combines path-based part-of-speech tagging and state-based part-of-speech tagging has higher accuracy in word and sentence units than in each method.

In particular, it can be seen that the performance was very good to improve the accuracy of the sentence unit. Comparing 'Path 2+ state' and 'Path 3+ state', the 'Path 3+ state' method showed a high accuracy rate, but the average weight per word was slightly higher. The heuristic rule did not have much effect on lowering the average weight per word, but was found to be effectively used to increase the accuracy. Finally, the path-based part-of-speech tagging using the trigram and the state-based part-of-speech tagging results, and postprocessing, the 'path 3 + state + post-processing' method showed the highest accuracy in word and sentence units. In particular, the proposed method showed a very good performance in improving the accuracy of sentence units.

The present invention implements the part-of-speech tagging method of Marcken and Weischedel as the part-of-speech set used by the proposed method and compares and evaluates the same by applying the same experimental corpus. The results are shown in Table 5 below.

Average weight per word Accuracy of each system Marcken Weischedel The present invention 1.36 99.41% 99.36% 1.31 99.33% 99.15% 1.29 99.07% 98.55% 1.12 98.71% 97.51% 1.09 98.45% 97.28% 99.22%

As can be seen from Table 5, Weischedel's method showed higher accuracy than Marcken's method when maintaining the same number of weights in a word after part-of-speech tagging, and the proposed method had the highest accuracy when the average weight per word was 1.09. Showed.

Currently, the proposed method does not have the ability to adjust the average weight and accuracy rate per word. This problem can be solved by applying Marcken's method or Weischedel's method to the path-based part-of-speech tagger or state-based part-of-speech tagger used in the proposed method. Simple solution. ENGCG2 is a rule-based part-of-speech tagging method, which requires a lot of manual work and expense to construct part-of-speech tagging rules, and was excluded from the above evaluation.

However, the results of Table 5 show that the proposed method can have better performance than the K-best part-of-speech tagging system using the conventional statistical based approach. It can also be seen that the use of a statistics-based approach and heuristic rules alone can yield the high performance reported by ENGCG2. The advantages of the proposed method compared to the rule-based method of ENGCG2 are as follows. First, it is easy to implement because it is possible to use a large amount of corpus recently. Second, it does not require the cost and effort to establish a large part-of-speech tagging rule. Third, it is easy to change and expand the system due to the change of part-of-speech set or application field.

As described above, the statistics-based K-best part-of-speech tagging apparatus and method according to the present invention has a similar performance to that of the rule-based part-of-speech tagger ENGCG2, which recently reported the best performance, and compared with the rule-based part-of-speech tagger. It is easy to expand and manage, and saves the cost and effort of constructing the large parts of speech tagging rules needed for developing high-precision parts of speech tagger such as ENGCG2, and applies simple heuristic rules to limit the accuracy of statistics based parts of speech tagging. There is an advantage to overcome.

Claims (6)

In the statistics-based K-best part-of-speech tagging device, State-based part-of-speech tagger that performs part-of-speech tagging for each word in a stemmed sentence from a primitive corpus using a predetermined state-based part-of-speech tagging method; A path-based part-of-speech tagger that performs part-of-speech tagging for each word in the morphologically analyzed input sentence using a predetermined path-based part-of-speech tagging method; And Merging the state-based part-of-speech tagged result with the path-based part-of-speech tagged result, removing the inappropriate part of speech of a specific word that has been assigned two parts of speech from the merged result, or assigning an appropriate part-of-speech, even if one part of speech is assigned And a post-processing unit configured to correct the tagging result to obtain a part-of-speech tagged corpus. According to claim 1, The post-processing unit, A statistical-based K-best part-of-speech tagging device, characterized by using predetermined heuristic rules. The method of claim 2, wherein the heuristic rule, Statistics-based K-Best part-of-speech comprising at least a state-based part-of-speech tagger and a rule constructed manually by analyzing the results of the path-based part-of-speech tagger and a new tagma rule automatically extracted by analyzing a part-of-speech tagged corpus. Tagging device. In a statistical-based K-best part-of-speech tagging method, which obtains a part-of-speech tagged corpus from a raw corpus using a statistics-based approach, A part-of-speech tagging step of performing tagging according to each of state-based part-of-speech tagging and path-based part-of-speech tagging methods belonging to the statistic-based approach in the input sentence morphologically analyzed from the source corpus; A merging step of merging two parts of speech tagging results; And Removing a part-of-speech part of the word that has been assigned two parts of speech from the merged result, or assigning a suitable part of speech, and correcting an error tagging result even if one part of speech is assigned, to obtain a part-of-speech tagged corpus. A K-best part-of-speech tagging method characterized by the above-mentioned. The method of claim 4, wherein the post-treatment step, A K-best part-of-speech tagging method characterized by using a predetermined heuristic rule. The method of claim 5, wherein the heuristic rule, Statistics-based K-Best part-of-speech comprising at least a state-based part-of-speech tagger and a rule constructed manually by analyzing the results of the path-based part-of-speech tagger and a new tagma rule automatically extracted by analyzing a part-of-speech tagged corpus. Tagging method.