KR20010028946A - Morphological analysis method and apparatus used in text-to-speech synthesizer - Google Patents

Abstract

PURPOSE: The method for interpreting morpheme used in a document interpreter of a voice synthesizer and an apparatus thereof are provided to enhance naturalness and clearness of a voice synthesizer by deciding kind of a sentence and kind of each paragraph composing a sentence, thereby creating a pronunciation suited thereto. CONSTITUTION: A document preprocessor comprises a sentence extracting unit(20), a user preapplying unit(24) and special pattern processor(26). The sentence extracting unit(20) receives a document and extracts the document as a token list type. The user preapplying unit(24) replaces a word row by a replacement word row corresponding to the word row if the word row stored in user dictionary database(22) is included in the token list typed document. The special pattern processor(26) decides a kind of extracted sentence and decides a kind of pattern by extracting a special pattern, and modifies a position of space. The token list typed document extracted from the document preprocessor is transmitted to a morpheme interpretation unit. The morpheme interpretation unit comprises a morpheme combination creating unit, a reserve list creating unit, a morpheme interpretation list creating unit and a morpheme error modifying unit. The morpheme combination creating unit receives the token list typed document from the document preprocessor and creates all sorts of combinations of morphemes. The morpheme interpretation list creating unit selects a reserve morpheme list having a value multiplying a connection weight value of a morpheme and a word as the last morpheme interpretation list.

Description

Morphological analysis method and apparatus used in text interpreter of speech synthesizer

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech synthesizers, and more particularly, to a morpheme analyzer and a method used in a document parser of a speech synthesizer.

The document parser of the speech synthesizer extracts a sentence from the input document, and generates a syntax tree and pronunciation information through morphological analysis and syntax analysis of the sentence. The syntax tree and pronunciation information generated by the document interpreter are input to the rhyme processing unit to influence the rhythm generation of the sentence and the selection of the composition unit.

Church's Korean Patent Application No. 89-1364 'Part-of-Speech Decision and Usage Method', which presents a part-of-speech tagging model for English, is not appropriate for the Korean language, which is a deadlock, and does not fully reflect real-world language phenomena because only statistical information is used. There is a drawback that the accuracy of part-of-speech tagging is poor.

In order to solve this problem, a method of using a rule approach in addition to a statistical approach has been proposed. The technique used in Lim Hee-seok's “Korean part-of-speech tagging model using linguistic knowledge and statistical information” uses a method of resolving or reducing ambiguity using positive and negative linguistic knowledge before using statistical information. Here, ambiguity means that there are several morphological results of a word. For words that are not resolved by linguistic knowledge, parts of speech are assigned as the result of statistical part-of-speech tagging. Finally, the correction knowledge is used to correct the result of the emergency. In this technique, negation knowledge serves to reduce the number of possible morphological interpretations, but affirmative knowledge generates only part-of-speech tagging results, and only highly accurate rules are used. Therefore, there is a problem that a rule such as' the vocabulary A of part-of-speech A is likely to be linked to the vocabulary B 'of the part-of-speech B' cannot be used. In addition, since the knowledge used in this method is meaningful only when extracted from a large amount of corpus, the knowledge extracted from a small amount of corpus is difficult to apply.

The present invention was created to solve the above problems, and can be read by a computer that records a morpheme analysis method, a morpheme interpreter, and a morpheme analysis program to enable a document interpreter to generate correct pronunciation information through morpheme analysis and part-of-speech tagging. Its purpose is to provide a record carrier.

1 is a block diagram showing the overall configuration of a morpheme analyzer according to the present invention.

FIG. 2 is a block diagram illustrating a detailed configuration of the document preprocessor of FIG. 1.

3 is a block diagram showing a detailed configuration of a morpheme analysis unit in FIG. 1.

4 is a flowchart illustrating a document preprocessing process for morpheme analysis according to the present invention.

FIG. 5 is a flowchart showing the detailed pattern processing step of FIG. 4 in detail.

6 is a flowchart illustrating a morpheme analysis process according to the present invention.

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

1, the morpheme analyzer used in the document interpreter of the Korean speech synthesizer according to the present invention includes a document preprocessor 10 and a morpheme analyzer 12.

According to FIG. 2, in the morpheme analyzer according to the present invention, the document preprocessor 10 includes a sentence extractor 20, a user dictionary application unit 24, and a special pattern processor 26. The sentence extraction unit 20 receives a given document and extracts the sentence in the form of a token list. The user dictionary application unit 24 converts the word string stored in the user dictionary database 22 prepared in advance into a substitution string corresponding to the word string when the word string is included in the sentence in the form of a token list. The special pattern processing unit 26 determines the type of the extracted sentence, extracts a special pattern of a predefined type in the sentence, determines the type of the pattern corresponding thereto, and corrects the position of the blank for natural break reading. . The sentence in the form of a token list extracted by the document preprocessor 10 is transferred to the morpheme interpreter 12.

According to FIG. 3, in the morpheme analyzer according to the present invention, the morpheme analysis unit 12 includes the morpheme combination generation unit 30, the candidate list generation unit 32, the morpheme analysis list generation unit 34, and the morpheme error correction unit 36. It is configured to include).

The morpheme combination generator 30 receives a sentence in the form of a token list processed with a special pattern from the document preprocessor 10 and generates all possible morpheme combinations for each token using a morpheme dictionary. The candidate morpheme list generator 32 extracts candidate morpheme lists for each word using left and right connection information by the longest matching method among all possible morpheme combinations. The morpheme analysis list generator 34 multiplies the link weights, which are weights multiplied by the concatenation of the morphemes and the parts of speech, according to the words, the context probability, the lexical probability, and the probability that the morpheme analysis yields a correct result. The best candidate morphological list is selected as the final morphological interpretation list. The morpheme error corrector 36 corrects errors that may occur in the final morpheme interpretation list using knowledge of morphological analysis correction.

Hereinafter, the operation process of the present invention will be described in detail.

As shown in FIG. 4, the document preprocessing process proposed by the present invention includes a sentence extraction step (400), a user dictionary application step (410), and a special pattern processing step (420).

First, a sentence is extracted from an input document by using the method of the prior application, Korean Patent Application No. 97-24231 (context extraction using contextual information and regional document form), and the extracted sentence is a character constituting a blank position and a word. Tokens are divided into tokens according to the type of the token list (step 400). Here, the token is divided by the type of characters that constitute a word. Types of tokens classified in the present invention are classified into Korean, Chinese, symbols, numbers, and foreign languages according to the types of characters constituting the token.

When the string registered in the user dictionary database 22 is in a sentence in the form of a token list, the string is replaced with the replacement string in the user dictionary. An example of the configuration of the user dictionary database 22 is as follows.

Co. Company

H / W hardware

Monday

Mi-Sun Mi-Sun

… …

If the raw string in the first entry of the user dictionary database 22 appears in a token list form sentence, replace that string with the destination string appearing in the second entry. That is, when the string '(month)' appears in the token list form sentence, the token corresponding to '(month)' is deleted and the token corresponding to 'Monday' is inserted at the position. Therefore, the voice synthesizer synthesizes a voice called 'Monday' when the string '(month)' appears in a sentence in the form of a token list. You can also use the user dictionary database to correct misuse of punctuation. If there is a ‘미. 일’ in a sentence in the form of a token list, the sentence is replaced with ‘미 · 일’ without ‘.’.

A detailed process of the special pattern processing step 420 of FIG. 4 is illustrated in FIG. 5. The sentence type is determined for the sentence in the form of a token list to which the user dictionary is applied (step 500), the position of the special pattern of the predefined type is estimated (step 510), and the type of the pattern is determined accordingly. According to the type of word, the type of word that each token is to be configured is determined, and the position of the blank is corrected (steps 520 to 540).

Types of sentences distinguished by the present invention are general sentences, title sentences, enumerated sentences, English sentences, and name sentences. The title sentence is a sentence corresponding to the title of the document and is determined by the form of the sentence before and after. An enumerated sentence is a sentence in which similar kinds of words are listed, such as '가나다 라마 바사'. An English sentence refers to a sentence composed of English, and a name sentence refers to a Korean name constituting a sentence. Name sentences are common sentences in texts such as novels and essays. This kind of sentence influences the method of morpheme interpretation and pronunciation. In the case of English sentences, punctuation, symbols, and numbers should be generated in English, and in case of name sentences, consonant stories should not occur between syllables of names. The enumerated sentences should be read one word at a time. In this way, the type of sentence affects the pronunciation generation. How the morpheme interpretation method differs according to the type of sentence will be described with reference to FIG. 6.

When the type of sentence is determined, a special pattern of a predefined type is extracted in the sentence (steps 510 to 540). The special pattern extracted in the present invention is a formula, an email address, a URL, a file name, a repeating pattern, and the like. These special patterns are mainly English / numeric / symbols. Therefore, if the English / numeric / symbol token node continues without a space in the token list form sentence, it is considered that there is a possibility of a special pattern (step 510).

If there is a possibility that the pattern is a special pattern, it is first determined whether the pattern is a mathematical pattern (step 520). The formula pattern is a pattern consisting of numbers and mathematical symbols such as '3 * (4 + 3) = 21'. For the token nodes determined as the mathematical pattern, the word type is 'formula', and the formula symbol and the numbers Insert spaces in the text to produce pronunciation and natural break readings.

If it is not a mathematical pattern, it is determined whether the e-mail address, a Uniform Resource Locator (URL), or a file name is processed (step 530). The email address is followed by the user identifier (uliphin) and the "@" symbol, such as "uliphin@sait.samsung.co.kr," followed by the domain name (sait.samsung.co.kr). In this case, insert a space where the specific symbol appears, and set the word type to 'e-mail address' to generate proper pronunciation for the symbol that appears in the pattern. For the URL, the protocol (http) and the symbol ': //' appear, such as 'http://www.samsung.co.kr,' followed by the site's location (www.samsung.co.kr). In the case of URLs, the e-mail address and the symbol reading method are the same, so the word type is 'e-mail address' and a space is inserted for each symbol. The file name is a directory name (c) followed by the symbol ': \', followed by the file name (config) and the symbol '.' Followed by an extension (sys), such as 'c: \ config.sys'. . In this case, the directory name may be omitted. In this case, the email address and the symbol reading method are the same, so the pattern type is 'e-mail address', and a space is inserted after the specific symbol.

If it is not an email address / URL / filename pattern, it checks whether the same type of string is a repeating pattern without spaces, and if there is a repeating pattern, inserts a space between repeated patterns to create a natural break. (Step 540).

If a pattern that is likely to be a special pattern is not processed in steps 520 to 540, a word that reads the word type “letter by letter” is read so that it is judged as a meaningless pattern and the numbers and symbols appearing in the pattern can be read one by one. ' For example, a pattern such as '344-06-087488' is a pattern that requires reading each number one letter.

The token list of sentences that have undergone a special pattern processing step in the document preprocessor 10 is transferred to the morpheme interpreter 12 as shown in FIG. 1. Morphological analysis in the present invention is a general term for morphological analysis and avalanche logging. The detailed operation of the morpheme analysis unit 12 is shown in FIG. 6.

In the present invention, the method of morpheme interpretation varies according to the type of sentence indicated by the input token list (step 600). When the type of sentence is a special sentence such as a name sentence or an enumerated sentence, the morpheme analysis is performed based on predefined information (step 680). For example, in the case of a name sentence, since the sentence itself consists of one person's name, the morpheme part of speech becomes a proper noun.

For the title / general / English sentence, the method of morphological interpretation of the token varies according to the type of token constituting the sentence (step 610). In operation 640, a single candidate stem list is generated for numbers, symbols, and foreign language tokens. Stemmed parts of numbers and foreign language tokens become numbers and foreign languages, respectively, and symbol tokens are assigned predefined stem parts according to their types.

For the Hangul and Hanja tokens, all possible stemming combinations are made using the stemming dictionary database 625 (step 620). In the present invention, in order to reduce the number of morpheme nodes generated, the morpheme parts of speech which can be determined only by the front / back context among the possible morpheme parts of speech are merged into one to reduce the number of morpheme nodes generated.

In the case of ancillary linkages and equivalent / slave linkings, it is determined by the parts of the morphemes that follow, rather than by probability, so it is integrated into one node. Of these two parts-of-speech, the part-of-speech that is representative is taken as the part of speech of the morpheme, and the detail part of speech is used to express the integrated part-of-speech. For example, the word ‘-고’ in ‘하는 하다’ has parts of speech that are both parallel and auxiliary. A part-of-speech of '-고' is an auxiliary linking end when a part of the following morpheme is a linkable verb, but otherwise, it is an equivalent linking end. After all, whether the parts of morphemes are equivalent linkages or auxiliary linkings is determined by the context of the front and back, and not stochastic. Accordingly, the two parts-of-speech are merged into one morpheme to generate a morpheme list, and the integrated part-of-speech is selected in the step 670 of modifying the morpheme analysis result using a rule, which is the final step of morpheme analysis.

In operation 620, a plurality of candidate morpheme lists are extracted from the morpheme combination generated in operation 620. This is to avoid calculating the probabilities for all possible cases for real time processing. In the present invention, it is preferable to generate a candidate morpheme list by the longest matching method. At this time, the extended left and right access information database 635 is used, and when generating candidate morpheme lists, lists of similar types are combined and processed.

Existing left and right connection information is related only to the morpheme parts of speech, but the extended left and right connection information used in the present invention includes lexical information in addition to the information on the connection possibility between the morpheme parts of speech. Vocabulary information is typically part of speech that is not connectable, but for a particular vocabulary, information is usually included. For example, verb verb suffixes generally follow only nouns. Therefore, the existing left and right access information has a type of access information called <noun + verb suffix>. However, some verbalization suffixes may follow adverbs or endings. Verbization suffixes such as stri / trida can follow the linking ‘아 / 어’ and verbs suffixes such as ‘다다 / distance 다’ may follow an adverb. The extended left-right connection information proposed in the present invention includes lexical information such as <a/word* link + drop/trida* verb suffix> and <adverb + put / distance * verb suffix>.

In the present invention, a speech synthesizer is semantically different from the case in which one noun is finally generated as the same morphological interpretation list as in the case of <noun + noun suffix> and the noun is divided into several nouns. If possible, the candidate morphological lists are merged into one if they can be integrated without affecting the reading and pronunciation generation.

Informatization * nc.2 + * jxc.0 ----- (1)

Info * nc.0 + sum * xn.0 + * jxc.0 ----- (2)

In the case of (1) and (2) in the list of candidate morphemes corresponding to 'informatization', the noun 'informatization' in list (1) is shown in the (2) list of <information * noun> and <hwa * noun The suffixes> are separated. After all, list (2) becomes the same list as (1) after noun derivation, so only list (1) is created when creating candidate morpheme list.

Info * nc.0 + * jxc.0 ----- (3)

* Nc.0 + bo * nc.2 + * jxc.0 ----- (4)

In the case of (3) and (4) among candidate morphological lists corresponding to 'information', 'information' in list (3) is expressed as 'jeong' and 'beam' in list (4) . Although the list is semantically different in terms of structure, both lists are in the form of <noun + survey>, which does not affect reading, and the same pronunciation is generated when the pronunciation is generated. Thus, integrating the two lists into one does not affect the naturalness and clarity of the speech synthesizer. Therefore, only list (3) is generated as a candidate stem list.

The most likely list of candidate morpheme lists generated in step 630 is determined as the final morpheme analysis list (step 650). In the present invention, to generate the final morphological analysis list, the context probability, the lexical probability, and the connection weight for the candidate morpheme list are extracted from the context probability / lexical probability / connection weight database 655. The final morphological analysis list is the candidate morphological list in which the product of the extracted context probabilities, lexical probabilities, and link weights is the best. Context Probability includes intra-word context probability and inter-word context probability. In-word context probability is the probability of linking two previous morpheme parts of speech with the current morpheme part-of-speech, and the inter-word context probability is the probability of transitioning from the last morpheme part of previous phrase to the first morpheme part of the current word. Vocabulary probabilities are the probability of having a current morpheme in a given current morpheme part of speech. Link weights are weights that multiply the linking of morphemes and parts of speech that are likely to be linked so that morphological analysis can produce the correct results, and there are link enhancement weights and link weakening weights. This weight is determined by the connection weight rule. The connection weight rule has the following form.

<(Old morpheme) * (old morpheme) (consolidated) (current morpheme) * (current morpheme) => Weight ＞

'+' Means morphological links within words, and '/' means morphological links between words. For example, <? * Side fire investigation /? * The dependent weighting rule of dependent verb => 10 ＞ is that if the last morpheme part of the previous word is a subtest, and the current morpheme, that is, the first morpheme part of the current word, is a dependent verb, the context probability is multiplied by a weight of 10 to select this morphological list. Increase your chances of becoming. The connection weight can solve the problem that occurs when the general context probability only considers the part of speech, and it can correct the case where the context probability differs from the actual language phenomenon due to the small size of the corpus.

For general / title / English sentences, steps 610 to 650 are repeatedly performed for all tokens included in the token list type sentence (step 660).

Finally, the morpheme analysis result generated in step 650 is corrected using the morphological analysis correction knowledge 675 (step 670). This is a process for correcting errors that occur when only statistical information is used. Morphological analysis correction knowledge 675 includes knowledge about the parts that frequently cause errors.

Meanwhile, the above-described embodiments of the present invention can be written as a program that can be executed on a computer. And, it can be implemented in a general-purpose digital computer for operating the program using a medium used in the computer. The media may be stored such as magnetic storage media (e.g., ROM, floppy disk, hard disk, etc.), optical reading media (e.g., CD-ROM, DVD, etc.) and carrier waves (e.g., transmission over the Internet). Media.

The recording medium is a module for receiving a given document to extract a sentence in the form of a token list; A module for retrieving a corresponding target string from the user dictionary database and replacing the word string in the sentence in the token list form when the raw string extracted from the user dictionary database is included in the sentence in the token list form; A module for determining a type of a sentence in the form of a token list to which a user dictionary is applied, extracting a predetermined type of special patterns in the sentence, and correcting a predetermined blank position according to the type of the special pattern; And a module for interpreting the morphemes constituting the sentence in the form of a token list processed with a special pattern, and performing a part-of-speech tagging to generate word-level morpheme interpretation lists as program code that can be executed on a computer.

The module for generating a word stem morpheme interpretation list may include a module for generating all possible morpheme combinations using a morpheme dictionary for each token constituting a token list processed with a special pattern; A module for extracting candidate morphological lists for each word using left and right access information from all the possible morpheme combinations; From the candidate morpheme lists of words, the candidate morpheme list is obtained by multiplying the concatenation weight, which is the weight multiplied by the concatenation of morphemes and parts of speech, according to the probability that the context probability, the lexical probability, and the morpheme analysis yield the correct result. A module for selecting from the analysis list; And program code for executing a module to correct errors that may occur in the final morphological interpretation list using morphological analysis correction knowledge.

Functional modules for implementing the present invention as described above can be easily implemented by programmers in the art to which the present invention belongs.

According to the present invention, first, since the type of a given sentence and the type of each word constituting the sentence can be determined to generate a sound corresponding to the sentence, the naturalness and clarity of the speech synthesizer are increased.

Second, by inserting appropriate spaces for special patterns where numbers / symbols / English are repeated without spaces, the speech synthesizer can be naturally read.

Third, the candidate morpheme analysis list is increased by generating the candidate morpheme list using the extended left and right access information, and the candidate morpheme list is optimized by integrating the similar parts of speech and the candidate candidate morpheme list. Reduce time.

Fourth, in addition to lexical probabilities and context probabilities, link weighting rules can be used to improve the accuracy of morphological analysis even if a small amount of corpus is used.

Claims (11)

In the morpheme interpreter used in the text interpreter of the speech synthesizer, A sentence extracting unit which receives a given document and extracts a sentence in the form of a token list; And a user dictionary database composed of a plurality of source strings and a plurality of object strings corresponding thereto, and when the source string extracted from the user dictionary database is included in the token list form sentence, A user dictionary application unit for retrieving and replacing the word string in a sentence of the token list form; And a special pattern processing unit that determines a type of a sentence in the form of a token list to which the user dictionary is applied, extracts a predetermined type of special patterns within the sentence, and corrects a predetermined blank position according to the type of the special pattern. Pretreatment unit; And And a morpheme analysis unit configured to receive a sentence in the form of a token list processed with a special pattern from the document preprocessing unit, and interpret morphemes and tagging parts of speech to generate word-level morphological analysis lists. The method of claim 1, wherein the morpheme analysis unit A morpheme combination generation unit that receives a sentence in the form of a token list processed with a special pattern from the document preprocessor and generates all possible morpheme combinations for each token using a morpheme dictionary; A candidate morpheme list generator for extracting candidate morpheme lists for each word using left and right access information by the longest matching method among all possible morpheme combinations; From the candidate morpheme lists of words, the candidate morpheme list is obtained by multiplying the concatenation weight, which is the weight multiplied by the concatenation of morphemes and parts of speech, according to the probability that the context probability, the lexical probability, and the morpheme analysis yield the correct result. A morpheme analysis list generation unit for selecting the analysis list; And A morphological analyzer comprising a morphological error correction unit for correcting errors that may occur in the final morphological analysis list by using morphological analysis correction knowledge. The method of claim 2, wherein the left and right connection information used in the candidate stem list generator is A morpheme interpreter comprising information on linkability between morpheme parts of speech and lexical information about parts of speech that are generally non-connectable parts of speech but that can be linked to a predetermined vocabulary. In the morpheme analysis method used in the text interpreter of the speech synthesizer, (a) receiving a given document and extracting a sentence in the form of a token list; (b) if the raw string extracted from the user dictionary database is included in the token list form sentence, fetching a corresponding object string from the user dictionary database to replace the word string in the sentence in the token list form; (c) determining a kind of a sentence in the form of a token list to which the user dictionary is applied, extracting a predetermined type of special patterns in the sentence, and correcting a predetermined blank position according to the type of the special pattern; And and (d) interpreting the morphemes constituting the sentence in the form of a token list processed with the special pattern and tagging the parts of speech to generate word-word morpheme interpretation lists. The method of claim 4, wherein the type of sentence in step (c) A morpheme interpretation method comprising a title sentence corresponding to a title of a document, an enumerated sentence listing similar kinds of words, an English sentence in English, a name sentence composed of Korean names, and a general sentence other than the above. The method of claim 4, wherein in the step (c), A morphological analysis method comprising: a mathematical pattern consisting of numbers and mathematical symbols, an email address pattern consisting of an email address, a sentence consisting of a URL or a file name, and a repeating pattern in which a string of the same type is repeated without a space. The method of claim 4, wherein step (d) (d1) generating all possible morpheme combinations using the stemming dictionary for each token constituting the token list processed with the special pattern; (d2) extracting candidate morphological lists for each word using left and right access information in all possible morpheme combinations; (d3) a candidate morphological list of which the product of the concatenation weight, which is the weight multiplied by the concatenation of the morphemes and parts of speech, of the candidate morpheme lists among the candidate morpheme lists according to the phrases, is the highest. Selecting as the final morphological analysis list; And and (d4) correcting errors that may occur in the final morphological analysis list using morphological analysis correction knowledge. The method of claim 7, wherein the left and right connection information used in the step (d2) is A morpheme analysis method comprising information on linkability between morpheme parts of speech and lexical information about parts of speech that are generally unconnectable parts of speech but that can be linked to a predetermined vocabulary. 8. The method of claim 7, wherein when the sentence in the form of a token list processed with the special pattern is one of a name and an enumerated sentence, the morphological analysis corresponding to the type of the sentence is performed instead of the steps (d1) to (d4). And generating a word-by-word morpheme analysis list. In a computer-readable recording medium recording a morphological analysis program used in a text interpreter of a speech synthesizer, (a) a module for receiving a given document and extracting a sentence in the form of a token list; (b) a module for retrieving a corresponding target string from the user dictionary database and replacing the word string in the sentence in the token list form when the raw string extracted from the user dictionary database is included in the sentence in the token list form; (c) a module for determining a type of a sentence in the form of a token list to which the user dictionary is applied, extracting a predetermined type of special patterns in the sentence, and correcting a predetermined blank position according to the type of the special pattern; And and (d) a module for interpreting the morphemes constituting the sentence in the form of a token list processed with a special pattern and tagging the parts of speech to generate word-level morphological interpretation lists. Record carrier. The method of claim 10, wherein the module (d) (d1) a module for generating all possible morpheme combinations using the morpheme dictionary for each token constituting the token list processed with the special pattern; (d2) a module for extracting candidate morphological lists for each word using left and right access information in all possible morpheme combinations; (d3) a candidate morpheme list having the highest product value obtained by multiplying a concatenation weight, which is a weight multiplied by a concatenation of morphemes and parts of speech, according to the probability that the context probability, the lexical probability, and the morpheme analysis yield a correct result. A module for selecting the final morphological analysis list; And (d4) A computer-readable recording medium having recorded thereon a morphological analysis program, comprising a module for correcting errors that may occur in the final morphological analysis list using morphological analysis correction knowledge.