KR20010107111A - Natural Language Question-Answering System for Integrated Access to Database, FAQ, and Web Site - Google Patents

Abstract

Unlike a search system that indexes documents on a website and presents the searched results, the present invention distributes a user's query to each area by a search system integrating FAQ list search and interactive DB search as well as document processing on a website. An integrated query-response search system and method capable of optimally searched response processing are provided. The present invention comprises the steps of (i) morphological / partial parsing of the input natural language query; (Ii) determining whether the analyzed query is a query suitable for the interactive DB search method and searching for a response using the interactive DB search method if the query is suitable for the interactive DB search method; (i) not suitable for the interactive DB search method; Applying bottom-up natural language information analysis to the query; (V) determining whether the query analyzed by the bottom-up natural language information analysis is a query suitable for the FAQ list retrieval method and retrieving the response using the FAQ list retrieval method; Determining whether or not the user requests a website search and searching for a response using the website query-response search method; (i) integrating the responses obtained from the interactive DB search, FAQ list search, and website search. An integrated query-response search method comprising the step of selecting a response.

Description

Natural Language Question-Answering System for Integrated Access to Database, FAQ, and Web Site}

The present invention relates to an integrated question and answer search system for searching all interactive DBs, Frequently Asked Qusetion (FAQ) lists, and websites. In particular, the present invention is a query that can freely search for desired information by processing all searches as natural language question-answers. -The response system and method.

Recently, an information retrieval system that processes a myriad of web document information on a website, extracts only information corresponding to a user's request, and provides the information to a user is widely used. In addition, an Internet information retrieval system capable of searching the Internet even in everyday natural language rather than a word format is widely provided. However, in general, it is very difficult to accurately extract the information requested by the information requester from a large collection of web documents, and even though a natural language search is possible, there are many shortcomings for the current search system to correctly obtain a response to a specific query desired by the information requester. .

In order to solve the above problems, an object of the present invention is a search system that integrates a FAQ list search and an interactive DB search as well as document processing on a website, unlike a search system that indexes only documents on a website and presents searched results. It is to provide an integrated query-response search system and method capable of processing an optimally searched response by distributing a user's query to each area.

In the present invention, a bottom-up technique such as statistical information retrieval technique and statistical language analysis, and a top-down technique such as language knowledge base and question-response modeling are integrated and used.

1 is a structural diagram of an interactive DB, a FAQ list, and a website search system according to the present invention;

2 is an integrated query-response search system for searching all interactive DBs, FAQ lists, and websites.

3 is a method flow diagram of an integrated query-response retrieval system.

4 is a flow chart of a search phase of an interactive DB search system.

5 is a flowchart of a search step of the FAQ list search system.

6 is a flow chart of a search phase of a website query-response system.

In order to achieve the above object, the integrated question-answer search system according to an embodiment of the present invention

(Iii) an interactive DB search system for searching an interactive DB;

(Ii) a FAQ list retrieval system for retrieving a list of answers to frequently asked questions due to queries and user requests not found in the interactive DB retrieval system;

(Iii) a website query-response system that executes a query-response search on a website in response to a user's request and a query not found in the interactive DB search system or the FAQ list search system.

In order to achieve the above object, the integrated question-answer search method according to another embodiment of the present invention

Input natural language query

(1) morpheme / partial parsing the input natural language query; (Ii) determining whether the analyzed query is a query suitable for the interactive DB search method and searching for a response using the interactive DB search method if the query is suitable for the interactive DB search method; (i) not suitable for the interactive DB search method; Applying bottom-up natural language information analysis to the query; (Iii) determining whether the query analyzed by the bottom-up natural language information analysis is a query suitable for the FAQ list retrieval method and retrieving the response using the FAQ list retrieval method if the query is suitable for the FAQ list retrieval method; And (v) searching for a response using a website query-response search method by determining whether the user does not conform to the FAQ list search method or if the user requires a website search; (i) the interactive DB search, a FAQ list search, Incorporating the responses from the website search to select the best response.

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

1 is an overall structure diagram of an integrated natural language query-response search system for analyzing a user's natural language query, searching for an interactive DB, a list of FAQs, and a website to generate a response. As shown, the integrated natural language query-response system includes an interactive DB search system 10 for searching an interactive DB; A FAQ list retrieval system 20 for retrieving a list of answers to frequently asked questions due to queries and user requests not found in the interactive DB retrieval system 10; A web site query-response system 30 that executes a query-response search on a web site in response to a user's request and a query not found in the interactive DB search system 10 or the FAQ list search system 20.

The interactive DB retrieval system includes: a morpheme / syntax analyzer 110 for determining a mathematical relationship between a morpheme analysis for analyzing a user's natural language query as a morpheme and a morpheme analyzed in the morpheme analysis; A granter 120 for giving a PLO tagr and a semantic code to the parsed phrase in the parser 110; A Lexico-syntactic pattern matcher 130 for matching the syntax having the PLO tagger and semantic code given by the grantor 120 to an SQL sentence; A searcher 140 that accesses a suitable DB for the query described by the matched SQL statement in the pattern matcher 130; And [responder 150] generating a response using the data from searcher 140.

In addition, the FAQ list retrieval system 20 includes a bottom-up natural language information searcher 21, a top-down information filter 22, and a FAQ editor 23. The bottom-up natural language information searcher 21 includes a morpheme syntax dictionary and statistical language information. A morpheme / syntax analyzer 210 for analyzing the user's natural language query based on the morpheme and syntax; A compound noun processor 211 which processes a morpheme and phrase-parsed query into a compound noun based on a single word, a compound noun dictionary, and air information between nouns; Information retriever 212; And a semantic discriminating matcher 213 that performs keyword matching using Korean wordnet semantic information with respect to a user query.

The top-down information filter 22 classifies a question sentence into about 30 kinds of Lexico-syntex patterns and question forms expressing lexical, part-of-speech, syntax, and semantic information in a regular expression form. ); A word refiner 221 may be configured to construct a dictionary including words dependent on a specific search region and adjust weights according to characteristics of the region using the dictionary.

In addition, the FAQ editor 23 updates the index used in the statistical natural language analyzer, the relational information of nouns and verbs used in the semantic filter, and the specialized domain dictionary used in the word refinement filter, and is used semi-automatically in the question type filter. It includes a knowledge base updater 230 for updating the Lexico-syntactic pattern and question type definition.

In addition, the website query-response system 30 includes a bottom-up natural language information searcher 31 and a top-down natural language information searcher 32 for receiving a search result from the bottom-up natural language information searcher 31 and searching. The natural language information searcher 31 may include a morpheme / syntax analyzer 310 for analyzing a user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; A compound noun processor 311 for processing the query analyzed in morpheme and syntax into a compound noun based on a single word, a compound noun dictionary, and air information between nouns; An information searcher 312 that performs a search using a stopword dictionary and weighted nouns of the index; And a semantic discriminating matcher 313 for performing keyword matching using Korean WordNet semantic information on the user query, and the top-down correct sentence extractor 32 includes a natural language of the user based on the morpheme syntax dictionary and statistical language information. A morpheme / syntax analyzer 320 that parses the query into morphemes and phrases; A pattern matcher 321 for assigning a PLO tagger and semantic code to the parsed phrase and determining a question type using a Lexico-syntactic pattern to which the PLO information and semantic code are assigned; Top-down information filter 322 using heuristics, such as semantic similarity, ISA format syntax structure, clue words, and position within a document.

Next, an integrated natural language query-response retrieval method for the interactive DB, the FAQ list, and the website configured as described above will be described in detail with reference to FIGS. As shown in FIG. 2, in step S11, the natural language query inputted from the user is analyzed into a morpheme and a partial syntax structure which is a minimum unit having one meaning in Korean grammar using a morpheme / syntax analysis dictionary and statistical language information. . This analyzed query is then distributed to a suitable search area, i.e. interactive DB search, FAQ list search, website search system 10, 20, 30, in step S12.

The query distributed to the interactive DB search system 10 is assigned a semantic code including schema information of the DB to the Korean WordNet information / domain dictionary in step 13, and the query with the semantic code is assigned to the query sentence. Structured Query, an interactive database query statement that is pre-stored through a Lexico-syntactic pattern determiner 130 that includes lexical, part-of-speech, PLO information, syntax information, and semantic code, and is described by regular expressions. Language). The SQL query thus determined finds the most optimal response in the database through the interactive database in step S14. When the interactive DB search system 10 is used as a separate system, it has a structure as shown in FIG.

On the other hand, the query list distributed to the FAQ list and the website query system 20, 30 processes the compound nouns using the single word and the compound noun dictionary in step S15, and then searches using the weights of the stopword dictionary and the indexed nouns. Using the bottom-up natural language information retrieval method, it is classified into FAQ query and website query.

Here, the query analyzed by the FAQ list query is given the semantic code of the Korean WordNet information / domain dictionary. The query word to which this semantic code is assigned includes a lexical, part-of-speech, PLO information, syntax information, and semantic code of the query sentence, and passes through a top-down FAQ filter 22 through a Lexico-Syntectic pattern determiner described in regular expression. Find the answer (S16). When using this FAQ list retrieval system 20 as a separate system, it has a structure as shown in FIG.

On the other hand, the query analyzed by the website correct answer document is given a PLO recognition and semantic code in step S17, and determines the pattern for determining the question type using the Lexico-Syntectic pattern, and the semantic similarity, ISA format. The correct answer sentence is found by passing through a top-down correct sentence extractor 32 including a top-down information filter 322 using heuristics such as a syntax structure, a word, and a position in a document. When using this website query-response system as a separate system 30, it has a structure as shown in FIG.

Then, in step S18, the best answer is found by integrating the interactive DB response, the FAQ response obtained from the top-down FAQ filter 21, and the correct answer sentence obtained from the top-down correct sentence extractor 32.

3 shows a procedural sequence for searching an interactive DB, a FAQ list, and a website in the integrated natural language question-answering system. As shown in FIG. 3, a user query is input into the integrated natural language question-response system according to the present invention (S21). The integrated natural language query-response system performs morpheme and partial parsing on the query input in step S21. Then, it is determined whether the user query analyzed in step S21 is suitable for the interactive DB search.

If the query is suitable for the interactive DB, the interactive DB search is performed (S22). Then, it is passed to the FAQ list retrieval module, and in step S23 it is determined whether the user requires a FAQ list retrieval to obtain more information than the response obtained in the interactive DB retrieval, and if so, the user query in step S25 Perform a list search. On the other hand, if not required, it is passed to step S28, which is an integrated response selection step.

In step S26, it is determined whether the user requests a website search, and if so, the website is searched to extract an appropriate response (S27). If the user does not require a website search, the search results are presented to the user through the integrated response selection step S28.

On the other hand, if the user query is not suitable for the interactive DB in step S21, bottom-up natural language information analysis is performed (S24). It is determined whether the analyzed user query is suitable for FAQ list search (S24). If appropriate, the FAQ list is retrieved in step S25. However, if it is not suitable for the user's FAQ list search, the website is searched and an appropriate response is extracted in step S27. Even when the user requests a web site search, the web site may be searched to extract an appropriate response (S26).

In this way, the integrated search result for the user's query is appropriately presented to the user through the selection of the integrated response (S28).

For example, if the user query of the integrated natural language query-response retrieval system is "What is Kim Jung-min's phone number?", The relevant interactive DB is searched to provide the desired information. How do I get after-sales service? ”If the answer is a question frequently asked by users, the FAQ list is searched for an appropriate response. In addition, if a user query is a query that can be answered by a specific document on the website, such as "Where is the nearest agency in Sillim-dong?", The website is searched and an appropriate document is provided as an answer.

Now, the interactive DB search system 10, the FAQ list search system 20, and the website search system 30 will be described in detail.

First, an interactive DB search system 10 and a search step will be described based on FIG. The query of the user input to the interactive DB search system 10 according to the present invention is subjected to morphological analysis and partial syntax analysis to generate a predetermined syntax unit (eg, noun phrase) (S110). The generated syntactic unit includes PLO information, which is information about which noun phrases belong to proper nouns, categories, people, places, and institutions, and extended semantic codes by including DB schema information in Korean WordNet and domain dictionaries. It is given (S120). After the PLO recognition and semantic code assignment process, Lexic-Syntic pattern matcher including the lexical, part-of-speech, PLO information, syntax information, and semantic code of the query sentence is described in the form of regular expression. The user query is mapped to a predefined SQL statement (S130). When the SQL statement is selected, a response is extracted through DB access (S140), and a proper response sentence is generated by using a predefined response frame (S150).

Next, a frequently asked question, a FAQ list search system 20 and a search step for easily finding answers to the FAQ will be described based on FIG. 5. Since the structure of the FAQ list retrieval system 20 has been described above, a detailed description thereof will be omitted. The FAQ list retrieval system obtains a suitable query-response pair via a bottom-up natural language information retriever 21 and a top-down information filter 22.

Bottom-up information retrieval analyzes the query through morpheme / syntax analysis (S210), and processes a compound noun using a single word and a compound noun dictionary (S211). After that, the search is performed using the stopword dictionary and the weight of the indexed noun (S212). A keyword matching search is performed by semantic discriminating matching using Korean WordNet semantic information (S213).

Therefore, in the case of using another keyword of similar meaning, which is a problem in a system using only the keyword matching search method by semantic discrimination matching using wordnet semantic information, the FAQ list search system 20 uses semantic classification and abstraction. Inference can be used to resolve errors from using other keywords.

On the other hand, the top-down information filter 22 improves the performance of the search by resolving errors resulting from information omissions other than nouns, and by refining words based on the dictionary considering the characteristics of the search target area. This top-down information filtering system accurately identifies user's question intention through lexico-syntectic pattern that expresses the sentence sentence, part-of-speech, syntax, and semantic information in the form of regular expression and the question type filtration that classifies 30 kinds of question forms. Identify and improve search performance (S220). In addition, the word reconstruction using a dictionary that includes words dependent on a specific search area is adjusted according to the characteristics of the search area (S221). This FAQ list retrieval system 20 has a FAQ editor, which constructs a robust FAQ list retrieval system 20 by using a collection of entered user's queries and a new question-answer on the FAQ list. Add a pair In addition, the FAQ editor 23 automatically updates the relational information of nouns and verbs used in indexing and semantic filtering used in statistical information retrieval systems, and specialized domain dictionaries used in word refinement. The Lexico-syntactic pattern and the question type definition used are updated (S230).

Next, the site query-response search system 30 is shown in FIG. As shown in FIG. 6, the user's natural language query performs top-down filtering 32 on the top N correct answer documents retrieved by the bottom-up natural language information searchers 21 and 31 shown in FIG. 5.

In the bottom-up natural language information searcher (21, 31), first, the user's natural language query is grouped into a certain syntax unit through morphological analysis and partial syntax analysis. Each syntax unit is given PLO information and a semantic code. The user query assigned the PLO information and the semantic code is passed to the pattern matcher module for determining the question type (S321). To determine the question type, the module uses the Lexico-Syntic pattern, which is described in regular expression form, including the lexical, part-of-speech, PLO information, syntax information, and semantic code of the query sentence in the Lexico-Syntax pattern S321). The pattern matcher reads the Lexico-syntactic pattern described in text form and filters the sentence that is correct according to the question type of the current query (S322). Table 1 is an example of Lexico-Syntectic pattern. In Table 1, the vocabulary starting with $ means Synset and its lower term of Korean WordNet to which the vocabulary belongs, and% means synonym.

Question type Person pattern (@Person |% person) (jCs | jOc) {0,1} (\?) {0,1} $ (@ person |% person) (jCs) {0,1} (\?) {0,1 } (. + (jOm) {0,1}) {0,1} who. * $ (@ person |% person) (. + (jOm) {0,1}) {0,1}% name (jCs | jOc) {0,1} (\?) {0,1} $ (@ person |% person) (. + (jOm) {0,1}) {0,1}% name (jCs) {0, 1} (What | How). * $ Who jCsWho jOm Symbol Description $: Hyponym%: Synonym {}: Repeat |: selection +: repeat one or more times *: repeat zero or more times

<Table 1> Lexico-syntatic pattern example

The top N documents retrieved using the bottom-up natural language information searcher (21, 31) also undergo a PLO recognition and semantic code assignment process. After this preprocessing of the user query and the retrieved document, the website query-response retrieval system 30 extracts a sentence estimated to be a correct answer from the retrieved document. To extract the correct answer, we use the question type, PLO information, semantic similarity, and various heuristics found in the pattern matcher module. The question type and PLO information serve as a filter to leave only the sentences that are likely to be correct and to remove the rest from candidates. Heuristics, such as semantic similarity and syntactic structure in ISA format, clue words, and position in a document, are used to score sentences that pass through the filter to select the sentences most likely to be correct. The website query-response retrieval system 30 also ranks the top N documents found by the bottom-up natural language information searcher (21, 31) by giving a low score to documents that do not contain the correct answer sentences. do.

The present invention is a search system that integrates FAQ list search and interactive DB search as well as document processing on the web, unlike a search system that displays only the documents on the web and presents the searched results. Organically integrated top-down natural language processing techniques, such as natural language processing technology, language knowledge base, and question-answer modeling, can provide users with convenient and accurate information.

Claims (11)

(Iii) an interactive DB search system for searching an interactive DB; (Ii) a FAQ list retrieval system for retrieving a list of answers to frequently asked questions due to queries and user requests not found in the interactive DB retrieval system; (Iii) an integrated query-response search system including a website query-response system that executes a query-response search on a website in response to a user's request and a query not found in the interactive DB search system or the FAQ list search system; . (i) a morpheme analyzer for morphologically analyzing a user's natural language query; (ii) a parser for determining a mathematical relationship between the analyzed morphemes; (i) a granter for assigning a PLO tagger and a semantic code to the analyzed syntax (Ⅳ) a Lexico-Syntectic pattern matcher for mapping the PLO tagger and the semantic code assigned distinctions into SQL statements; (Iii) a searcher for accessing a suitable DB for the query described by the SQL statement; (iii) an interactive DB search system comprising a response generator for finding an optimal response in the accessed DB. It includes a bottom-up natural language finder, a top-down information filter, and a FAQ editor. The bottom-up natural language information searcher, (i) a morpheme / syntax analyzer that parses the user's natural language query into morphemes and phrases based on morphological syntax dictionaries and statistical language information; (Ii) a compound noun processor for processing queries analyzed in case and phrase into compound nouns based on single words, compound noun dictionaries, and air information between nouns; (Iii) an information searcher for performing a search using a stopword dictionary and a weighted index of nouns; (Iii) a semantic distinguishing matcher which performs keyword matching using Korean wordnet semantic information for a user query, Top down information filter, (Ⅴ) expressing vocabulary, parts of speech, syntax, and semantic information in the form of regular expressions Question form filter which grasps Lexico-Syntex pattern and question form into 30 kinds to identify question intent; (Iii) includes a word refinement filter that constructs a dictionary containing words that depend on a particular search region and uses the weights to adjust weights according to the characteristics of the region, FAQ editor, (Iv) Updates indexes used in statistical natural language analyzers, relational information of nouns and verbs used in semantic filters, and specialized domain dictionaries used in word refinement filters; semi-automatic syntactical used in question type filters. FAQ list retrieval system that includes a knowledge base updater that updates pattern and question type definitions. Includes a bottom-up natural language finder and a top-down correct sentence extractor, The bottom-up natural language information searcher includes: (i) a morpheme / syntax analyzer for analyzing a user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; (Ii) a compound noun processor for processing morphemes and phrases into queries based on single words, compound noun dictionaries, and air information between nouns; (Iii) an information searcher for performing a search using a stopword dictionary and a weighted index of nouns; And (iii) a semantic discriminating matcher that performs keyword matching using Korean WordNet semantic information for the user query, Top-down correct sentence extractor, (ⅴ) stemming / syntax analyzer to analyze the user's natural language query based on the morpheme syntax dictionary and statistical linguistic information; (iii) PLO tagging and semantic code to give the parsed syntax PLO recognition and semantic code giver; (Iii) a pattern matcher for determining question types using Lexico-Syntectic patterns with PLO information and semantic code assigned syntaxes; And (iii) a top-down information filter using heuristics such as semantic similarity, syntax structure in ISA format, clue words, and position within a document. (i) stemming / partial parsing the input natural language query; (Ii) determining whether the analyzed query is a query suitable for the interactive DB search method and searching for a response using the interactive DB search method if the query is suitable for the interactive DB search method; (Iii) applying bottom-up natural language information analysis to a query that is not suitable for the interactive DB retrieval method; (Iii) determining whether the query analyzed by the bottom-up natural language information analysis is a query suitable for the FAQ list retrieval method and retrieving the response using the FAQ list retrieval method if the query is suitable for the FAQ list retrieval method; (v) retrieving the response with the website query-response retrieval method by determining whether it is not suitable for the FAQ list retrieval method or if the user requires a website retrieval; (Iii) integrating the responses obtained from the interactive DB search, the FAQ list search, and the website search to select the best answer. The method of claim 5, wherein the interactive DB search method comprises the steps of: (i) analyzing the user's natural language query in morphemes; (ii) determining a mathematical relationship between the analyzed morphemes; Assigning a tagger and a semantic code; (iii) mapping a division to which a PLO tagger and semantic code are assigned to an SQL statement; (Iii) accessing a suitable DB for the query described by the SQL statement; (iii) generating a response. The method according to claim 5 or 6, wherein the FAQ list retrieval method is used. Bottom-up natural language information search, (i) analyzing the user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; (Ii) processing the query analyzed in case and phrase into compound nouns based on single words, compound noun dictionaries, and air information between nouns; (Iii) performing a search using a stopword dictionary and a weighted index of nouns; And (Iii) performing keyword matching on the user query using Korean WordNet semantic information, Top-down information filtration, (Iv) classifying the Lexicon-Syntex pattern and the question form in which the sentence is expressed in a regular expression form of vocabulary, part-of-speech, syntax, and semantic information into a plurality of branches to identify the intention of the question; And (Iii) constructing a dictionary containing words dependent on the search region and using the same to adjust weights according to the characteristics of the region, Edit FAQ, (Iv) Update the indexes used in statistical natural language analyzers, relational information of nouns and verbs in semantic filtration, and specialized domain dictionaries used in word refinement filtration. And updating the question type definition. The method according to claim 5 or 6 or 7, wherein the website query-response retrieval method is as follows. Bottom-up natural language information search, (i) analyzing the user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; (Ii) processing the query analyzed in morpheme and syntax into a compound noun based on a single word, a compound noun dictionary, and air information between nouns; (Iii) performing a search using a stopword dictionary and a weighted index of nouns; And (Iii) performing keyword matching on the user query using Korean WordNet semantic information, Top-down correct sentence extraction, (Iii) analyzing the user's natural language query into a morpheme and syntax based on a morpheme syntax dictionary and statistical language information; (Iii) assigning a PLO tagger and semantic code to the parsed phrase; (Iii) determining the question type using Lexico-Syntectic pattern with the PLO information and the semantic code assigned syntax; And (Iii) performing top-down filtering using heuristics such as semantic similarity, syntax structure in ISA format, clue words, and position within a document. (i) analyzing the user's natural language query in morphemes; (ii) determining mathematical relationships between the analyzed morphemes; (i) assigning PLO tags and semantic codes to the parsed phrases; (i) PLO Mapping a tag and a semantic code assigned to the SQL statement; (Iii) accessing a suitable DB for the query described in the SQL statement; (iii) generating a response. Bottom-up natural language information search, (i) analyzing the user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; (Ii) processing the query analyzed in case and phrase into compound nouns based on single words, compound noun dictionaries, and air information between nouns; (Iii) performing a search using a stopword dictionary and a weighted index of nouns; And (Iii) performing keyword matching on the user query using Korean WordNet semantic information, Top-down information filtration, (Iv) classifying the Lexicon-Syntex pattern and the question form in which the sentence is expressed in a regular expression form of vocabulary, part-of-speech, syntax, and semantic information into a plurality of branches to identify the intention of the question; And (Iii) constructing a dictionary containing words dependent on the search region and using the same to adjust weights according to the characteristics of the region, Edit FAQ, (Iv) Update the indexes used in statistical natural language analyzers, relational information of nouns and verbs in semantic filtration, and specialized domain dictionaries used in word refinement filtration. And a FAQ list retrieval method that includes updating the question type definition. Bottom-up natural language information search, (i) analyzing the user's natural language query into a morpheme and a phrase based on a morpheme phrase dictionary and statistical language information; (Ii) processing the query analyzed in morpheme and syntax into a compound noun based on a single word, a compound noun dictionary, and air information between nouns; (Iii) performing a search using a stopword dictionary and a weighted index of nouns; And (Iii) performing keyword matching on the user query using Korean WordNet semantic information, Top-down correct sentence extraction, (Iii) analyzing the user's natural language query into a morpheme and syntax based on a morpheme syntax dictionary and statistical language information; (Iii) assigning a PLO tagger and semantic code to the parsed phrase; (Iii) determining the question type using Lexico-Syntectic pattern with the PLO information and the semantic code assigned syntax; And (Iii) performing top-down filtration using heuristics such as semantic similarity, syntax structure in ISA format, clue words, and position within a document.