KR100597435B1 - System and method for classfying question based on hybrid of information search and question answer system - Google Patents

Abstract

The present invention relates to a hybrid-based question classification system and method in an information retrieval and question answering system. The present invention relates to a question work name processing for recognizing a work name included in an input question, to morphological analysis of a question, object name recognition, and lexical meaning. The question language analysis process converts individual vocabulary into meaningful codes through tagging process, and the rule-based question classification classifying the types of correct answers required by a question by using meaningful LSP form codes of the question and predefined question classification rules. Using the process and the statistical information constructed from the learning document tagged with the correct answer type of the question, the statistics-based question classification process for classifying the LSP code of the question, and the results of the rule-based question classification and the statistics-based question classification Information inspection by going through the process of determining the type of correct answer to finally determine the type of correct answer to the user's question From the vast documents that are the search results for the user's questions in color and answering questions, only the correct answers the user wants can be focused, improving the performance of the information retrieval and question answering system.

Information search, question and answer, title recognition, hybrid question classification

Description

SYSTEM AND METHOD FOR CLASSFYING QUESTION BASED ON HYBRID OF INFORMATION SEARCH AND QUESTION ANSWER SYSTEM}             

1 is a view showing a conventional search process,

2 is a diagram illustrating a question classification process using a query type classifier based on a conventional support vector machine;

3 is a diagram showing a hybrid-based question classification system and a classification method according to an embodiment of the present invention;

4 is a view showing various types of correct answers for question classification of a user of the present invention;

5 is a diagram showing a relationship between a dictionary of works already constructed and a unique code of the present invention;

6 and 7 are diagrams illustrating a filtering relationship using the built-in left / right context information of a work name;

8 is a view showing a language analysis process for a question of the present invention,

9 shows an example of an LSP rule,

10 shows an example of a question corpus.

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

10: Question work name processing unit 11: Work name and recognition context DB

20: question language analysis unit 21: language analysis knowledge DB

30: Rule-based question classification unit 31: Question classification rule DB

40: Statistics based question classification unit 41: Question corpus statistical information DB

50: question type determination section

The present invention relates to a hybrid-based question classification system and method in an information retrieval and question answering system. More specifically, an information retrieval and question classifying a question by finding a type of correct answer required by a question in a parameter included in the question. A hybrid based question classification system and method in an answering system.

In an information retrieval system, a user's query is understood as a sequence of keywords, and the relationships between the keywords are represented by Boolean operations. The keyword-based information retrieval system uses weights using information such as keyword frequency (TF) or inverse document frequency (IDF) assigned to the indexing process when the keyword exists in the document. . Accordingly, the result of the search is the entire document related to the keyword, and finding the part desired by the user in the document requires the effort of the user. However, the user wants the search system to find the answer contained in the document directly, and the question answering system wants to classify the type of correct answer required by the question to satisfy the user's purpose.

In the past, a method has been proposed for a service in which an artificial intelligence system analyzes a user's question input in natural language to effectively present information existing on the Internet. Through the expansion and classification of user questions, 'Definition', 'Legal', 'Medical', 'Education' and 'Internet' are classified.

In addition, a document retrieval system and a question answering system have been proposed, where keywords are extracted from the question for the question answering system, which are non-critical about key forms keywords and supplementary information on the central subject represented by the question. Extract the form keyword. For example, "Where are the winners of the 2002 FIFA World Cup?", "2002", "Host", "FIFA", "World Cup", "Winner", "Country", "Where", etc. Get Based on this, a keyword type is assigned to each search keyword, and a meaning expression such as a date expression in 2002 and an organization name in FIFA is given. In addition, the verbs are classified into main verbs and auxiliary verbs according to their syntactic properties, and the questions are classified using the interrogative verb included in the question. The meaning expression of each word is hierarchical. In case of date, it has year / month / day / time structure, and in case of place, it has national / provincial prefecture / city / provincial village / address level.

The process and result of "Where is the winner of the 2002 FIFA World Cup?" As shown in FIG. 1, a keyword form is assigned using a semantic attribute for each word, and a question form is classified by using a semantic form of individual keywords such as the questionnaire 'where' and 'country'. We are finding out question.

In addition, a query type classifier based on a support vector machine for Korean question-and-response systems has been proposed. This query type classifier supports the support vector machine through feature extraction and feature selection, feature weight assignment and learning from the question corpus for question classification. We create a and use it to determine the type of question. This is largely divided into a query learning part and a query classification process, and the detailed flow is shown in FIG. 2.

To classify the query type, each word is subjected to the process of recognizing the name of the object and attaching the meaning marker. Each vocabulary is classified into one of the following four words.

-Proper noun: name of person, country, city, institution

-Common nouns: job title, hobby, job title, etc.

Unit nouns: km, m, kg, g, mg, etc.

-Other: Special Words Needed for Q & A

As shown in FIG. 2, the individual words in the question in the learning query set are encoded through object name recognition and semantic labeling, and a vector is formed using the same, and the support vectors for each query type are learned. Use this to classify questions.

However, this type of support vector machine based question type classifier for the Korean question and answer system is a statistics-based model, which requires the construction of a large-scale learning question. . In addition, by using only the statistical model, it was difficult to expand the new question correct answer type, and it was difficult to tune the question classification system.

On the other hand, a method has been proposed for a service that effectively analyzes the user's question input in natural language and presents information on the Internet. The classification system of the question is defined by definition, law, 'Medical', 'education' and 'internet' are very simple, and the question itself is a way to find the corresponding category rather than the type of correct answer the question requires. In this section, semantic attributes are assigned to individual vocabulary and questions are classified by combining semantic attributes and question sentences, and the questions are classified using heuristic information rather than formal and systematic methods such as concrete rules or statistical models. Doing.

As such, in the existing methods, the names of various works included in the question (names of works in artistic fields such as film, drama, fiction, drama, opera, music, etc.) consist of more than one word, and the individual vocabulary in the analysis of the question. There was a problem with the accuracy of the question classification by processing the question without any awareness of the meaning.

Accordingly, an object of the present invention is to solve the above problems of the prior art, the recognition of the name of the work included in the question, the vocabulary, the inclusion and exclusion relationship of important vocabulary, parts of speech, interrogation, semantic classification of the vocabulary and The present invention provides a hybrid-based question classification system and method in an information retrieval and question answering system that classifies questions by finding the type of correct answer required by a vocabulary individual type.

Hybrid based question classification system in the information retrieval and question answering system of the present invention for achieving the above object, the question title processing unit for recognizing the name of the work included in the input question; A question language analysis unit for converting a question into a meaningful code through morphological analysis, entity name recognition, and vocabulary semantic tagging process; A rule-based question classification unit for classifying the types of correct answers required by the question using meaningful LSP type codes of the questions and predefined question classification rules; A statistics-based question classifier configured to classify the LSP code of the question using statistical information constructed from a learning document tagged with the correct answer type of the question; And a question correct type determination unit that finally determines the type of the correct answer to the user's question using the results of rule-based question classification and statistical-based question classification.

On the other hand, the hybrid based question classification method in the information retrieval and question answering system of the present invention comprises: a question work name processing step of recognizing the work name included in the input question; A question language analysis step of converting the individual words into meaningful codes through morphological analysis, entity name recognition, and vocabulary semantic tagging process; A rule-based question classification step of classifying a type of correct answer required by a question using a meaningful LSP type code of the question and a predefined question classification rule; A statistics-based question classification step of performing classification on the LSP code of the question using statistical information constructed from a learning document tagged with the correct answer type of the question; And determining a type of a question correct answer that finally determines a type of a correct answer to a user's question by using the results of rule based question classification and statistical based question classification.

As described above, the system and method proposed by the present invention are to determine the semantic class of correct answers required by the question by using the semantic and statistical characteristics of the vocabulary appearing in the user's question.

That is, the present invention, in the information retrieval and question answering system, defines a semantic classification of the correct answer required by the question in advance for the user's natural language question, and automatically classifies the type of the correct answer required by the user's question in real time. Instead of finding the right document for the question, it is a question-categorization method that helps you find the correct answer in the document. For example, for the question "Where is the company that runs Everland?" For the "Company name" and "A book made as part of the distribution of Hunminjeongeum?" Finding the type and classifying the question, for this purpose: 1) predefine the type of correct answer for the question the user is likely to ask, 2) be included in the question, and individual words have no meaning, but are combined Recognizes the name of a work that acts as a meaningful unit, and 3) classifies the question types through elaborate and flexible rules for each type of correct answer for each question, and 4) does not determine the rules by statistical methods. A method and apparatus for automatically classifying a question classifying the correct answer type of a question.

Hereinafter, a hybrid based question classification system and method in an information retrieval and question answering system of the present invention will be described in detail with reference to the accompanying drawings.

3 is a diagram illustrating a hybrid based question classification system and a classification method according to an embodiment of the present invention. As shown in FIG. 3, the hybrid-based question classification system of the present invention includes a question work name processing unit that recognizes a work name (a title composed of one or more words such as a book name, a play / movie title, a music name, etc.) largely indicated in a question. (10), a question language analysis unit 20 which performs morphological analysis of the question, object name recognition, lexical semantic attachment and lexical semantic classification based on the concept network, and a rule-based question classification unit using a manual classification rule. (30), a statistics-based question classification unit 40 using statistical information learned from documents tagged with the correct answer type of question, and finally a correct answer type determination unit 50 of the question.

Here, the question work name processing unit 10 further includes a work name table for recognizing the work name and a work name and recognition context DB 11 storing the work name left / right context rules. The question language analysis Part 20 further includes a language analysis knowledge DB (object name dictionary, object name recognition context, lexical concept network, mutual information for classification of nouns) 21 for question language analysis, and the rule-based question classification unit. (30) further provides a query classification rule DB (31) based on the Lexico-Semantic Pattern (LSP) for rule-based question classification, and the statistics-based question classification unit 40 has a statistics-based question classification using Naive Bayesian. Question corpus statistical information DB 41 automatically learned from the learning document is further provided for.

Hybrid-based question classification method using this system, largely the question title processing step (S10), question language analysis step (S20), rule-based question classification step (S30), statistics-based question classification step (S40), question correct answer type The determination proceeds to step S50.

As a result of the above process, the result of the user's question classification may be various types of correct answers shown in part in FIG. The processing method and the knowledge used for each step are as follows.

-Title processing step (S10)

In a system for information retrieval and question and answer, the user's question can be expressed words such as the name of the work. For example, in a question such as "Who was the heroine of Gone with the Wind?", The user would ask the question without displaying special parentheses for Gone with the Wind. Ham refers to the titles that are widely used in various genres such as plays, movies, novels, music, and dramas. At this time, if you search documents by using 'wind', 'disappear', etc. as individual keywords in the search step, you may get unwanted answers. In the work name processing step 10, the question is converted into a work name using a work name dictionary and a unique code previously constructed as shown in FIG. For example, in "What is the genre of the movie ET?", "ET" and "genre" are in the title dictionary, and they are automatically converted to "What is the hat 49126 of the movie Hat 24191?" 'Genre' may be in the actual work name, but is not the work name in the context of the question, and is filtered using the work name left / right context information constructed as in FIGS. 6 and 7. Thus, in the end, it produces something like "What is the genre of movie Hat 24191?" The work name and recognition context DB 11 can be semi-automatically constructed from the web using the work name dictionary.

-Question language analysis step (S20)

This class is used to classify questions by linguistic analysis. 8 is a diagram illustrating a language analysis process for a question of the present invention. As shown in FIG. 8, in the morpheme analysis step S201, the parts of speech of each morpheme are determined using the morpheme dictionary 2011, and when the morpheme is a question mark, information on the morpheme is given. In the entity name recognition step (S202), about 160 predefined semantic classifications are defined as the entity name, and recognition thereof is performed. As shown in FIG. 4, the individual names are classified in detail for human names, academic names, theories, artifacts, organization names, names, culture / civilization, dates, times, quantities, events, animals, plants, materials, and terminology. . The entity name dictionary 2021 is a dictionary of more than 160 entity classifications. In the lexical meaning tagging step (S203), the noun vocabulary concept network 2031 is used to assign a concept to each noun. In the lexical meaning classification step (S204), only one lexical meaning of information attached in the lexical meaning tagging step (S203) is determined. To this end, the mutual information 2041 between the vocabularies in one sentence is obtained from the large corpus and utilized. For example, in the question “Does the door act as the gate in the temple?”, The door is used to mean '#shape', '#facility', '#unit', etc. To be discerned.

Rule-based question classification step (S30)

For rule-based question classification, LSP (Lexico-Semantic Pattern) question classification rules are used. Each vocabulary included in the question is coded in LSP format according to the characteristics of the vocabulary at the language analysis stage. The LSP form refers to the process of converting a word included in a question into a part-of-speech, vocabulary, entity name, question, and semantic code. That is, the question is converted into an LSP, and the question type is classified by comparing with a predefined LSP rule. Several operators are defined for this purpose and the following operators are used. Write a rule using the operators below, and construct a question based on the operators below.

-^ Vocabulary: Real Vocabulary Information

* Ex) plants-> ^ plants, occupation-> ^ occupation

* Matches the actual vocabulary shown in the question.

Parts of Speech: Part of speech information

* Ex) Term->! Nc, Dar->! Pa,->! Jc, did->! Ep

* Match the parts of speech in the vocabulary.

-Question: whether the question is a vocabulary

* Ex) what-> @what, when-> @when

-& Object name: vocabulary is one of the object types

* Eg UK-> & LC_COUNTRY, Lark-> & AM_BIRD

* Matches the entity type to which the vocabulary presented in the question belongs.

-# Meaning: Expression of meaning of vocabulary

* Ex) Door-> #Facilities, austerity-> #Reduction

* Matches the expression of meaning in the vocabulary

% Vocabulary: vocabulary inclusion

* Example) University->% School, Elementary School->% School

* Matches partial inclusion relations in vocabulary

^ $: Special symbol to end the question

Each operator can be combined by the following selection and inclusion operators.

* [@What]: Information in ‘[’ and ‘]’ can be omitted

* {@What}: The information in "{" and "}" is important for classifying questions

* <@What>: Information within ‘<’ and ‘>’ should not be listed

* ‘|’: Selectable

An example of the LSP rule according to the above expression method is shown in FIG. 9. Through the linguistic analysis step, the question is transformed into LSP form and compared with the established LSP rules to classify the type of correct answer required by the question. For example, "What novel did Hemingway write?" Goes through a linguistic analysis step and converts it into a column such as "& PS_NAME! Jx @ which # book! Jc! Pv! Ep! Ef" and compares it to the LSP rule, "@ Since a #book ”partially matches, it will be classified as a search for“ AF_WORKS ”. A question can be partially coincident with several rules, each of which is weighted according to the meaning of the LSP, and sorted the result of the sum of the weights to be considered the final classification of the question.

-Statistics based question classification (S40)

When classifying questions, you may not be able to classify on an LSP basis. For example, a question such as “What is the smallest species of ornamental tropical fish?” Is the information that is important for classifying the correct answer type, separated by other words. In such a case, it is difficult to classify the pattern, and perform statistical-based question classification that considers the qualities included in the question as a vector form. where q is the question, at is the correct type of question, t is the information contained in the question (object name, semantic tagging, part-of-speech, question mark, vocabulary, etc.), and n is the size of the entire learning question. ) Finds the type of the correct answer at for the given question q. To this end, using the learning document (Fig. 8) attached to the correct answer type required by the question to the question, and to learn the statistical information that t has at with at each q, based on this, using the Naive Bayesian model P (at | Find the value of q).

--- 식(1)

--- Equation (1)

In this case, considering the case where the probability cannot be calculated directly and the probability is 0, the correct answer type for the question is classified by modifying the above equation as follows.

--- Equation (2)

-Determination of the correct type of question (S50)

If a rule-based question classification is used, the rule-based question classification value is used when the weight is greater than or equal to the threshold value. However, when the threshold value is less than the threshold value, the statistical question classification result is used. However, in the statistical question classification, if the threshold value obtained through learning is lower than the threshold, the classification of the question is abandoned, and the meaning expression of the vocabulary is returned as the result value of the question classification from the vocabulary at the end of the question.

Although the present invention has been described in more detail with reference to some embodiments, the present invention is not necessarily limited to these embodiments, and various modifications can be made without departing from the spirit of the present invention.

As described above, the hybrid-based question classification system and method in the information retrieval and question answering system according to the present invention can reduce the document containing the correct answer desired by the user, or process such as entity name recognition and paragraph search from the actual document. It can help to extract the correct answer directly, thereby improving the accuracy of the IR or question answering system.

Claims (14)

A question work name processing unit recognizing a work name included in the input question; A question language analysis unit for converting a question into a meaningful code through morphological analysis, entity name recognition, and vocabulary semantic tagging process; A rule-based question classification unit for classifying the types of correct answers required by the question using meaningful LSP type codes of the questions and predefined question classification rules; A statistics-based question classifier configured to classify the LSP code of the question using statistical information constructed from a learning document tagged with the correct answer type of the question; And Question answer type determination unit that finally determines the type of answer to the user's question by using the results of rule-based question classification and statistics-based question classification Hybrid based question classification system in the information retrieval and question answering system comprising a. [Claim 2] The information search and question according to claim 1, wherein the question work name processing unit further comprises a work name table for recognizing the work name and a work name and a recognition context DB storing the work name left / right context rules. Hybrid-based Question Classification System in Response System. The method of claim 1, wherein the question language analysis unit further comprises a language analysis knowledge DB including an entity name dictionary, an entity name recognition context, a lexical concept network, and mutual information for discriminating noun meanings for question language analysis. Hybrid-based Question Classification System in Information Retrieval and Question and Answer Systems. The method of claim 1, wherein the rule-based question classification unit further includes a question classification rule DB (31) based on a Lexico-Semantic Pattern (LSP) for rule-based question classification. Hybrid based question classification system. The information retrieval and question answering system of claim 1, wherein the statistics-based question classification unit further includes a question corpus statistical information DB automatically learned from a learning document for statistics-based question classification using Naive Bayesian. Hybrid based question classification system. A question work name processing step of recognizing a work name included in the input question; A question language analysis step of converting the individual words into meaningful codes through morphological analysis, entity name recognition, and vocabulary semantic tagging process; A rule-based question classification step of classifying a type of correct answer required by a question using a meaningful LSP type code of the question and a predefined question classification rule; A statistics-based question classification step of performing classification on the LSP code of the question using statistical information constructed from a learning document tagged with the correct answer type of the question; And A step of determining the type of question correct answer that finally determines the type of answer to the user's question by using the results of rule-based question classification and statistics-based question classification. Hybrid based question classification method in the information retrieval and question answering system comprising a. The method of claim 6, wherein the question title processing step comprises: Converting a question into a work name using a pre-built work name dictionary and a unique code; Filtering an area other than the work name by using the work name left / right context information constructed on the converted question; And Outputting the filtered question Hybrid based question classification method in the information retrieval and question answering system comprising a. The method of claim 6, wherein the question language analysis step comprises: Determining the parts of speech of each morpheme using a morpheme dictionary; Performing recognition on an entity name using an entity name dictionary in which a plurality of predefined semantic classifications are defined as entity names; Assigning a concept to each noun using a noun vocabulary concept network; And Determining the lexical meaning of only one of the information attached in the lexical semantic tagging process through mutual information between the vocabularies in a sentence from a large corpus Hybrid based question classification method in the information retrieval and question answering system comprising a. The method of claim 8, wherein the individual name includes a person's name, academic name, theory, artifact, organization name, place name, culture / civilization, date, time, quantity, event, animal, plant, substance, terminology. Hybrid based question classification method in information retrieval and question answering system. The method of claim 6, wherein the rule-based question classification step comprises: In the question language analysis step, the information is coded into an LSP form according to the characteristics of each vocabulary included in the question, in the form of an LSP, and an information retrieval comprises an operator included in the question to classify the question types. Hybrid based question classification method in question answering system. The method of claim 6, wherein the statistical based question classification step, Using the learning document with the correct type of question required for the question, the statistical information of each variable is learned, and the correct type of question for the given question is found by using the following equation (1). Hybrid-based Question Classification in Information Retrieval and Question Answering Systems.

--- Equation (1) Where q is the question, at is the correct answer type of the question, t is the name of the individual included in the question, semantic tagging, parts of speech, question sentences, vocabulary, and the like, and n is the size of the entire learning question. 12. The method of claim 6 or 11, wherein the statistical based question classification step, Hybrid based question classification method in an information retrieval and question answer system, characterized by classifying the correct answer type for a question by considering the case where the probability is 0.

--- Equation (2) The method of claim 6, wherein the determining of the correct type of question comprises: Hybrid based question classification method in information retrieval and question answering system, characterized in that only the question classification satisfying the threshold among rule based question classification and statistical based question classification is selectively determined as the correct answer type of question. The method of claim 6 or 13, wherein the determining the correct type of question If the result obtained through learning in rule-based and statistical-based question classification is less than or equal to the threshold value, the classification of the question is abandoned and the meaning of the vocabulary is expressed from the last word of the question. Hybrid based question classification method in information retrieval and question answering system, characterized in that it is returned as a value.

Images