KR102540939B1 - System and method for improving the adequacy of natural language searches - Google Patents

Abstract

The present invention provides a relevance improvement system and relevance improvement method for natural language search in which a score obtained by adding a phrase score as well as a token score is given to a document to be matched with a search term, and sorted in descending order according to the assigned score and provided to the user in order. Suggest. The natural language search relevance improvement system includes a tokenization module, a document filter module, a phrase weighting module, and a document score computation module.

Description

System and method for improving adequacy of natural language searches {System and method for improving the adequacy of natural language searches}

The present invention relates to a natural language search system, and more particularly, the relevance of a natural language search in which a sum of a token score and a phrase score is assigned to a document to be matched with a search term and sorted in descending order according to the assigned score. It is about the improvement system and the adequacy improvement method.

Natural language refers to the language that humans use for communication in everyday life to distinguish it from programming language or machine language used by computers. Natural language processing (NLP) refers to a process of interpreting natural language, which is a human language, so that a computer, which is a machine, can understand it, and allowing the computer to process it.

A search is performed using a natural language as a search term, and a correlation between a given query, that is, a natural language entered as a search term, and a document is evaluated, and how often a term included in the query appears in each document is evaluated.

Republic of Korea Patent Registration No. 10-2256007 (May 18, 2021) 'System and method for searching documents and providing responses through natural language queries' tokenizes natural language queries when a natural language query is input from a user, and tokenized natural language queries It describes a technique for selecting documents by performing a similarity test with all documents in the database using .

The order of service provision of the selected documents, that is, which document is best provided to the user among a plurality of selected documents and which document is to be provided next, is not considered.

If the order of the documents to be served is not determined, there is a disadvantage in that a general document is provided first, rather than a document in which the best answer to the query is written, increasing the user's search time.

In addition, even if a certain principle is set, when the principle is not the best, the same disadvantages as when the order of serviced documents are not determined will be included.

The technical problem to be solved by the present invention is to improve the appropriateness of natural language search by assigning a score obtained by adding not only a token score but also a phrase score to a document to be matched with a search term, sorting in descending order according to the assigned score, and providing the user in order. It is to propose a system.

Another technical problem to be solved by the present invention is to improve the appropriateness of natural language search by assigning a score obtained by adding not only a token score but also a phrase score to a document to be matched with a search word, and sorting in descending order according to the assigned score and providing it to the user in order. It's about suggesting a way.

The technical problems to be achieved in the present invention are not limited to the above-mentioned technical problems, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

A system for improving relevance of natural language search according to the present invention for achieving the above technical problem includes a tokenization module, a document filter module, a phrase weighting module, and a document score calculation module.

The tokenization module collects natural language search words included in a query input by a user and separates them into tokens. The document filter module filters documents including tokens separated by the tokenization module among a plurality of documents. The phrase weighting module generates a phrase by combining two tokens among a plurality of tokens and assigns a weight to each phrase. The document score calculation module calculates a token score and a phrase score using the search word input by the user, the filtered document, the phrase, and the weight.

The method for improving the relevance of natural language search according to the present invention for achieving the other technical problem is performed by the system for improving the relevance of natural language search according to claim 1, and the tokenization module is included in the query input by the user. Collecting natural language search words and separating them into tokens, filtering, by the document filter module, documents including tokens separated by the tokenization module among a plurality of documents, and by the phrase weighting module, two tokens among a plurality of tokens generating phrases and assigning a weight to each phrase, and calculating a token score and a phrase score by the document score calculation module using the search word entered by the user, the filtered document, the phrase, and the weight. include

The technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

The relevance improvement system and relevance improvement method of natural language search according to the present invention considers not only the token score but also the phrase score in parallel, so that search and relevance can be performed more effectively than the existing method of improving the relevance of search using only the token score. There is an advantage of enabling service provision of searched information.

The effects obtainable in the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description below. will be.

1 is an embodiment of a system for improving relevance of natural language search according to the present invention.
2 is an embodiment of a method for improving relevance of natural language search according to the present invention.

In order to fully understand the present invention and the advantages in operation of the present invention and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings describing exemplary embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings. Like reference numerals in each figure indicate like members.

1 is an embodiment of a system for improving relevance of natural language search according to the present invention.

Referring to FIG. 1 , the natural language search relevance improvement system 100 according to the present invention includes a natural language search module 110, a tokenization module 120, a document filter module 130, and a phrase weighting module 140. , a document score computation module 150 and a document sorting module 160.

The natural language search module 110 searches for natural language included in a user query.

For convenience of explanation, it is hereinafter assumed 'Q = [method and system for improving relevance in natural language search]' as an example of a searched query.

The tokenization module 120 separates a natural language search term included in a query input by a user into tokens.

For example, it may be possible to tokenize based on spacing in a search term sentence, such as query Q = [natural language, search, relevance, improvement, method, and system] separated into tokens.

The document filter module 130 filters documents including tokens separated by the tokenization module among a plurality of documents. At this time, it is possible to filter (or select) all documents including the token regardless of the number of tokens included, but it is preferable to filter documents containing more than a preset reference token (number).

The phrase weighting module 140 generates a phrase by combining two tokens among a plurality of tokens and assigns a weight to each phrase.

For example, phrase groups P = [{(in natural language search), 6.0}, {(relevance in search), 5.0}, {(improving relevance), 4.0}, {(how to improve), 3.0}, {(method and), 2.0}, {(and system), 1.0}].

The first phrase “in natural language search” is given a weight of 6.0, and the second phrase “relevance in search” is given a weight of 5.0.

The document score calculation module 150 calculates a token score and a phrase score using a search word input by a user and filtered documents, phrases, and weights.

Equation 1 below is an equation for calculating a token score.

는 쿼리에서 i(i는 변수) 번째 토큰,

는 검색된 문서에서 매칭된 토큰 수,

은 1.2를 기본값으로 설정되어 있는 TSP(term saturation parameter), b는 디폴트(dafault)로 취하는 상수인 0.75, dl은 검색된 문서의 길이, avgdl은 전체 문서의 평균 필드의 길이, docCount는 문서의 총 개수,

는 쿼리에서 i번째 토큰(

)를 포함하는 문서 빈도의 역수 그리고

는 해당 토큰(

)을 포함하는 문서의 개수를 각각 의미한다. In Equation 1, D is the document to be compared, Q is the search word entered by the user,

is the i (i is a variable) token in the query,

is the number of matched tokens in the retrieved document,

is the term saturation parameter (TSP), which is set to 1.2 as a default value, b is a constant of 0.75 as default, dl is the length of the searched document, avgdl is the average field length of all documents, and docCount is the total number of documents ,

is the ith token in the query (

) and the reciprocal of the frequency of documents containing

is the corresponding token (

) means the number of documents containing each.

dl will be the number of tokens when documents are tokenized, and avgdl will be the average number of tokens per document when all documents in the index are tokenized.

If document frequency is the basis for judging how often a specific term (token) appears, the reciprocal of document frequency (IOnverse document frequency) recognizes frequently occurring words as less important words and tries to lower the weight. reflects the attempt

Referring to Equation 1, the token score will be an index for evaluating how often a term (token) in a query appears in each document.

That is, for the token score of Equation 1, the weight of frequently appearing words is lowered, the meaning of the number of tokens according to the length of the document is reflected, and whether the keyword in the query frequently appears in the document is converted into a score.

In the present invention, in addition to the token score of Equation 1, a phrase score described below is additionally calculated to improve the appropriateness of natural language search.

Equation 2 below is an equation for calculating a phrase score.

는 문서에서 j번째(j는 변수) 프레이즈(phrase)의 개수를 확인하는데 토큰과 토큰 사이의 거리가 주어진 거리 이하인 프레이즈의 개수이고, weight는 가중치이다. In Equation 2,

is the number of phrases where the distance between tokens is equal to or less than the given distance, and weight is the weight.

In Equations 1 and 2, n and m are natural numbers.

The document sorting module 160 sorts in descending order of scores for each document obtained by adding a token score and a phrase score. Scores per document will allow the user to provide the order of documents containing the most search terms.

2 is an embodiment of a method for improving relevance of natural language search according to the present invention.

Referring to FIG. 2 , the natural language search relevance improvement method 200 according to the present invention includes a natural language search step 210, a tokenization step 220, a document filter step 230, and a phrase weighting step 240. , document score calculation step 250 and document sorting step 260.

In the natural language search step 210, a natural language included in a user query is searched.

In the tokenization step 220, the tokenization module 120 collects natural language search terms included in the query input by the user and separates them into tokens.

In the document filtering step 230, the document filter module 130 filters documents including tokens separated by the tokenization module among a plurality of documents.

In the phrase weighting step 240, the phrase weighting module 140 generates a phrase by binding two tokens among a plurality of tokens and assigns a weight to each phrase.

In the document score calculation step 250, the document score calculation module 150 calculates a token score and a phrase score using the search word input by the user and filtered documents, phrases, and weights.

In the document sorting step 260, the document sorting module 160 sorts in descending order of scores for each document obtained by summing the token score and the phrase score.

In the step of separating into tokens (220), the search word is tokenized based on spaces in the search word sentence, and in the step of filtering the document (230), documents containing tokens equal to or greater than a predetermined standard token are filtered, and weight is assigned to the phrase. In the assigning step 240, it is preferable to generate two tokens adjacent to each other as one phrase.

The token score is based on the search term entered by the user, the documents to be matched against the token, the number of matched tokens in the filtered documents, the length of the documents retrieved, the average length of the documents retrieved, and the inverse data frequency for the tokens included in the query. (Inverse Document Frequency).

The phrase score is preferably calculated using the number of matched phrases in the document and the number of phrases in which the distance between tokens in the document is equal to or less than a predetermined distance.

In the above, the technical idea of the present invention has been described together with the accompanying drawings, but this is an illustrative example of a preferred embodiment of the present invention, but does not limit the present invention. In addition, it is obvious that anyone skilled in the art can make various modifications and imitations without departing from the scope of the technical idea of the present invention.

110: natural language search module
120: tokenization module
130: document filter module
140: phrase weighting module
150: document score calculation module
160: document sorting module

Claims (14)

A tokenization module that collects natural language search words included in a query entered by a user and tokenizes them based on spaces;
a document filter module for filtering documents that include more than a predetermined standard number of tokens from among a plurality of documents separated by the tokenization module;
a phrase weighting module for generating a phrase by binding two tokens adjacent to each other among a plurality of tokens and assigning a weight to each phrase;
a document score calculation module that calculates a token score and a phrase score using a search word input by a user, the filtered document, the phrase, and the weight; and
A document sorting module for arranging scores for each document obtained by summing the token score and the phrase score in descending order;
The token score is based on a search word entered by the user, documents to be matched with the token among the documents, the number of matched tokens in the filtered documents, the length of the retrieved documents, the average length of the retrieved documents, and the tokens included in the query. The ratio of the length of the individual document to the average length of the filtered document is reflected for the number of matched tokens in the individual document included in the filtered document. It is characterized by applying a weight and an inverse document frequency for the number of matched tokens according to the document length,
The phrase score,
Calculation using the number of matching phrases in the document and the number of phrases in which the distance between tokens in the document is less than or equal to a given distance
Relevance improvement system for natural language search.
delete delete delete delete delete delete It is performed by the system for improving the adequacy of natural language search described in claim 1,
The tokenization module collects natural language search words included in a query input by a user, tokenizes them based on spaces, and separates them;
filtering, by the document filter module, documents including a predetermined standard number of tokens or more of tokens separated by the tokenization module among a plurality of documents;
generating, by the phrase weighting module, two tokens adjacent to each other among a plurality of tokens to generate a phrase and assigning a weight to each phrase;
calculating, by the document score calculation module, a token score and a phrase score using the search word input by the user, the filtered document, the phrase, and the weight; and
A document sorting module sorting scores for each document obtained by summing the token score and the phrase score in descending order;
The token score is based on a search word entered by the user, documents to be matched with the token among the documents, the number of matched tokens in the filtered documents, the length of the retrieved documents, the average length of the retrieved documents, and the tokens included in the query. The ratio of the length of the individual document to the average length of the filtered document is reflected for the number of matched tokens in the individual document included in the filtered document. It is characterized by applying a weight and an inverse document frequency for the number of matched tokens according to the document length,
The phrase score,
Calculation using the number of matching phrases in the document and the number of phrases in which the distance between tokens in the document is less than or equal to a given distance
A method for improving the relevance of natural language search.
delete delete delete delete delete delete

Images