KR20060122276A - Relation extraction from documents for the automatic construction of ontologies - Google Patents

Abstract

A method for extracting concept relation from documents for automatic ontology construction is provided to automatically construct ontology requiring high cost and time by extracting the concept relation from a document set with a hybrid type for combining relation information extracted through a symbolic method and a statistical method. The symbolic method automatically extracts and uses a generalized syntax pattern. The statistical method constructs correlation information by clustering association information and fixes a name of the relation based on the constructed correlation information. The symbolic method is combined with the statistical method.

Description

Relation extraction from documents for the automatic construction of ontologies

1 is a diagram illustrating a hybrid relational extraction combining a symbolic method and a statistical method according to the present invention.

2 is a flowchart illustrating a symbolic method used in the present invention and illustrating a flow of extracting a concept from a document.

3 is a flowchart illustrating a statistical method used in the present invention, illustrating a flow of extracting a concept from a document.

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

1: Documentation-Professional documents in the field of ontology construction

List of Candidate Concepts-Candidate Concepts Used in Ontology Construction

10: symbolic method module

20: Statistical Method Module

30: hybrid method module

31: Ontology

101: Seed Concepts-Examples of Relationships Selected by Experts

110: Document Preprocessing Module-Preprocessor using natural language processing techniques in documents

120: emergence of seed concepts context extraction module

121: tagged context-120 contextual information extracted through the module

130: generalized syntax information pattern generation module

140: new seed generation module

210: Preprocessing Module-Information preprocessor used for statistical methods to read documents

211: word table-word information that appears in the document

212: Sentence information-print out sentence by sentence

220: context extraction module-context information extraction module including candidate concepts

221: context

230: Association rule extraction module-Association rule generation module between candidate concepts and concepts

231: Association Rules

240: pattern clustering module

241: Association Pattern Information

250: Relationship Naming Module

251: extracted relationship information

The present invention relates to a technique for automatically constructing an ontology in a document, and more particularly to a technique for automatically extracting a concept and a relationship between concepts. In Korea, there is a technology for automatically constructing an ontology in an encyclopedia, but there is no technology in the field of automatically extracting ontology from specialized documents. The present invention has been started based on the technology that is going on abroad in this field. The automatic extraction of relations from documents has been studied in two ways: symbolic approach and statistical approach.

The symbolic approach to constructing a concept from a document uses grammatical elements such as the grammar of the sentence and patterns with specific meanings. For example, in a document such as 'dogs and other animals', we assume that 'dog' will represent a sub-concept of 'animal', and we make this a generalized pattern called 'As and other Bs' Find the parent-child relationship between the obtained entities of A and B.

NP ₀ such as {NP, NP, ..., (and | or)} NP-> hyponym (NP _s , NP ₀ )

such NP ₀ as {NP,} * {(or | and)} NP-> hyponym (NP _s , NP ₀ )

NP {, NP} * {,} (or | and) other NP _0- > hyponym (NP _s , NP ₀ )

NP ₀ {,} including {NP,} * {(or | and)} NP-> hyponym (NP _s , NP ₀ )

NP ₀ {,} especially {NP,} * {(or | and)} NP-> hyponym (NP _s , NP ₀ )

Hearst's work involved defining six patterns for extracting subrelations from a document and integrating the results with WordNet, as shown above. However, the extraction of the relationship by the pattern has a disadvantage that it is difficult to find the expected relationship due to the variety of language usage methods.

The Snowball system, developed by the University of Columbia, USA, provides a way to extract specific relationships from documents. In this system, concept pairs that describe relationships of interest to the user are used as examples to extract the context between concepts from the document in which the pair appears. The extracted contents are found through clustering, and the concept pairs are extracted using this pattern to extract concept pairs for a specific relationship. The result of the Snowball system is a set of concepts related to the pairs of concepts that make up a particular relationship from the document and the relationships that connect it, which can be used as important information in building an ontology. However, the use of this system requires that the accuracy of concept pairs affect the results in the selection of concept pairs used as input, and the classification of specific concepts must be clear. To do this, the Snowball system uses named entity taggers for people's names, agencies, and locations to extract clearly related concepts or relationships to entities. Therefore, this system is extremely limited in use and has the disadvantage that there is only one relationship extracted from one concept pair. In addition, related patterns for relationships created in Snowball are limited to apply in documents that have various variations to fixed patterns.

 The symbolic approach to constructing concepts from documents extracts concepts and relationships based on the statistical distribution of a document set and builds relationships based on them. The statistical distribution of a document set includes the frequency of words, the frequency of documents, and the dependencies between words. The statistical distribution is processed by applying a term-weighting scheme, and clustering and Correlation information is extracted through data mining techniques.

Developed by the Karlsruhe University and the FZI Research Center in Germany, the KAON system is a tool to help generate ontologies from documentation, and uses a statistical approach to discover the relationships between the ontology concepts. In a strict sense, the KAON system uses both approaches to ontology construction, but this paper focuses on the function of presenting the relationship between the concepts of the system. How to build ontology in KAON system is as follows. First, we use a low level parser to process the document and extract a noun clause containing more than one word to be used as a concept. It then looks for generalized association rules between the selected concepts and suggests the possibility of relationships between them. Finally, if there is an existing ontology, it extends the ontology by adding the relationships found in the ontology, or constructs an ontology from concepts and relationships. The KAON system judges whether to define a relationship as a user by presenting potential candidates as a relationship between concepts rather than directly extracting the relationship between concepts and providing them in the name of the support and confidence relationship of the association rule. And naming. The ontology generation method using this system requires human participation in the generation and maintenance of the ontology and does not provide intuitive information for explaining the relationship between the extracted concepts. That is, there is a problem that a lot of effort and time is required to generate the ontology using the KAON system. In addition, the association information found by the statistical method does not develop into a relationship, but has a problem of simply providing the association degree information existing between the concepts.

The present invention is to solve such problems, and in the symbolic method, the method focuses on a technique of automatically constructing syntax pattern information in order to solve the difficulty of designating a pattern to be constructed by an administrator. At this time, in order to solve the problem that the syntax pattern information to be constructed does not include various syntaxes that appear variously in the document information of the ontology building field, the syntax pattern information is generalized to find more relations.

In the statistical method, in order to overcome the limitations of the existing technology of simply providing the association degree information, the association pattern information is composed, and the related association pattern information is clustered to provide group information by grouping similar relationships. In addition, the relationship names of these groups are suggested to give more meaning to the relationships.

It is an object of the present invention to provide a hybrid combining method that combines ontology relationship information extracted by a symbolic method and relationship information extracted by a statistical method.

In the present specification, the document refers to various professional text documents in the field of ontology construction.

Furthermore, an object of the present invention is to enable to automatically construct an ontology applicable to any field, not an ontology of a specific field.

The present invention for achieving the above object is largely divided into a symbolic approach and a statistical approach, and provides a method of combining the results of this method. As shown in FIG. 1, the document receives the specialized documents (1) of the field to be automatically constructed and the candidate concept list (2) that can be used in the ontology, as shown in FIG. 1, respectively. The results extracted after the method module 20 are combined through the hybrid method module 30 and converted into an ontology 3 to automatically build the ontology.

The symbolic approach attempts to solve the problems described above based on the existing Snowball system. FIG. 2 is a flowchart illustrating a symbolic approach, which includes a document preprocessing module 210, an appearance content extraction module 120 of seed concepts, a generalized syntax information pattern generation module 130, and a new seed generation module ( 140).

The document preprocessing module 110 receives the professional documents 1 as inputs, extracts sentence information and part-of-speech tagging information in each word, and passes the extracted content of the next seed concepts to the module 120. At this time, a well-known bullytacker was used for tagging of parts of words.

The appearance content extraction module 120 of the seed appearance concepts extracts the appearance content of the seed in the document based on the result of the document preprocessing module 110 and the seed concepts 101 designated by the administrator. In this case, when two or more seeds appear in a sentence at the same time, it is determined as meaningful information and extracted as appearance contents. That is, the words appearing in a sentence can be considered to have a relationship with each other. Only information determined as meaningful appearance contents is output to the tagged context 121 and then passed to the next generalized syntax information pattern generation module 130.

여기에서 t는 tuple을 의미하며, concept1과 concept2는 후보 개념들로서 두 개념에 관련되는 정보를 패턴으로 기술한 것이다. lv, mv와 rv는 단어들의 벡터를 의미하는데. lv는 개념1과 개념2의 연관정보 왼쪽에 출현하는 단어들로 만들어진 벡터가 되며, mv는 두 개념 사이에 출현하는 단어들로 만들어진 벡터이며, rv는 두 개념의 오른쪽에 출현하는 단어들로 만들어진 벡터이다. 이들 벡터를 만드는 과정에서 문서에 출현할 수 있는 다양한 형태를 지원해주기 위하여 본 발명에서는 패턴의 일반화 작업을 진행하였다. 일반화 규칙은 아래의 표에 기술한다.The generalized syntax information pattern generation module 130 generates a generalized pattern based on the tagged document 121. The pattern is expressed as a vector of 5-tuples that holds the following five pieces of information:

Here, t means tuple, and concept1 and concept2 are candidate concepts that describe information related to the two concepts in a pattern. lv, mv and rv mean vectors of words. lv is a vector made up of words that appear to the left of the association between Concept 1 and Concept 2, mv is a vector made up of words that appear between the two concepts, and rv is made up of words that appear to the right of the two concepts Vector. In the present invention, in order to support various forms that can appear in a document in the process of making these vectors, the pattern generalization work was performed. Generalization rules are described in the table below.

Word / Pawn category example  is, are, am, was, were  BE  is-> [BE]  Noun  No change  Kodak-> Kodak  adjective  ADJ  pretty women-> [ADJ] women  adverb  remove  well->  Article  DT  the-> [DT]

According to the generalization rule above, the generalized syntax information pattern generation module 130 generates syntax information patterns and passes them to the new seed generation module 140.

The new seed generation module 140 receives syntax information patterns from the previous module and clusters the syntax information patterns using a single-pass clustering algorithm. The similarity calculation used at this time is performed using the following equation. The similarity calculation for comparing the pattern information of ts and the pattern information of tp is determined by the value of the following Match (ts, tp).

If the similarity value calculated in the above equation exceeds a certain threshold, it is expressed as the same cluster. Representative syntax pattern information representing the clusters in each cluster created by the above method is extracted by the vector sum method. At this time, based on the extracted representative syntax pattern, the actual document is searched again and the concepts applied to the representative syntax pattern are extracted. The concepts created at this time become concepts and concepts about the relationship to be found in the present invention. This information is combined with existing seed concepts, and iteratively performs symbolic approach to extract more relational information.

The statistical approach includes a preprocessor 210, a context extractor 220, an association rule extractor 230, a pattern clustering module 240, and a relationship naming module 250. The preprocessor 210 divides documents into sentence units in order to efficiently handle a large amount of documents, and expresses each sentence in a continuous form of IDs of words. At this time, the words extracted from the document find the root using KStemmer to improve the clustering performance, and use the B + -tree to maintain the information of the word-word ID.

The context extractor 220 generates a concept indexes structure before extracting the context. The concept index structure uses the ID of the starting word of a concept as an index. In addition, by maintaining a list of concepts starting with an index and IDs of the concepts, the searching time of the concepts appearing in the sentence is shortened. Also, in order to give priority to concepts containing long words, the list of concept IDs in the concept index structure is sorted in ascending order of the number of words included in the concept.

The association rule extractor 230 finds an association rule composed of two concepts from the context information 221 extracted by the context extractor 220. The association rules found in this process are the relationships between the important concepts that the document set describes. Since the apriori-algorithm proposed to find association rules aims to find the association of a product from a set of products for a purchased product, it allows to have multiple products in the conditions and result parts of the association rule. However, the present invention simplifies and uses the existing apriori-algorithm because the association rule consists of two concepts.

The pattern clustering module 240 finds a pattern describing the relationship between the two concepts from the content generated in the previous step and the association rule through clustering. The pattern mentioned here refers to the centroid of the cluster. For clustering, the present invention uses a simple single-pass clustering algorithm. Clusters created through this clustering process contain one or more contents. But each cluster must have objectivity in order to be a pattern for explaining the relationship between the two concepts. Therefore, if the number of contents included in the cluster is smaller than a certain threshold value, the cluster is removed to perform filtering on the cluster. Patterns that describe the relationship between the two concepts are found by clustering the contents of each association rule, which can complement the constraint that there is only a single relationship between the two concepts, which is a disadvantage of the existing SnowBall system. It can be a way. The association pattern information 241, which is the result of the pattern clustering module 240, is grouped into related association patterns by grouping the association pattern information, and a relationship name of an unnamed relation group 1 is given to them.

The relationship naming module 250 names the relationship between the two concepts by using the pattern and the contents generated by the above-described processes. The patterns found through the pattern clustering module 240 are specialized, each describing only the relationship between the two concepts, and can create many relationships that are not important. Therefore, the relationship naming module 250 performs clustering of patterns to find a generalized pattern as a previous step of naming to solve this problem. This clustering helps to find general relationships for patterns generated from similar content. When clustering is done to find a general pattern, a name is assigned to a relationship represented by each cluster. This task uses the centroid of a cluster to calculate the score for each piece of content in each pattern, and names the highest score in the name as a name that represents the relationship between the two concepts. This is a good way to provide intuitive information about the relationship between the two generated concepts by using the weights of the words in the general pattern and the sequence of words in the content of the pattern.

The hybrid approach focuses on combining the relational information extracted from the symbolic and statistical approaches. In general, it is customary to place more weight on the semiotic result with the epistemological part of the person than the statistical result, so the present invention also puts a higher weight on the semiotic result. Based on the symbolic results, the statistical results are combined. In the statistical result, unlike the symbolic result, the name of the relationship is not fixed and the candidate relationship name of each group is presented. Therefore, for each group in the statistical results, the concept information belonging to each group is compared with the symbolic results, so that the concepts belonging to a specific group in the statistical results have a high similarity with the specific relationship to the symbolic results. In this case, the relational name of the symbolic result is given to the group of statistical results. If this group does not belong, nameless relationship group1 forms the final ontology by naming the relationship in this form and stating the candidate group name presented by statistical method together.

The popularization of the web due to the development of the Internet has led to the active participation of machines in the sharing and retrieval of information, but most of the information on the web has been prepared for human understanding. This was a major obstacle to the machine's handling of user requests. As a way to solve this problem, a semantic web for expressing information in a form that can be understood by a machine has been proposed. However, building and maintaining the ontology that constitutes the logical foundation of the Semantic Web requires a lot of time and money. Therefore, the present invention invented a method for extracting a relationship between concepts, which is a part of ontology construction process, from a document set.

Furthermore, the ontology construction work, which required high cost and time, can be automatically switched by the expert's manual labor.

It is also expected to play a role in speeding up the coming of the semantic web on the Internet.

Claims (6)

In the method of automatically extracting the ontology relationship from the document, a1) a symbolic method of automatically extracting and using generalized syntax patterns; a2) a statistical method of clustering related information to build relationship related information and naming the relationship based thereon; a3) hybrid method combining symbolic and statistical methods; Ontology automatic building method comprising a. The method of claim 1, wherein the method a1) extracts representative syntax patterns that can represent various relationships from seeds for various relationships. The method of claim 1, wherein the method a1) generalizes the syntax pattern to apply the generalized syntax pattern. The ontology automatic construction method of claim 1, wherein the method a2) extracts the association information through the grouping of the association patterns with respect to the association rule. The ontology automatic construction method according to claim 1, wherein the method a2) assigns names to the relationships through the relationship association information. The method of claim 1, wherein the method a3) combines statistical results based on the symbolic results.

Images