Research On Concept And Relation Extraction Of Chinese Domain Ontology

As a shared conceptual model, Ontology plays an increasingly role in many areas.Such as artificial intelligence, knowledge engineering, information retrieve, semanticweb and so on. However, the manual construction of ontology is a complex taskwhich takes considerable time and cost, and needs experts’ participation. It can nolonger meet the needs of ontology’s applications. Therefore, the ontology learningtechnology, automatic or semi-automatic ontology construction, is becoming a hot spotof research now.Ontology learning obtains expectant ontology in an automatic or semi-automaticway from existing data sources. Always statistics, machine learning, natural languagetechnology and many other area of science can be used. The main task of ontologylearning consists of automatic or semi-automatic acquisition of every element containedin ontology. Currently, research on this area focuses on concept extraction and theirrelations extraction. Statistics-based approaches are mainly adopted in the existingontology learning. These methods combine domain relevance and domain consensus toextract the concepts, and use association rules method to extract the relation pairs. Mostof the methods are based on extracting concepts which are high-frequency in thedomain. The result contains more redundancy, and poor accuracy.To solve problems above, the statistics of log-likelihood ratio is introduced, andthen filter the extraction result of tradition method to improve the accuracy. In theprocess of concept extraction, we use a method which combines the domain relevanceand domain consensus to get the initial concepts in the field, and then use thelog-likelihood ratio to measure the domain’s importance of concepts, filter redundantconcepts and get the final concepts. In the process of non-taxonomical relationextraction, we use the log-likelihood ratio method for extracting semantic relationswhich based on using the association rules to acquire concept pairs, and try to get therelational labels. By using VFICF (Verb Frequency-Inverse Concepts Frequency) metricto extract domain verbs as candidate relation labels, and using log-likelihood ratiomethod to map between relation labels and concepts, we obtain non-taxonomicalrelations.In addition, the method for extracting taxonomic relation is discussed. First, thismethod uses the context of concepts to construct the Vector Space Model, and uses thesemantic similarity obtained by the cosine between vectors to represent the distance oftwo concepts. According to the spirit of hierarchical clustering, an extraction method based on the minimum spanning tree is used to get the taxonomical relations.In order to verify the model’s validity, an ontology learning system is constructed.This system compares traditional ontology learning method and Ontology learningmethod based on log-likelihood ratio filtering. Experimental results show that theimproved model can effectively improve the precision of concept extraction and relationextraction. So it verifies the proposed learning method is effective.