Abstract:Domain ontology is an efficient and reasonable display form of domain concepts and their relationships. In the process of building domain ontology, the problem often encountered is that ontology concept is complete but concept relations are complex and diverse and artificial tag cost too much. Using unsupervised Relation Extraction algorithm on rich Web information related with domain Ontology concepts solve previous problem. But the traditional method based on unsupervised learning does not take into account the situation of a single sample with more concepts, leading to the final incomplete results. We used Web information in trafffic field to construct ontology, introduced the weight of sample concept relation pair to a traditional unsupervised learning approach-Kmeans to solve this problem and achieved good results through experiments.