Chinese Entity Relation Extraction Based On Multi-Agent Strategy

With the development of the technology and www, the text information is dramatically increasing. To extract useful information from multiple texts is an important work. The relation extraction is to identify if there is a relationship between two entities and to identify the relationship. The entity is the information that extracted from the text, such as the time, the person name, the organization. The relation between entities is predefined, such as the part-whole, the employment and so on, The relation extraction plays an important role in information retrieve. And it is an important part of the Q & A systems, the information extraction systems and machine learning systems.In this paper, an ensemble kernel method is present first, which combines the feature-vector based method and the shortest path tree (SPT) based kernel method. The ensemble kernel method is used to extract the Chinese entity relationship. The experiment is conduct on ACE RDC 2005 corpus. The F-score of the ensemble kernel method is 68.5%, which is higher than the single kernel method by 4.36% and 17.37%. It proves that the ensemble kernel method is better than the two single methods, and the two single methods are complementary. This paper compares the method with other kernel based relation extraction systems, and it proves that the ensemble kernel method that combines the feature-based kernel method and structure kernel method is better than the system that adds the entity information into the tree kernel as the tree nodes.Then, this paper incorporate the multi-agent strategy into the relation extraction system to solve the problem that all relation types use one global feature set and that feature set cannot cover the character of each relation types. Firstly, this paper chooses different sensitive feature set and the learning classifier to form the base agent for each relation types; then the communication between basic agents is conducted to optimize the result of each agent; at last, this paper combines the result of the basic agents to form final result of the relation extraction. The experiment is conduct on ACE RDC 2005. And the F-score is 69.72%, which is higher than the F-score of the ensemble kernel method by 1.22%. The experiment result shows that the multi-agent strategy is an effective method that can improve the relation extraction. At last, when compared with other relation extraction systems, the Multi-Agent system presents better performance. And it proves that the Multi-Agent strategy is effective in relation extraction.