Research On Entity Attribute Extraction Based Upon Semantic Analysis

Entity Attribute Extraction(Slot Filling task in TAC-KBP)aims at extracting the specific attributes of a given entity in a large scale corpus.Two fundamental stages are included: retrieving relevant documents of the given entity from the corpus and extracting attributes from the relevant documents.Our research addresses the name ambiguity issue in retrieval stage and the challenge of lack of hand labeled training data in extraction stage.(1)Research on Co-reference Resolution based Entity Archiving ApproachThe basic retrieval model used in Slot Filling is subject to the name ambiguity issue and leads to a low precision result.Hence,we propose a cross document co-reference resolution(CDCR)based retrieval approach.Firstly,we use pseudo-relevant feedback to collect documents as supplementary of the reference document.Secondly,CDCR is used to filter irrelevant documents from the basic retrieval result.Experimental results show that our approach improves the performance of document retrieval.(2)An Attribute Extraction Approach based on Definition Semantic ConstraintsFocusing on the challenge of lack of labeled training data for attribute extraction,we propose a novel approach considering the semantic constraints within attribute definition.We take the definition of an attribute as constraints to slot provenances.Features describing the semantic relevance to attribute definition are employed to classify the attribute provenance,from which we extract the attribute value.Evaution results demonstrate that our approach outperforms the baselines.(3)Refining Attribute Extraction based upon Information SelectionKeep on refining the definition semantic constraints based attribute extraction approach by focusing on “mention overlap” problem.We firstly conduct syntactic tree pruning to filter irrelevant semantics.Secondly,we do key word identification to retain key semantics.Thus,we make the test sentence more concise and precise.Experimental results prove the effectiveness of these improvements.