Research On Open Knowledge Graph-based Question Answering System For Metallic Materials Domain

At present,the question answering(QA)has been applied to many specific domains,such as the medicine domain and the geography domain.With the scale of entity and relation in the open knowledge graph becoming larger,the open knowledge graph can provide richer information for QA in the specific domain.However,the information query about specific domain based on open knowledge graph is faced with some challenges.The data in the open knowledge graph do not specify which domain it belongs to.Additionally,the user's natural language query is structurally inconsistent with the domain knowledge in open knowledge graph.This paper proposes a domain-specific question answering method based on open knowledge graph.The proposed method designes a series of logical rules to transform the semantic parsing results of the user's question into the question triples.And the concepts in the triples are achieved multi-angle semantic expansion based on WordNet and Microsoft Concept Graph.The semantic expansion of a triple is realized according to the semantic expansion results of the concepts in this triple.And then the matching relations between the expanded triples and the specific domain knowledge are built.The main contributions of this paper are as follows:1)The logical rules are designed to transform the question into a kind of structured representation.The NLP tools Stanford Parser and Stanford CoreNLP are used for parsing the user's question.And according to the designed rules,the parsing result is integrated into the triple representation,realizing the structured representation of the user's question.2)The semantic query expansion strategy and the expansion result's filtering strategy are designed for the query expansion.The concepts in the question are expanded from multiple semantic angles(e.g.synonym,hypernym and hyponym)based on WordNet and Microsoft Concept Graph.And the different filtering strategies are designed for the expansion result of each semantic angle,based on the similarity calculation methods.3)The matching result is verified based on semantic tag information.Besides the similarity value,the semantic tag information is also used to filter the matching results of the concepts in the question.The combination of the similarity value and the semantic tag information can further improve the precision of the concept matching.4)A query relaxation strategy based on ontology structure is proposed for the unsuccessful matching.According this strategy,there conducts a new matching betweenthe concept in the question triple and the resource of designated type in the domain knowledge.Additionally,the similarity value threshold,which is used for judging whether a concept is matched with a certain resource in the domain knowledge,may be relaxed.Therefore,this proposed query relaxation strategy could make more matching relations between the question and the domain knowledge built.