Research On Question Answering System Based On Entity Relationship

Entity relationship extraction is an important branch of information extraction technology,which refers to the extraction of entities and the relationship between them from the free text.With Internet data expanding rapidly,the entity relationship extraction technology can transform it from unstructured data to semi-structured or structured data,which is of great significance for information retrieval and reuse.Based on the web search engine service,this paper builds an entity-based question answering system and makes an in-depth study on several core issues involved.The main work and contributions are as follows:(1)With the help of web crawler and information extraction technology,set up a QA system of entity relations.Based on four core modules of front-end,search and recall,relationship classification and database,achieve two kinds of solution to answer queries,which are local search and online retrieval respectively.(2)Put forward a more suitable solution to the trigger word discovery problem.After analyzing the insufficiency of the pre-method based on the word activation,we pointed out the consistency between the trigger discovery task and the keyword discovery task.The two methods of tf-idf and chi-square detection were migrated to this task and achieved a good filtering effect.(3)As for the relationship classification,we put forward the instance-adaptive attention mechanism.It improves the previous attention mechanism in the relationship classification.By using CNNs of multi-length to generate deciding factors from the shortest dependency path of entities,it enhances the explanatory and robustness of the model and achieves more excellent results on the data set.(4)Propose a new task of generating relation samples,interpret it as a labeled text generation problem and propose preliminary solutions.Labeled text generation provides less input information.This paper explores how to introduce label information effectively and proposes two training approaches based on cross-entropy and enhanced learning theory respectively.