| Question answering system(QA)is a system used to answer natural language questions raised by humans.It aims to correctly return the required answers in the system based on questions,and is widely used in the fields of information retrieval and information extraction.In response to the fact that China has a large population,complex social relations,and an endless stream of various civil and criminal cases,this thesis has carried out relevant research around the process of QA,and built a QA for legal judgment documents based on deep learning to help legal personnel understand the case and make decisions.The work of this thesis is mainly composed of the following parts:1.Due to the large number of legal judgment documents,it is not realistic to find the answer directly from the document database based on the question.In response to this situation,a document retrieval method based on the diversity model is proposed,which combines the siamese recurrent neural network and the TF-IDF algorithm.The siamese recurrent neural network model can extract deep semantic information.The siamese structure can reduce network parameters,and the recurrent neural network has certain advantages in the extraction of context information.The TF-IDF algorithm can extract the shallow features of the text.The fusion of the siamese recurrent neural network and the TF-IDF algorithm can more effectively recall documents with high relevance.Related experimental results prove the effectiveness of the document retrieval method based on the diversity model.2.In view of the complex reasoning questions in legal judgment documents,a complex reasoning question parsing method is proposed.This method comprehensively considers the influence of entity words and relationship types on question sentences,and combines part-of-speech analysis and syntactic analysis to achieve the effect of simplifying sentences.Among them,the relationship extraction model is based on the siamese model,and the adversarial training is added,which can effectively classify the relationship types.The experimental results show that the parsing method of complex reasoning questions can improve the effect of the QA.3.In view of the fact that there are unanswered questions in legal judgment documents and some questions are general questions,an answer extraction model with a branch of question type judgment is proposed.Because BERT has powerful feature extraction capabilities,the BERT model is used as the basic answer extraction model.In order to improve the answer mechanism,the question type judgment branch and multi-task joint training are introduced.Related experiments prove that the answer extraction model with the branch of question type judgment can effectively improve the answer mechanism of the QA,and help to improve the effect of the QA. |
