Design And Implementation Of Question Answering System Based On MRC

In MRC-based question-answering systems,the "Retriever-Reader"architecture is an effective approach that consists of two components:the retriever and the reader.The retriever is responsible for retrieving candidate texts from a large text corpus that may contain the answer,while the reader analyzes and comprehends these texts to provide the answer.This system has wide applications in natural language processing,such as intelligent customer service,search engines,and intelligent assistants.This article designs and implements a "Retriever-Reader" architecture based on MRC for question-answering systems.The main work includes:(1)For the Retriever,this article uses named entity recognition and word vector search to construct an entity linking method to extract texts that may be related to the answer from the knowledge base.The experimental results on the Dureader_checklist dataset show that the retrieval performance is improved compared to the traditional Elasticsearch method.(2)For the Reader,based on the MRC task,the baseline model XLNet is adopted,and the TF-IDF method is used for data augmentation and highlighting of text.The experimental results on the Dureader_checklist dataset show that the model performance is improved.(3)The design and development of the MRC-based question-answering system are completed.Through the requirements analysis of the questionanswering system,the Chinese Wikipedia is used as the knowledge base,and the MRC-based question-answering method is implemented as the core design,deploying the system to TorchServe.The system is tested for both functional and non-functional requirements,and the test results show that the system’s overall performance meets the expected performance standards.