Two-stage Document Filtering and Asynchronous Multi-granularity Graph Multi-hop Question Answering
Multi-hop question answering aims to predict the answer to a question and the supporting facts for the answer by reasoning over the content of multiple documents.However,current multi-hop question answering methods aim to find all documents related to the question in the document filtering task,without considering whether all these documents are useful for finding the answer.Therefore,we propose a two-stage document filtering approach.In the first stage,the documents are scored and a small threshold is set to obtain as many relevant documents as possible to ensure a high recall of documents.In the second stage,the inference path of the question answer is modeled,and the documents are extracted again based on the first stage to ensure high accuracy.In addition,we propose a novel asyn-chronous update mechanism for answer prediction and supporting fact prediction for multi-granularity graph composed of documents.The proposed asynchronous update mechanism divides the multi-grain graph into heterogeneous and homogeneous graphs to perform a-synchronous updates for better multi-hop inference.The performance of the proposed method is better than that of the current mainstream multi hop question answering method,and the effectiveness of the proposed method is verified.
国家科技期刊平台
NSTL
万方数据
维普
