Open Domain Question Answering With Multi-hop Reasoning

The traditional search engine can no longer well meet the user's information needs by returning multiple related document links according to the user's query,and the user is more inclined to provide a possible answer to their question directly by the information system.Multi-hop problem requires the model to infer along multiple intermediate entities to get the answer,which is more general.By studying the reasoning mechanism of multihop problem in open domain question answering,the following work is mainly included:A one-stage model combining knowledge graph and text information is constructed.Single-stage model uses the correlation degree between the problem and the relationship between entities to explicitly construct the inference path for the problem,and uses graph relation network to model the different relationships between entities and simulates the reasoning process of multi-hop problem.Furthermore,a multi-stage model is proposed,which uses the method of vector search,so that the retrieval and reading operations can be carried out alternately in the multi-hop reasoning process.Inverse file indexing and product quantization are used to pre-index the document vector.According to the new problem vector,the approximate vector search technique is used to obtain the new document and knowledge graph information related to the current context and update the problem subgraph accordingly,so as to overcome the dependence of single-stage model on the recall performance of retrieval module to some extent.According to the experiment on multiple data sets,the model has achieved high evaluation indexes on multiple data sets,and the influence of each parameter on the experimental results is compared.Experimental results show that the model can achieve high accuracy while ensuring real-time performance and low resource consumption.