Multiple-choice Question Answering Based On Commonsense Knowledge

With the rapid development and popularization of the mobile Internet,question answering systems have been successfully implemented in many industrial fields and have achieved good economic benefits and social value.Commonsense knowledge,as the research focus of massive cognitive information,plays an important role in question answering systems.It is also constantly highlighted and has great research prospects.The Commonsense Question Answering(CQA)task is to study how to obtain relevant commonsense knowledge,and predict correct answers through question understanding and knowledge reasoning.This thesis focuses on the research of CQA tasks in the supervised and unsupervised scenarios.At present,the following problems in CQA tasks need to be solved urgently:in supervised scenarios,the current research work focuses on optimizing and improving the knowledge reasoning strategy of the model,while ignoring problems such as insufficient knowledge coverage and knowledge noise,which will seriously damage the model’s degradation in the knowledge reasoning stage resulting in prediction bias.In unsupervised scenarios,most current research methods focus on designing manual rules for specific tasks to improve the quality of generated knowledge,which leads to restricted knowledge types and weak transfer capabilities between different tasks.Therefore,the thesis makes deep explorations and studies regarding these challenges,and summarizes the work as follows:1.Aiming at insufficient knowledge coverage and noise problems in supervised CQA tasks,the thesis proposes a knowledge-enhanced graph contrastive learning model(KE-GCL).First,the model integrates the contextual descriptions for the entities in the QA pair into the corresponding knowledge subgraph,to achieve multi-source knowledge fusion;then,the model proposes an adaptive sampling strategy to generate the knowledge-enhanced view of the current subgraph,and simultaneously constructs the positive and negative graph pairs;finally,the model performs edges’ scattering and nodes’ aggregation to update and reason over the knowledge graph.2.To solve the problem of restricted knowledge types and weak transfer capabilities in the unsupervised CQA task,the thesis designs Prompt-based Knowledge Generation Network(PKGN).First,the model performs unsupervised contrastive learning through the Dropout enhancement strategy to capture the subtle differences between questions and learn better representations;then,the model uses instruction prompts to generate question-related knowledge statements;finally,the model leverages the text matching model to perform knowledge reasoning and answer prediction.In this thesis,a large number of experiments have been carried out on three commonsense question answer datasets,including CommonsenseQA,OpenbookQA,and SocialIQA.Through quantitative analysis and qualitative comparisons,the feasibility and effectiveness of KE-GCL and PKGN models in supervised and unsupervised directions are further verified.The experimental results show that the proposed models,with good robustness and generalization capabilities,in the thesis are constantly better than the current baselines in both supervised and unsupervised scenarios.