Research And Implementation Of Commonsense Reasoning Technology Based On Knowledge Fusion

With the continuous development of intelligent technologies,question answering techniques(Q&A)based on machine reasoning have been widely used in multiple fields,which provides a convenient form of human-computer interaction through semantic understanding of contextual information.However,in the commonsense Q&A scenarios,existing methods still face limitations in understanding,expressing and reasoning about human commonsense,since the contextual information is not explicitly included in the question text.Therefore,this thesis addresses the key issues such as the lack of commonsense,weak reasoning ability and poor interpretability of existing models.Besides,this thesis investigates knowledge fusion techniques for multiple-choice commonsense Q&A tasks with a large-scale pre-training language model as the benchmark.Specifically,the main research work of this thesis is as follows:(1)We propose a pre-training model fusion inference method based on knowledge attention.This thesis focuses on knowledge retrieval methods for Wikipedia commonsense knowledge texts,Wiktionary concept interpretation,and external commonsense Q&A training sets.Besides,we propose a knowledge purification method based on knowledge contribution degree strategy,and an enhanced inference method based on knowledge attention.A multi-source knowledge fusion model is designed using the above methods with the pre-training model ALBERT.Experiments demonstrate that this method can effectively fuse multi-source external knowledge,improve the model’s ability to characterize contextual semantics,and thus significantly improve the model’s commonsense reasoning ability.(2)We propose a graph neural network-based method for joint inference of structured knowledge.This thesis focuses on the extraction method of structured knowledge,graph knowledge representation method and context fusion inference method.In the semantic encoding stage of the pre-training model,a knowledge extraction method based on relational weights is proposed to output a complete representation of knowledge jointly with unstructured knowledge.In the encoding stage of graph neural network,a knowledge representation method based on improved graph convolutional network is proposed to obtain structured graph embeddings by aggregating nodes and relational feature information.Experiment results illustrate that combining the contextual information encoded in the above two stages can achieve an effective fusion of semantic space and symbolic space information,which proves the advantages of this method in interpretable reasoning.(3)A knowledge fusion-based commonsense knowledge Q&A demo is implemented based on B/S architecture.The system uses the Q&A model as the core algorithm module to assist users to answer commonsense questions and provide valid inference evidence information.Through functional testing and Q&A system demonstration,it is verified that the system has application potential in the field of knowledge Q&A.