Research And Implementation Of Commonsense Reasoning Technologies Based On Multiple Knowledge Fusing

The improvement of software and hardware performance of contemporary computers makes it possible to train complex models based on massive text data.In the field of natural language processing,a series of pre-training models have been born.These models have learned a wealth of knowledge information from huge open corpora,thus significantly improving the performance of multiple task datasets in machine translation,intelligent dialogue,question answering and other fields.However,when they are used for commonsense reasoning tasks,the performance is not very good.This is because none of these pre-trained models themselves fail to incorporate and incorporate common sense knowledge for text reasoning.Aiming at this problem,this thesis conducts research on the key technology of commonsense reasoning based on the fusion of multiple knowledge bases.As one of the key verification tasks to realize intelligent reasoning,common sense question answering needs to combine external knowledge to find the relationship between questions and options through knowledge information.Therefore,it is very important to test the model's ability to acquire external knowledge and integrate knowledge for reasoning.Around this point,this thesis mainly carries out the following research work:(1)The fusion of multiple knowledge bases is realized and a standard commonsense question answering research framework is proposed.By analyzing the respective characteristics of multiple knowledge bases,entity alignment is carried out between two structured knowledge bases based on semantic similarity matching.Then,according to the three stages of knowledge acquisition,transformation and utilization,a standardized commonsense question answering research framework is designed based on the pre-training model ALBERT.Experiments demonstrate the effectiveness of the strategy of integrating knowledge bases and the normative research framework.(2)An adaptive key knowledge acquisition method is proposed and additional corpus is introduced to enhance text semantic representation.The knowledge acquisition method combines two strategies of retrieval and generation to acquire key evidence knowledge.On this basis,dictionary information is introduced to help the model understand the difference between different options under the same relationship,and at the same time,wiki argument triples are used.Further enhance the semantic representation of text.Finally,based on the multiple-choice model of Transformer open source,a common sense question answering model based on text semantic representation enhancement is constructed.Experiments demonstrate the effectiveness of adaptive knowledge acquisition algorithms and methods for enhancing text semantics.(3)A text reasoning method based on an attention mechanism with option-aware networks is proposed.By constructing different encoded text sequences and corresponding reasoning networks,the fusion reasoning of knowledge and task texts is studied and realized.Finally,a text inference module is added to the multiple-choice model,which further improves the inference performance of the model on the basis of(2).(4)Based on the research work of this thesis,a scientific knowledge question answering system is implemented.The system uses the model proposed in this thesis,based on the B/S architecture,to achieve a complete question and answer function,and with a friendly user interface,to meet the actual application requirements.