| Reading comprehension requires understanding not only what is stated in a text,but also what is between the lines.In the field of natural language processing,many pre-trained models have been born that learn from the open corpus with profound knowledge,thus improving performance on multi-domain datasets such as Q&A and intelligent conversations.However,enabling machines to understand common knowledge about life situations is less effective.This is because existing models fail to introduce and integrate common sense knowledge to achieve self-understanding of the machine itself.To address this problem,this thesis conducts research on machine reading comprehension based on common sense knowledge of life situations.As one of the key tasks to achieve intelligent machine comprehension of common sense knowledge,common sense of life situations originates from stories in life,from which problems are found and the degree of correlation with options is discovered.Therefore very much test the machine’s ability to understand,analyze the semantics and partially reason about the life context knowledge.Around this relevant content,this thesis mainly carries out the following research work.(1)A free-form polynomial Chinese machine reading comprehension dataset is selected for polynomial machine reading comprehension of common knowledge of Chinese life contexts,which is published by international students from Cornell University.In this thesis,the BERT model is used to improve the input,coding and output layers on the base model.The reliability is finally verified on the test set by marking the context,questions and answers as a new sequence through serialization in the input layer,adding a multi-headed attention mechanism to the context,questions and candidates in the interaction layer,and modifying the occupancy of the mask mechanism.The experimental results show that the improved model on the base model achieves 62.28%、63.33%and 63.07%accuracy in C_D~3,C_M~3 andC~3.After modifying the percentage of the masking mechanism,the full word masking presented a better result than the masking mechanism of 15%in terms of accuracy by 0.95%,2.52%and1.77%,respectively.(2)For polynomial machine reading comprehension of common sense in English life situations,Cosmos QA,a large-scale dataset for everyday narrative stories,and SWAG,a dataset rich in foundational common sense reasoning,are chosen to verify the generalization ability of different methods on the dataset.In this class of tasks,this thesis uses the Co-matching model based on semantic matching for semantic matching,the DCMN~+model for dual collaborative matching networks,and the pre-training models GPT,BERT and the fused multi-headed attention mechanism BERT model.The generalization performance of the fused multi-headed attention mechanism BERT model is further demonstrated by validating different methods on the dataset. |
