Reading Comprehension Model Based On Two-way Attention Mechanism And Conditional Random Field

Teaching machines to understand human language is an elusive task,which is a longterm challenge for artificial intelligence.At the same time,it has become an important research field in industry and academia.How to build an artificial intelligence system to read and understand texts and answer questions is the core task of natural language processing,and high-performance reading and understanding system will be one of the key technologies in question answering and dialogue applications.In recent years,with the rapid development of artificial intelligence technology and the full promotion of computing resources,artificial intelligence has shown a strong vitality,and the use of neural networks to achieve machine reading comprehension model is the current popular trend,at the same time,the task of machine reading comprehension needs to have a strong ability of natural language understanding and semantic analysis.At present,most of the mainstream models are based on hierarchical network structure,and adopt different strategies in different layers.They are presentation layer,fusion layer and result layer,and fusion layer is the most important component structure.In this paper,the algorithm of reading comprehension model is improved as follows:(1)The model of machine reading comprehension in this paper is based on dynamic word vector.At present,the traditional word vector is static word vector,which is based on context free semantic features.It can't solve the problem of polysemy of a word.At the same time,it can't generate different word vectors according to the context.For this reason,this paper uses the dynamic word vectors obtained from the current pre training language model training,and the final results show that the effect of the presentation layer on the final model is improved.(2)In the fusion layer,the model in this paper uses the two-way attention mechanism to deeply fuse the text and the problem code.Compared with the traditional interaction layer,it is simple and the fusion has weak correlation,which results in the poor effect of the model.This paper begins to integrate the two-way attention mechanism,and the model of this paper also introduces the self-attention mechanism to further strengthen the text's table specific ability.The final results show that the interaction of the fusion layer has a higher ability of language understanding,and can improve the effect of model prediction.(3)Finally,on the output layer of the model,we use conditional random field to predict the model,and at the same time,we test on two open data sets,one with answer and the other without answer.The final results show that the model has improved in the actual effect.