Research On Key Technologies Of Grammatical Error Diagnosis And Argument Reasoning In Chinese Auxiliary Learning

Chinese Auxiliary Learning is a research hotspot and difficulty in Chinese natural language processing.In recent years,the increasing number of Chinese learners has increased their application value.Numerous Chinese learning scenarios have spawned various sub-problems such as grammatical error diagnosis,sentence analysis,and argument reasoning.From a linguistic point of view,these problems can be divided into two categories:grammar and semantics.This thesis chooses grammatical error diagnosis and argument reasoning as the representative of grammar and semantics.It summarizes the deficiencies and difficulties of current research status,and proposes improvement methods for them.The main work is as follows:(1)In view of the difficulties in the lack of training data in grammatical error diagnosis,this thesis proposes a pre-training method based on transfer learning to reduce the dependence of the model on training data.In view of the same difficulty,in the argument reasoning,this thesis adopts the method of transfer training data from other languages.(2)In view of the difficulties in the complexity of Chinese,in the grammatical error diagnosis,this thesis proposes a self-attention mechanism based on the bidirectional language model to obtain the vector expression of the sentence and enhance the expression ability of the model.The difficulty of the complexity of Chinese still exists in argument reasoning.In this thesis.,the dot product attention mechanism is used to improve the encoder,increase the quantity of information,and make the subsequent similarity calculation results more accurate.(3)In view of the difficulty in the great difference between the number of errors and the correct position in grammatical error diagnosis,this thesis uses some improved decoders(conditional random field,weighted softmax)to improve the decoding performance.Based on the above improvements,this thesis designs new models for grammatical error diagnosis and argument reasoning.Experiments show that the two new models outperform some traditional practices in their respective key evaluation indicators.