Research And Application Of Multi-Choice Reading Comprehension For Samll Samples

In the intelligence of human beings,understanding natural language and obtaining effective information is one of the basic abilities of human beings,and the task of machine reading comprehension is derived from it.This task provides articles and questions for the machine,and evaluates the machine’s understanding of natural language through the answers.Compared with other types of reading comprehension tasks,multiple-choice reading comprehension needs to summarize,summarize and reason the original text in order to obtain the correct answer.Therefore,it has higher requirements for semantic understanding and has important research value.However,in practical application,it is difficult to obtain large-scale and high-quality data because the field involved in machine reading comprehension task is relatively limited.Without enough training data,how to improve the performance of reading comprehension model in the field of small samples has good theoretical research and practical significance.Based on the traditional machine reading comprehension model,this thesis introduces open domain data as the source domain for cross domain knowledge transfer to the small sample domain,and combines the machine reading comprehension model with two knowledge transfer methods:domain adaptation and knowledge distillation to solve the multi-choice reading comprehension task in the small sample domain.The main work and innovations of this thesis are as follows:Firstly,a reading comprehension model based on semantic interaction and domain adaptation(SIDA)is proposed.Firstly,from the perspective of improving the semantic understanding ability of the model,a reading comprehension model under semantic interaction(SI-MRC)is proposed.Based on human reading strategies,SI-MRC model fully interacts with the semantic relationship among articles,questions and options in multi-choice reading comprehension.Secondly,SIDA model takes the open domain data set as the source domain to migrate knowledge in the small sample domain.Starting from the different data distribution between the source domain and the target domain,the common low-dimensional common features are realized by sharing the embedded layer.At the same time,the featurebased domain adapter is introduced to align the high-level features of the source domain and the target domain.The experimental results on relevant data sets show that the accuracy of SIDA model in two types of knowledge transfer tasks is 62.2%and 70.2%,which is 8.0%and 16.0%higher than the baseline model.At the same time,the ablation experiment verifies that semantic interaction can improve the semantic understanding ability of the model,and the domain adaptation method can be effective for small sample domain reading comprehension tasks.Secondly,in order to further enhance the auxiliary effect of open domain on small sample domain,based on the previous work,this thesis proposes a reading comprehension(MLKD)model based on multi-level knowledge distillation.Firstly,aiming at target distillation,a target distillation method integrating domain weight is proposed.The samples of source domain and target domain are modeled to obtain domain correlation,and domain correlation is used as knowledge fusion weight to balance the knowledge transferred from source domain to target domain,so as to carry out knowledge fusion more pertinently.Secondly,for feature distillation,a feature distillation method based on semantic interaction flow is proposed.The semantic interaction flow in reading comprehension task is abstracted and defined,and described by semantic interaction flow matrix.By minimizing the semantic interaction flow matrix corresponding to teacher model and student model,the student model can learn the specific solution process.The experimental results show that the accuracy of MLKD model on the two tasks is 64.5%and 72.4%,and the ablation experiment verifies the effectiveness of the two different levels of distillation methods on small sample domain reading comprehension tasks.