Research On Machine Reading Comprehension Model Based On Curriculum Learning

Machine reading comprehension is one of the important topics in the field of natural language processing.Machine reading comprehension technologies are involved in intelligent question answering,dialogue generation,text summarization,and other fields.At present,the training of machine reading comprehension models generally uses random sampling and does not consider the influence of the sequential organization of sample sentences on the model.In machine reading comprehension tasks,the difficulty of training examples may be different.For example,some examples can be solved by simple vocabulary retrieval,while others require complex semantic reasoning.Then the weights of samples with different difficulties should be different during training.In order to solve this problem,some researchers have introduced the idea of Curriculum Learning(CL)into machine reading comprehension tasks.Curriculum Learning is to sort the examples from simple to complex,and then present them to the model in a certain order during training.When using technology of CL to learn modeling,you need to consider how to define the difficulty of samples,and select the sample based on the difficulty.However,as far as I know,In machine reading comprehension tasks using technology of CL for the definition of sample difficulty are generally based on data sets,and the relationship between the sample difficulty and the model is not fully considered,which leads to the difficulty evaluation highly depending on the data set.Based on these,this thesis uses a cross-review method to construct a difficulty evaluation method for reducing reliance on data in difficulty evaluation.The cross-review method does not fully consider the relationship between the difficulty of the sample and the model,so this thesis further introduces the Item Response Theory(IRT)to model the relationship between the difficulty of the sample and the model.This thesis validates and analyzes the machine reading comprehension model based on Curriculum Learning and IRT method on the RACE data set.The experimental results show that the introduction of the idea of Curriculum Learning can improve the accuracy of the machine reading comprehension model.At the same time,combining the IRT model on the basis of Curriculum Learning further improves the performance of the model.Experimental results show that the effective combination of IRT model and Curriculum Learning can effectively improve the performance of the machine reading comprehension model.