Research On Extractive Machine Reading Comprehension

In recent decades,with the development of artificial intelligence theory and technology,the task of machine reading comprehension(MRC)has received more and more attention from all circles.Machine Reading Comprehension means using computers to model algorithms,which requires machines to have similar reading ability like humans.It can answer questions by reading and browsing text,analyzing semantics,and logical reasoning.This task can measure the machine's understanding of natural language,which has important research significance and application value.At present,most traditional machine reading comprehension models use an end-to-end "coding-interaction-prediction" framework to simulate the process of human reading,which have achieved good performance on some public datasets,but the research has found that these models only contain a single-layer interaction between the text and the question,so there will be insufficient interaction between them and impossible to dig deeper into the relationship between them.Moreover,in real scenarios,there are few or even unavailable marked data,and the cost of marking unmarked data is very high,which is also a difficulty in the task of machine reading comprehension.So how to train the model in the absence of data is a huge challenge.To solve the above problems,this paper proposes a machine reading comprehension model based on a multi-level aggregated attention mechanism,it can enhance the depth of interaction between the text and the question,and a semi-supervised machine reading comprehension framework based on few samples to solve the problem of insufficient training data.The main work and contributions of this paper include:(1)A machine reading comprehension model based on a Multi-Level Aggregated Attention Mechanism(MLAA)is proposed.This model simulates the idea of human reading comprehension.It not only uses the aggregate attention mechanism to extract important features of text and question interaction from different angles,in order to compensate for a single attention mechanism,but also uses a multi-level interactive structure to deepen the interactive representation between the text and the question,and mines the deep association between them,which provides more effective information for the final answer prediction.The experimental results show that compared to other traditional machine reading comprehension models,MLAA has achieved better performance on the SQuAD.(2)A Semi-Supervised machine reading comprehension Framework based on few samples(SSF)is proposed.When machine reading comprehension data is scarce,how to introduce effective information from a large amount of unmarked data to assist model training is crucial.Based on this,the framework uses the semi-supervised idea,and on the basis of few marked samples,it adds the regularization loss of unmarked data(The gap between the predicted answer with unmarked data and the answer predicted by the enhanced data)to the total loss function,which improves the ability of the machine reading comprehension model to learn with few samples.This paper extracted 10 texts and 100 texts from the SQuAD to construct a small amount of marked dataset Mark-SQuAD(containing 2764 data)and a large amount of unmarked dataset UnMark-SQuAD(containing 17407 data).The experimental results show that the SSF framework is effective in machine reading comprehension tasks under the condition of few samples.