Research On Multi-task Learning Algorithm In Machine Reading Comprehension

Machine reading comprehension is a very important research direction in the field of natural language processing,especially in the realization of human-computer interaction is an extremely critical technology.Because of its complexity,machine reading comprehension has always been an important aspect of evaluating machine natural language understanding.At the same time,due to the rapid development of the times and the continuous accumulation of various text information,how to use these text resources to make the question and answer process more realistic is also a topic with practical application value.In the actual question answering system,to complete the question and answer of the whole field,a large amount of knowledge information needs to be involved,and the collection of this knowledge information requires years of accumulation.The problems faced by machine reading comprehension and human reading comprehension are similar However,in order to reduce the difficulty of the task in reality,many machine reading comprehension data sets currently studied have certain limitations,and the data samples are relatively small.When there is not enough training data in the target domain,how to use the data outside the domain to enhance the model effect,that is,the multi-task learning method to enhance the model performance of the main task is very important.Therefore,the study of multi-task learning methods in machine reading comprehension is of great signnificance.Recently,the rapid development of deep learning has continuously brought new breakthroughs to machine reading comprehension technology.This paper studies the difficulties in segmentation-based machine reading comprehension tasks,and proposes a method that combines machine reading comprehension models and multi-task learning to solve this problem.The main work and contributions of this article include:First,in the task of machine reading comprehension in the general field,a method based on the traditional machine reading comprehension model and multi-task learning with soft constraints is proposed.The model uses a multi-level context representation encoder with different granularities for text representation.The attention mechanism is responsible for connecting and fusing the information in the context and question words.The high-speed network mechanism applies different neuron weights to each training sample to achieve soft parameters Sharing,formulating different answer layers for different task data sets.Through experiments on multiple public data sets,the experimental results show that the model proposed in this paper performs better than the traditional model in the field of machine reading comprehension tasks.Second,in the task of machine reading comprehension in a single field,a method of combining multi-task learning based on pre-trained language models and hard constraints is proposed.The model uses a pre-trained.and designed network structure to do language model tasks,uses a large amount of unlabeled text,and then extracts a large amount of linguistic knowledge into the network structure,using a hard constraint method to make multiple tasks share these Network parameters,and different answer layers for data sets for different tasks.By conducting experiments on multiple data sets,the experimental results show that the model proposed in this paper has better performance in a single field.