Research On Temporal Question And Answering Methods For Machine Reading Comprehension

In recent years,machine reading comprehension(MRC)is an important research topic in the field of natural language processing.The data set for MRC task contains a variety of questions,of which the temporal questions are common and difficult.Temporal questions require machines to understand temporal semantic information based on the common sense knowledge of time in the real world,and analyze and reason with questions and materials to accurately answer relevant questions.However,due to implicitness and diversity of temporal expressions,the existing reading comprehension models can not fully understand the semantic information of temporal questions and temporal expressions,which affects the effect of answering temporal questions.Therefore,this paper mainly focuses on the method of answering temporal questions of extractive reading comprehension,and the main work is as follows:(1)The temporal expression recognition method combining BART and template is studied.Currently,temporal expression recognition is mainly based on sequential neural networks,a large amount of data needs to be labeled manually,and the attention to fine-grained and implicit temporal information is insufficient.In this paper,eight types of temporal expressions are identified by combining BART and template.This method uses the template to acquire more knowledge to alleviate the need for labeling data and uses the generative pre-trained language model BART to improve the accuracy of temporal expression recognition.Relevant experiments show that this method can effectively identify the temporal expression,and the F1 value for ours increased by 1.22% compared with the baseline models.(2)The question answering method based on question understanding and temporal semantic enhancement is studied.In this paper,temporal questions are more accurately understood by the model through multiple strategies such as problem classification,temporal-constrained information extraction,interrogative word recognition and external knowledge fusion,and the use of differential attention mechanisms based on subtraction to enhance the representation of temporal semantic information.Relevant experiments show that this method can effectively learn the temporal characteristics of the questions,and the F1 value for ours increased by 2.72% compared with the baseline model.(3)A prototype system of automatic answers to reading comprehension temporal questions is designed and implemented.Based on the above research methods,the temporal question-answering system for Chinese machine reading comprehension is designed and implemented.The system mainly includes data preprocessing,question classification,temporal expression recognition,automatic question answering and other modules.Through specific examples,the system is tested and the relevant pages of this system are displayed.To sum up,research has been conducted on the understanding strategies of temporal semantic information,and the method of answering temporal questions in MRC has achieved certain results in answering real questions.In the future,it is necessary to further strengthen the reasoning ability of models for temporal relationships between multiple events.