Study On Answer Ranking For Machine Reading Comprehension

With the development of intelligent automobile technology and the support of the Ministry of Education on students' practice activities,the scale of various kinds of smart car competition for students has been gradually expanded.In recent years,with the development of deep learning,many traditional problems in the field of natural language processing,such as word segmentation and named entity recognition,have made breakthroughs in ideas and solutions.At the same time,as new technologies emerge,new natural language processing tasks are constantly being produced.The problem of machine reading comprehension derived from the question answering task has gradually become a popular topic in recent years.As metioned above,the machine reading comprehension problem is a branch of the question answering task.This problem also requires the machine to answer the question.It needs to rely on the semantic comprehension and reasoning to help the model answer questions or sort the answers according to the given context environment.In the problem of answer ranking for machine reading comprehension,if we can propose some optimization or improvement ideas based on the existing methods,it has a heuristic meaning for the machine reading comprehension problem itself.In response to this problem,the content of this research mainly includes the following two aspects.First,the study focuses on apriori features and different fusion methods on the problem of answer ranking for machine reading comprehension.In order to help the model understand the context more fully,we can provide apriori information at the model input,such as co-occurrence features of words in questions and answers,named entity features,and character features,as well as researching different fusion methods for questions,articles,and candidate answers.These methods are able to better sort and choose answers.The validity of the apriori features and our fusion method proposed in this paper has been checked based on the RACE dataset on this research,and the validity of the proposed fusion method is further proved on the SQuAD dataset.Second,the research focuses on the influence of sentence reasoning on the answer ranking problem for machine reading comprehension.On this research we propose an option gating network based on sentence information combined with the current common sentence inference problem.The network can help the model achieve better results in sentence inference,thus helping the model to improve the answer ranking for questions.The relevant experiments in this section have achieved the best results on the sub-dataset of RACE,the RACE-H dataset,within this research.