Description Question Answering In Reading Comprehension

Machine reading comprehension?MRC?aims to make a machine read texts like human being,and answer a series of related questions through understanding.In recent years,MRC has been focused in the field of artificial intelligence and NLP.Currently,most research on MRC pay more attention on the questions with short answers.However,the description questions with long answers are common in real life,so it is necessary to study this kind of questions.Generally,description questions are with high semantic abstraction,and their corresponding answers are expressed with concrete words.Meanwhile,their answers are often composed of many continuous sentences,so they have discourse natures.Therefore,understanding of questions and discourses are important for answering description questions.The main works of this paper are as follows:?1?The answering strategy based on question understanding is explored.This paper enhances understanding of questions in the model by identifying three types of information: question type,question topic,question focus.Firstly,the question type is classified by CNN and keywords;secondly,the question topic and focus are obtained through syntactic analysis;finally,the information on questions is integrated into the answering model.Experiments on relevant datasets show that: with the integration of question understanding,the performance of the answering model increases by 2%-10%,compared to the baselines.?2?The answering strategy based on discourse representation is studied.This paper get discourse representation by using hierarchical coding.At first,the sentence representation is obtained by using CNN or LSTM from word representations;then,the discourse representation is gained by Bi_LSTM from sentence representations;and the discourse representation is updated by interacting with question representation;finally,prediction layer predicts the start and end positions of the answer based on the sentence sequences.Experiments on relevant datasets show that: applying the hierarchical representation of the discourse,the metrics of Rouge_L and Bleu₄ of the answering model increases by about 4% and 2% respectively.Based on the above model,we make a further study on answering ‘how'questions,which have a large proportion in the dataset.We merge the temporal relationship into the model's prediction layer,as there are temporal relationships between sentences of the answer for such kind of questions.Experiments on ‘how' questions show that: the metrics of Rouge_L and Bleu₄ of the answering model increases by 1.08% and 0.97% respectively.?3?The answering strategy based on discourse and question understanding is implemented.In the answering model,the question understanding information and hierarchical representation of the discourse are integrated.Meanwhile,the document sorting information and the post-answering strategy are used.Experiments on relevant datasets show that: the metrics of Rouge_L and Bleu₄ of the answering model increases by about 9% and 11%respectively.The model based on the discourse and question understanding is effective in answering description questions.However,some further researches such as the understanding of abstract words and the identification of discourse relationship are needed.