Open-Domain Question Answering In Chinese Based On Neural Machine Reading Comprehension

Open-Domain Question Answering(ODQA)systems return the answer,identified from a vast document collection,as a response to a question asked in natural language.Re-cent years have seen considerable success in natural language understanding with the de-velopment of Neural Machine Reading Comprehension(NMRC).As ODQA and NMRC have common with each other,ODQA can be tackled based on NMRC.However,because of the characteristics of Chinese ODQA,there are several chal-lenges.First,in Chinese Natural Language Processing(NLP),word segmentation is often an indispensable step,but segmentation errors are unavoidable and would propagate to downstream tasks.Second,in open domain settings,there may be a large number of doc-uments related to the question.How to find the article that contains the right answer is the key to ODQA.Third,answers to a series of successive questions may be in different documents.The model should answer all of those questions correctly.This article aims to tackle above challenges.More specifically,first,we add deep contextualized word representations to the embedding layer to extract contextual informa-tion more effectively and contribute to the answer prediction.Second,we give a compre-hensive survey on the progresses of NMRC.The general architecture of NMRC models is decomposed into four modules and prominent approaches utilized in each module are introduced detailedly.Based on the survey,we design the RNN Transformer network(R-Trans)to handle the Chinese MRC task,which not only spends less time in training and inference but also improves answer prediction.Third,we extend R-Trans network to ODQA with the methods of taking the results of both MRC and question-document matching into consideration jointly when selecting relevant documents,which contributes to higher recall and mitigate the side-effect of incorrect document selection as well.In addition,we split a series of questions and introduce heuristic co-reference resolution methods so that our model can answer all of those successive questions correctly.