DLRCNeg: Deep Learning based Reading Comprehension by handling Negation

Abstract

Question Answer (QA) System for Reading Comprehension (RC) is a computerized approach to retrieve relevant response to the query posted by the users. The underlined concept in developing such a system is to build a human computer interaction. The interactions will be in natural language and we tend to use negation words as a part of our expressions. During the pre-processing stage in Natural Language Processing (NLP) task these negation words gets removed and hence the semantics gets changed. This remains to be an unsolved problem in QA system. In order to maintain the semantics we have proposed a novel approach Hybrid NLP based Bi-directional Long Short Term Memory (Bi-LSTM) with attention mechanism. It deals with the negation words and maintains the semantics of the sentence. We also focus on answering any factoid query (i.e. ’what’, ’when’, ’where’, ’who’) that is raised by the user. For this purpose, the use of attention mechanism with softmax activation function has obtained superior results that matches the question type and process the context information effectively. The experimental results are performed over the SQuAD dataset for reading comprehension and the Stanford Negation dataset is used to perform the negation in the RC sentence. The accuracy of the system over negation is obtained as 93.9% and over the QA system is 87%.