Generalization of Machine Reading Comprehension Using Dynamic-Critical Based Learning

Abstract

The purpose of this paper is to research the possibility of generalization of Dynamic-Critical, a reinforcement learning (RL) algorithm, to enhance the performance in machine reading comprehension (MRC) using Standford Question Answering Dataset (SQuAD). MRC is the ability for a machine to read and understand text, then infer appropriate answers to given questions. Various machine learning algorithms such as ensemble have been applied in the field. We enhanced natural language processing (NLP) and MRC performances in traditional used algorithms, R-Net, FusionNet, Conductor Net, and Mnemonic Reader, with the introduction of Dynamic-Critical. These models output the location of text as the answer to a question. When the location is different, their performance degrades. We applied Dynamic- Critical Reinforcement Learning to resolve this issue. According to experiment results, the four algorithms with Dynamic-Critical show better performance on average compared to existing SQuAD results. Hence this paper supports the possibility of generalizing the research result that RL algorithm contributes to MRC performance improvement.