Can Knowledge Enhance Reading Comprehension? An Integrated Approach with Semantic Lexicon

Abstract

The machine reading comprehension task (MRC) requires the model to answer die questions based on a piece of context. Over the past few years, more and more powerful models have been proposed based on various deep learning techniques. The MRC models based on deep learning is powerful and effective; however, most of them are focusing on changing the neural network structure. Apart from improving on the deep learning architectures, word embeddings are also essential elements for question answering systems and should not be neglected. Even a small improvement in word representation can lead to substantial performance differences in question answering task. The proposed approach comprises two modules that specialize the semantic representation of word representation and then pipe them to use in MRC models. Fundamentally, pre-trained vectors are retrofitted based on semantic lexicons (PPDB, WordNet, and FrameNet) beforehand. Our experiments on the Stanford Question Answering Dataset (SQuAD) reveal that integrating both a single and combined lexicon knowledge yields improvements over the only use of pre-trained embeddings.