Effects of Common Sense and Supporting Texts for the Important Words in Solving Text Entailment Tasks - A Study on the e-SNLI Dataset

The recent progress of Machine Learning and Deep Learning opens multiple application areas. Successfully adopting the ML model in production depends on its performance and explainability. This research describes our proposed methodology for exploring Explainable Natural Language Processing (NLP) models. Our work is motivated by the e-SNLI dataset published by Camburu et al. [1]. This dataset contains a large number of text samples, each sample comprising the premise, hypothesis, label, explanation, and a set of highlighted words. In this research, we proposed an approach to expand such a dataset and similar ones by incorporating common sense and supporting texts built from the important keywords in the premise and hypothesis. Later on, the performance evaluation is conducted to verify the effect of the supporting data on the text entailment tasks, especially on explanation text generation. Our preliminary results confirm the effectiveness of our approach to improving the entailment classification, while there is more room to improve the explanation text generation.