Improving Generation of Sentiment Commonsense by Bias Mitigation

Commonsense knowledge graphs (CSKG) are crucial for artificial intelligence systems to understand natural language. Recently, with the construction of COMET (Commonsense Transformer) and ATOMIC2020, a comprehensive coverage commonsense reasoning knowledge graph, CSKG research is increasingly vital in natural language understanding and reasoning. Since sentiment commonsense knowledge is understudied yet, our work focuses on improving the generation of sentiment commonsense in ATOMIC2020. We first show a problem in natural language generation that degrades the accuracy due to the unbalanced sentiment distribution in the dataset. Next, we propose the EDA (Easy Data Augmentation) and UDA(Unsupervised Data Augmentation) based methods that improve the accuracy through biased mitigation of the unbalanced dataset. Our experimental results show that EDA method has little effect on the accuracy, while UDA-based method has some accuracy improvements in ROUGE-I, ROUGE-2, and ROUGE-L.

I. Introduction

Commonsense knowledge graphs (CSKG) are essential for Natural Language Understanding(NLU) and reasoning tasks. Recently, COMET(COMmonsEnse Transformers) has been established(Bosselut et a1.,2019), and ATOMIC 2020, a new CSKG that includes general-purpose commonsense knowledge containing knowledge, also appeared(Hwang et al., 2021). However, since sentiment in commonsense knowledge has been rather under-studied, more research is needed(Li et a1.,2021) to improve NLP and reasoning. In this study, we propose improving the generation of sentiment text with sentiment relations datasets related to the emotional state of ATOMIC2020.