BERT for Sentiment Classification in Software Engineering

Abstract

Sentiment analysis (SA) has been applied to various fields of software engineering (SE), such as app reviews, stack overflow Q&A website and API comments. General SA tools are trained based on movie or product review data. Research has shown that these SA tools can produce negative results when applied to the field of SE. In order to overcome the above limitations, developers need to customize tools (e.g., SentiStrength-SE, SentiCR, Senti4SD). In recent years, the pre-trained transformer-based models have brought great breakthroughs in the field of natural language processing. Therefore, we intend to fine-tune the pre-trained model BERT for downstream text classification tasks. We compare the performance of SE-specific tools. Meanwhile, we also studied the performance of SE-specific tools in a cross-platform setting. Experimental results show that our approach (BERT-FT) outperforms the existing state-of-the-art models in terms of F1-scores.