Enhancing emotion detection in software engineering using a residual multi-embedding fusion network

Abstract

Emotions play a crucial role in the development of software, particularly in team dynamics, productivity, and decision making. Developer communications — such as bug reports, code reviews, and online discussions — often include emotional signals. But part of the difficulty in identifying these feelings lies in the technicality and informality of the words, and in the utter scarcity of even critical but rare emotions like fear and surprise. This study aims to improve the detection of both common and minority emotions in software engineering texts, with a focus on better identifying underrepresented classes. We introduce R-MEFN, Residual Multi-Embedding Fusion Network, a network model that employs multiple types of contextual word embeddings to represent the text. Residual connections serve to keep signals of subtle emotionality, especially ones associated with emotions that are infrequent. Cross-validation is performed to choose the best combination of embeddings to be fused. We evaluate R-MEFN on two real-world datasets (StackOverflow and Jira), comparing it to other prior approaches on these benchmarks, as well as to single-embedding and combined-embedding baselines. R-MEFN outperforms other methods that have been evaluated in the same benchmarks for multilabel emotion detection, showing particular improvements on rare classes while keeping a good performance on frequent emotions. Also, it outperforms all single-embedding baselines, as well as all combined-embedding baselines, where embeddings from multiple sources are simply concatenated, showing the strength of the fusion approach. The cross-validated integration of contextual embeddings allows R-MEFN to produce more balanced and expressive representations across all emotion categories. These findings show the effectiveness of using multiple contextual embeddings and residual learning for addressing class imbalance in emotion detection. We see R-MEFN as a useful starting point towards creating emotion aware tools that can allow software teams to track emotional dynamics in a project, and identify hidden risks.