Emotion analysis systems are expected by users to provide reliable results. Optimizing their robustness, in addition to accuracy, is particularly challenging in noisy environments, such as social media data or other forms of naturally expressed language. Our hypothesis is that robustness can be improved by guiding a language model to pay attention to multiple emotion signals in the input text. Therefore, requesting labels according to multiple emotion theories may improve robustness. To study this hypothesis, we run permutation experiments on input instances to test the robustness of LLM configurations based on basic emotions only, and jointly with valence–arousal, and cognitive appraisals. We find that adding valence–arousal and cognitive appraisal along with basic emotion instructions to the emotion classification improves robustness, with the effect being most pronounced for adding appraisals. In a further analysis, we find that this improvement is less susceptible to word order permutations and more to random word substitutions and deletions. These differences hold across multiple LLMs studied in our work.

This paper presents an original approach for SemEval 2025 Task 11. Our study investigates various strategies to improve Text-Based Multi-label Emotion Detection task. Through experimental endeavors, we explore the benefits of contextualized vector representations by comparing multiple BERT models, including those specifically trained for emotion recognition. Additionally, we examine the impact of hyperparameters adjustments on model performance. For Subtask A, our approach achieved F1 scores of 0.71 on the English dataset and 0.84 on the Russian dataset. Our findings underscore that (1) monolingual BERT models demonstrate superior performance for English, whereas multilingual BERT models perform better for Russian; (2) pretrained emotion detection models proving less effective for this specific task compared to models with reduced vocabulary and embeddings focused on specific languages;(3) exclusive use of BERT-based models, without incorporating additional methods or optimization techniques, demonstrates promising results for multilabel emotion detection.