Emotional intelligence in dialogue systems: methods, models and optimization

The development of modern dialogue systems necessitates the integration of emotional intelligence (EI) to move from mechanical information exchange to deep and contextually relevant human-computer interaction. Despite the demand for empathetic interfaces, existing methods for textual emotion analysis face serious limitations in accuracy, robustness to noise, and computational efficiency. Therefore, the aim of this study was to compare the fine-tuning strategies of the BERT family of models, such as head-only tuning, full fine-tuning, and the use of low-rank LoRA adapters, on datasets ranging from 40,000 to 240,000 messages.

During the experiments, it was found that the choice of the optimal strategy directly depends on the scale of the training sample: if for small data volumes (around 40,000), the advantage remains with full fine-tune, then when scaling the dataset, LoRA adapters lead, demonstrating the best balance of accuracy (F1 ≈ 0.882) and inference speed. To improve classification quality and minimize errors between semantically similar categories, class weighting methods and a focal loss function were successfully implemented. The statistical significance of the obtained results was confirmed using the McNemar test at p < 0.01.

The practical implementation of the research was the proposed architecture of a Telegram bot, integrating an emotion recognition module with a language model to generate empathetic responses in real time.

This work provides developers with clear guidelines for choosing optimization methods based on available resources, ensuring the creation of high-performance and secure commercial chat services capable of taking user communication with artificial intelligence to a whole new level.