Emotion Recognition for Human-Centered Conversational Agents

This thesis proposes a study on Emotion Recognition in Conversation to address the challenges of the task with a chatbot reference case study to enhance conversational agents’ ability to understand and respond appropriately to human emotion. The study consists of two phases. The first one involves the use of several baselines and the implementation of EmoBERTa to explore aspects of the task, such as preprocessing, balancing technique and context modelling tested on ERC benchmark dataset. The results reveal that the punctuation provides key information to the task, balancing techniques can provide marginal improvements if appropriately selected and context can provide additional information and suggest that a non-static context construction could be beneficial. In the second phase, the effectiveness of a Few-Shot learning method, SetFit, is explored in the context of ERC to face the scarce amount of real labelled data. An incompatibility with the given context definition of the architecture employed by the mentioned method called for an adaptation which proved to be ineffective. The performance of the SetFit method and finetuning are compared in a limited data regime. Finally, the study explores the capabilities of a trained model on a specific ERC dataset to adapt to limited data from a different domain using Transfer Learning and fine-tuning with inconclusive results. The findings and insight from this can lay the groundwork for future developments and studies in the growing field of emotional-aware conversational agents and the application of Few-Shot learning in this task.