Multi-Head Neural Network for Emotion Classification and Trigger Detection in Multi-Speaker Dialogues
Authors: Lorenzo Balzani, Alessia Deana, Thomas Guizzetti
In this project, we tackle the Emotion Discov- ery and Reasoning its Flip in Conversation (EDiReF) task, which focuses on predicting the emotion of each sentence in a dialogue and identifying any emotional shifts of speak- ers, or "triggers". Our methodology began with creating a BERT-based multi-head neu- ral network, which we then aimed to refine by incorporating a concatenation method, a transformer, and a large language model. The outcomes demonstrated below average capa- bilities in emotion prediction and the perfor- mance in identifying emotional triggers was less satisfactory. Through our efforts to en- hance the model with various improvements, we observed a marginal but significant improve- ment in the large language model’s ability to detect triggers. Still, more fine-tuning is re- quired to reliably recognize emotional changes within dialogues.