This repository contains the implementation and experiments for the Emotion Recognition in Conversations (ERC) and Emotion Flip Reasoning (EFR) tasks as part of the EDiReF shared task competition. The aim is to classify emotions in each utterance of a dialogue and identify trigger utterances that cause an emotion shift within the conversation. The project leverages transformer-based models, specifically BERT and ELECTRA, to achieve these goals.
We implemented and compared three different approaches to process the outputs from BERT and ELECTRA:
- Concatenation (concat): Concatenates the encoding of the dialogue with the target utterance.
- No Pooling (nopool): Uses the raw first token from the last hidden layer (the [CLS] token) as a representation of the whole dialogue.
- Extraction (extraction): Processes the entire dialogue in a single forward pass, extracting the target utterance's tokens from the last hidden state and concatenating them with the [CLS] token embedding.
Additional information regarding the results obtained and the specifications on the techniques used can be found in the file emotion_discovery_report.pdf.