metadata
language:
- hi
- en
tags:
- emotion-recognition
- code-mixing
- hinglish
- conversation
- semeval-2024
datasets:
- masac
metrics:
- f1
HingRoBERTa ERC — E-MASAC
Fine-tuned l3cube-pune/hing-roberta for Emotion Recognition in Conversation (ERC) on the SemEval-2024 Task 10 E-MASAC benchmark (Hinglish TV dialogues from Sarabhai vs. Sarabhai).
Performance
- Test wF1: 0.4852 (HingRoBERTa + SVM soft ensemble)
- Phase 1 alone: wF1 = 0.3901
- +12.5 pts above SemEval LSTM baseline
Architecture
Two-phase ensemble:
- Phase 1: Fine-tuned HingRoBERTa with Focal Loss + R-Drop, speaker embeddings, emotion-conditioned context
- Phase 2: SVM-RBF on 3,185-d features (Phase 1 embeddings + hand-crafted conversational features)
Labels
anger, contempt, disgust, fear, joy, neutral, sadness, surprise
Citation
Chakradhar Kotha, Konda Bhanu Sai Akhil, Venkata Siva Sai Krishna Pokuri IIIT Hyderabad | SemEval-2024 Task 10