meld-text-baseline

Text Emotion Recognition on MELD (BERT Embeddings + MLP Classifier)

This repository contains a text-only emotion classification model trained on the MELD (Multimodal EmotionLines Dataset) using pre-extracted BERT embeddings and a lightweight MLP classifier.

The model predicts one of 7 emotion classes from a single utterance.

---

Model Overview

• Base language model: bert-base-uncased
• Embedding extraction: CLS token (no pooling)
• Classifier: 2-layer MLP
• Training strategy:
  • BERT encoder is frozen
  • Only the classifier is trained

---

🗂 Dataset

• Name: MELD (declare-lab/MELD)
• Modality: Text (utterances)
• Setting: Multi-class emotion classification
• Splits: Train / Validation / Test (official MELD splits)

---

Training Details

• Loss: Cross-entropy
• Optimizer: Adam
• Max sequence length: 128
• Batching: Embeddings extracted in batches
• Evaluation metrics:
  • Accuracy
  • Macro F1-score
  • Per-class F1-score

---

Important Notes

• This model does not fine-tune BERT.
• It relies on offline embedding extraction.
• The provided weights correspond only to the classifier.

To reproduce results, the same BERT model must be used for embedding extraction.

---

Intended Use

• Baseline for multimodal emotion recognition
• Comparison with audio-only and fusion models
• Research and educational purposes

---

Limitations

• Context between utterances is not modeled
• Speaker identity is not used
• Performance depends on quality of extracted embeddings

---

Reproducibility

To reproduce the reported results, embeddings must be extracted using the same bert-base-uncased model with identical tokenization and maximum sequence length.

---