PetraMicanovic/meld-text-class-weights

Text Emotion Recognition Model

A text-based emotion recognition model trained on the MELD dataset. The model operates on dialogue utterances and serves as both a strong text-only baseline and the text encoder in a multimodal emotion recognition framework.

Model Summary

• Task: Text-based Emotion Recognition
• Dataset: MELD
• Backbone: bert-base-uncased
• Pooling: [CLS] token representation
• Classifier: MLP with class-weighted loss
• Classes: 7 emotion categories

Architecture

1. Text Encoder

   • Pretrained BERT (bert-base-uncased)
   • Outputs contextualized token embeddings

2. Utterance Representation

   • [CLS] token embedding
   • Represents the full utterance semantics

3. MLP Classifier

   • Fully connected layers
   • ReLU activation and dropout
   • Softmax output layer

Class Imbalance Handling

The MELD dataset is highly imbalanced across emotion classes. To address this, class weights are applied in the cross-entropy loss function, improving macro-level performance on minority emotions.

Training Details

• Tokenization: BERT tokenizer
• Max sequence length: 128 tokens
• Optimizer: Adam
• Loss: CrossEntropyLoss (with class weights)
• Metrics: Accuracy, Macro F1, Weighted F1

Usage

• Standalone text emotion classifier
• Text branch for early and late fusion in multimodal emotion recognition