An Innovative 3D Attention Mechanism for Multi-Label Emotion Classification

Due to considerations regarding future work and research progress, the release of data related to this project will be postponed.

Taxonomies

The current implementation focuses on the Original GoEmotions dataset, with future plans to expand to:

1. Original GoEmotions: 27 emotions + neutral
2. Hierarchical Grouping: Positive, negative, ambiguous + neutral
3. Ekman Emotions: Anger, disgust, fear, joy, sadness, surprise + neutral

Requirements

• torch==1.4.0
• transformers==2.11.0
• attrdict==2.0.1
• pandas
• matplotlib
• seaborn
• scikit-learn

Hyperparameters

Model Architecture

• Base Model: XLNet (xlnet-base-cased)
• Hidden Size: 768
• Number of Emotions: 28
• Attention Mechanism: Multi-head with emotion-specific attention layers

Data Preprocessing

• Tokenization: XLNet tokenizer
• Input Sequence Length: 128 tokens
• Train/Test Split: 80/20

Upcoming Features