Emotion Detection and Intensity Prediction

Project Overview

This repository is for our project as part of SemEval 2025 Task 11-A, focusing on emotion detection and intensity prediction from text. The goal is to classify emotions expressed in a text snippet and predict their intensity. The target emotions include:

• Joy
• Sadness
• Fear
• Anger
• Surprise
• Disgust

This project involves multi-label classification and ordinal intensity prediction for a single language (English).

---

Goals

1. Develop a machine learning model capable of:
   • Detecting multiple emotions in a text.
   • Predicting the intensity of each detected emotion.
2. Focus on a single-language (English) dataset to optimize accuracy.
3. Provide explainability and insights into the model's predictions using tools like SHAP.
4. Evaluate the model's performance using the F-score and explore comparisons with human judgments.

---

Tasks

1. Preprocess and explore the dataset to prepare it for model training.
2. Fine-tune transformer models (e.g., RoBERTa) for:
   • Emotion detection (multi-label classification).
   • Intensity prediction (ordinal regression or classification).
3. Conduct model evaluation and error analysis.
4. Perform explainability analysis to interpret the model's predictions.
5. Document findings, results, and challenges.

---

Dataset Specifications

• Source: Dataset provided by the SemEval 2025 Task 11-A competition (download from Codabench).
• Language: English.
• Structure: Text snippets with labels for emotions and optional intensity levels.
• Emotion Classes: Joy, sadness, fear, anger, surprise, disgust.
• Intensity Levels:
  • 0: No emotion.
  • 1: Low intensity.
  • 2: Moderate intensity.
  • 3: High intensity.

Outline of Plan

Phase 1: Data Preparation

• Download and preprocess the dataset.
• Explore label distributions and handle class imbalance.
• Tokenize text using RoBERTa's tokenizer.

Phase 2: Model Development

• Fine-tune RoBERTa for emotion detection and intensity prediction.
• Add dual-output heads:
  • Multi-label classification for emotions.
  • Ordinal regression or classification for intensity.

Phase 3: Evaluation

• Evaluate the model using:
  • F-score for emotion detection.
  • Mean Squared Error or Weighted F-score for intensity.
• Compare results with baseline models.

Phase 4: Explainability

• Use SHAP to analyze model predictions and identify key text features contributing to emotions and intensity.

Phase 5: Deployment and Reporting

• Develop APIs for emotion and intensity prediction.
• Document the model's performance, insights, and challenges.
• Prepare a final report for submission.

---

Getting Started

Dependencies

• Python 3.8+
• Hugging Face Transformers
• PyTorch or TensorFlow
• SHAP for explainability

To Do

• Set up the repository structure.
• Collaborate on preprocessing scripts and baseline models.
• Divide tasks among team members for efficient progress.

---

Links

• SemEval 2025 Task 11-A Details
• Competition GitHub Repository
• Guidelines about submission and dataset
• BERT Text Classification Tutorial
• RoBERTa Explanation
• SHAP for Text Analysis
• GoogleDoc for Documentation

---