Emotion Detection in Text: RNNs vs. Transformers

Overview

This project compares traditional Recurrent Neural Networks (RNNs) with modern Transformer-based models for emotion detection in text, with a focus on social media content. The research explores how these architectures handle the unique challenges of informal language, including sarcasm, mixed sentiments, and linguistic noise.

Key Features

• Comparative Analysis: Direct performance comparison between RNN (LSTM/GRU) and Transformer (BERT) architectures
• Social Media Focus: Specialized evaluation on noisy, user-generated text with informal language patterns
• Comprehensive Metrics: Evaluation across accuracy, F1-score, computational efficiency, and memory requirements
• Practical Insights: Identification of optimal use cases for each architecture based on deployment constraints

Technical Approach

Data Processing

• Specialized text cleaning for social media content (handling emojis, slang, typos)
• Emotion label normalization across datasets

Model Architectures

• RNN Baseline: Bidirectional LSTM with attention mechanism
• Transformer Model: Fine-tuned BERT-base with emotion classification head

Evaluation Framework

• Standard metrics (precision, recall, F1) across emotion categories
• Computational efficiency benchmarks (training time, inference speed)
• Error analysis on challenging cases (sarcasm, ambiguous expressions)