Three-Class Sentiment Analysis in Movie Reviews

The Problem

Standard sentiment analysis treats reviews as binary (positive/negative), but real opinions are more nuanced. A 3-star review isn't negative; it's neutral. Reframing sentiment as a three-class problem (bad/neutral/good) better reflects how people actually express opinions, but it's significantly harder for models to learn.

What I Built

A comprehensive evaluation of five deep learning architectures on 50,000 IMDb reviews reclassified into three sentiment categories. Includes systematic ablation studies isolating the contribution of each architectural component.

Architectures Evaluated

Model	Macro-F1	Key Characteristic
LSTM	0.721	Sequential processing, captures long-range dependencies
BiLSTM	0.738	Bidirectional context, better phrase understanding
CNN	0.712	N-gram feature extraction, fast inference
GRU	0.729	Simpler gating, fewer parameters than LSTM
BERT	0.784	Pre-trained contextual embeddings, attention mechanism

Ablation Studies

Rather than just reporting final scores, I isolated individual components to understand what actually drives performance:

• Embedding strategy: Pre-trained GloVe vs randomly initialized vs BERT subword tokenization
• Architecture components: Effect of attention layers, bidirectionality, and depth
• Hyperparameters: Learning rate schedules, dropout rates, batch size sensitivity
• Class imbalance: Impact of the neutral class being underrepresented in training data

Key Findings

• BERT's pre-trained contextual embeddings provide the largest single performance gain over all other approaches
• The neutral class is hardest to classify across all architectures (often confused with weak positive/negative)
• BiLSTM + attention narrows the gap with BERT significantly while using far fewer parameters
• CNN performs surprisingly well given its simplicity, especially on short reviews

Links

GitHub Repository | Research Paper (PDF)