Clothing Review Sentiment Classifier

Fine-tuned DistilBERT for 3-class sentiment classification (negative/neutral/positive) on ~48k clothing reviews from Kaggle ("Consumer Review of Clothing", Jocelyn Dumlao).

Results (test set, n=9,659)

Error analysis

Most errors occur at class boundaries, particularly 4-star reviews predicted as neutral. 72% of these errors are 4-star reviews, and 76% contain contrast words ("but", "however") vs 36% of correctly classified positives — suggesting the model is reasonably picking up on genuinely mixed sentiment (praise + a caveat) rather than making random mistakes. Some errors likely trace to label noise (e.g. reviews describing returns or damage rated 4-5 stars).

Labels

0 = negative, 1 = neutral, 2 = positive

Training

• Base model: distilbert-base-uncased
• 3 epochs, best checkpoint at epoch 2 (by validation macro F1)
• Class-weighted loss to handle imbalance (74% positive / 15% negative / 11% neutral)
• max_length=128, learning_rate=2e-5, batch_size=32