IMDb Sentiment Analysis

Compare LSTM, GRU, and LSTM–GRU models trained in TensorFlow on IMDb movie reviews.

Try a review

Predictions run in your browser. Your review is not sent to a prediction service.

The models use the first 160 tokens. Scores are uncalibrated estimates for positive and negative sentiment; there is no neutral class.

Model comparison

Evaluated on 9,916 test reviews using their true labels. Duplicate normalized reviews were removed before splitting. Vocabulary fitting uses training reviews only; the default model is selected by validation ROC AUC.

Model	Test accuracy	Test macro F1	Test ROC AUC

Test ROC curves for the three trained sentiment models

Results use one split and one training seed. New review sources, sarcasm, and mixed sentiment may behave differently.