All projects
Machine Learning/Natural language processing /Project 07 of 10 M.S. coursework

Sentiment Analysis with BiLSTM

A Bidirectional LSTM classifier that labels movie-review text as positive or negative, from tokenization and sequence preparation through held-out evaluation.

Python NLP Deep learning BiLSTM Text analytics M.S. coursework

Business impact

A team considering feedback triage can use the test counts to judge how much reading the classifier would create.

78.57%Accuracy on 98 test reviews
0.7789Positive-class F1
997Records after deduplication

Business question

Can a compact recurrent neural network classify sentiment in short movie reviews?

Approach

How the work was done

  • Parsed 1,000 individual IMDb excerpts and removed three exact duplicate pairs. Grouped splits keep identical cleaned texts together: 799 training, 100 validation and 98 test reviews.
  • Fitted the vocabulary on training text only and padded sequences to 50 tokens, covering all cleaned training reviews. The saved tokenizer and preprocessing metadata accompany the model.
  • Trained a 100-dimension embedding and 128-unit bidirectional LSTM with a 64-unit dense layer, dropout, sigmoid output, and early stopping.

Results

What the analysis shows

  • On the held-out test set, 77 of 98 predictions were correct: accuracy 78.57%, positive-class precision 0.7255, recall 0.8409 and F1 0.7789 at a 0.5 threshold.
  • The negative-review queue contained 47 reviews: 40 negative and 7 positive. It caught 40 of 54 negative reviews and missed 14.
  • The trained Keras model is included in the repository so it can be inspected or loaded without retraining.

From the project

Held-out errors and validation during training

Confusion matrix for 98 reviews: 40 negative and 37 positive classified correctly, with 14 negative and 7 positive reviews misclassified.

The manual-review workload

Negative predictions flag 47 of 98 reviews. The queue contains 40 actual negative reviews and 7 positive reviews; 14 negative reviews are missed.

View paired predictions Open full size
Training and validation curves across five epochs; validation loss selects the restored epoch 3 weights.

Training fit and unseen reviews

Training continued for five epochs and restored the epoch 3 weights with the lowest validation loss. The later divergence shows why fitting the training reviews alone is not enough.

View the source Open full size

Business relevance

What this makes easier to decide

The model offers a repeatable way to summarize sentiment from text, with a clear precision-recall tradeoff to consider before using the labels for customer-experience monitoring.