Performance IMprovement Suggestions

[Varunkumar2516]

Title: Improving Sentiment Analysis Accuracy (Currently 90%) – Suggestions Needed

Hello

I built a Movie Sentiment Analysis model using TF-IDF + Logistic Regression and achieved 90% accuracy on the IMDB dataset.

Current pipeline:

• Text cleaning (HTML removal, contractions, punctuation removal)
• Stopword removal (keeping negations)
• Lemmatization with POS tagging
• TF-IDF (max_features=45000, ngram_range=(1,2))
• Models tried: Naive Bayes, KNN, Logistic Regression, SVM, Decision Tree

Goal: Improve accuracy to ~92 to 95%+

What I’ve tried:

• Ensemble methods (did not improve significantly)
• Hyperparameter tuning

Questions:

1. Are there better feature engineering techniques I should try?
2. Would word embeddings (Word2Vec, GloVe) help here?
3. Any suggestions for handling tricky cases like negations better?

Here is my notebook:
https://github.com/Varunkumar2516/IMDb-Sentiment-Analysis-NLP-Project/blob/master/1%20IMDB_Sentiment_Analyzer_Notebook%20.ipynb

Any suggestions or feedback would be really helpful. Thanks!