WhatsApp Chat Analysis

This started as a course project and I kept adding things. You upload an exported chat (Android and iPhone exports use different formats, and the parser handles both) and it shows activity stats, timelines, heatmaps, a word cloud with Hinglish stopwords removed, and emoji usage. Then the ML parts: sentiment, LDA topics, toxicity with Toxic-BERT, and a reply graph that shows who talks to whom.

Indian group chats are mostly Hinglish, and most sentiment models are trained on English tweets. That mismatch was the interesting part.

I trained Logistic Regression and LinearSVC on Sentiment140 (TF-IDF), and also added multilingual BERT. Messages that aren't in English get translated with deep-translator. On top of that there's a small hand-made Hinglish slang list that overrides the model for strong words it would otherwise get wrong.

• Sentiment: LogReg / LinearSVC (TF-IDF, Sentiment140) plus nlptown/bert-base-multilingual-uncased-sentiment, with a Hinglish slang override.
• Topics with gensim LDA, toxicity with unitary/toxic-bert.
• Reply network with NetworkX, drawn interactively with PyVis.
• Hinglish stopword list for the word cloud, and deep-translator for non-English messages.
• pytest tests and a Dockerfile.

Test accuracy on Sentiment140: Logistic Regression 76.78% (trained on a 1M-tweet sample), LinearSVC 73.79% (trained on 50k; it overfits at 80.95% on train). On real Hinglish chats it's worse than that, because chat messages look nothing like tweets. I wrote that limitation into the README instead of hiding it.