#Sentiment Lab — Twitter Sentiment Classification Baseline

A TF-IDF + Multinomial Naive Bayes baseline for classifying tweets about gaming brands and platforms (Borderlands, Call of Duty, Nvidia, FIFA, and dozens of others) as Positive or Negative sentiment.

Takes the Twitter Entity Sentiment dataset — tweets labeled by brand/entity and sentiment class — and trains a classic ML baseline: clean the text, vectorize with TF-IDF, classify with Naive Bayes. This page reports what was actually built and measured, not an aspirational production pipeline.

• Data: A random 500-row sample from twitter_training.csv (the full dataset is much larger), reduced to 302 rows after dropping a missing review and restricting to the Positive and Negative classes only — Neutral and Irrelevant were excluded, so this is a binary classifier, not a full 4-class sentiment model
• Preprocessing: NLTK tokenization (RegexpTokenizer) with stopword removal
• Features: TF-IDF vectorization (302 samples × 1,661 features)
• Model: MultinomialNB, trained on an 80/20 train/test split (241 train / 61 test)

On the 61-sample held-out test set:

Overall accuracy: 62%. That's a modest result for a binary classification task — meaningfully above the ~54% majority-class baseline for this split, but not a strong classifier. It's reported as what it is: a first baseline on a small sample, not a finished system.

• The validation set was never evaluated. twitter_validation.csv is loaded and filtered to match the binary setup, but the trained model is never actually run against it — there's no validation accuracy to report, despite validation data being present in the repo.
• Trained on ~300 rows, not the full dataset. The 500-row sample (down to 302 after class filtering) is a small fraction of the available training data. A production baseline would train on the full set.
• Neutral and Irrelevant classes are dropped, not handled. Real-world sentiment classification needs to deal with ambiguous and off-topic text, not just discard it.

twitter_training.csv (full dataset)
  → Random 500-row sample → drop 1 missing review → 499 rows
  → Filter to Positive/Negative only → 302 rows
  → NLTK tokenization + stopword removal
  → TF-IDF vectorization (302 × 1,661 features)
  → 80/20 train/test split (241 / 61)
  → MultinomialNB → 62% test accuracy
  (twitter_validation.csv loaded and filtered, but not yet evaluated against the model)

Key engineering decisions:

• TF-IDF + Naive Bayes as a deliberate first baseline — a fast, interpretable starting point before reaching for anything heavier. The value of a baseline is knowing what "better" needs to beat, not being the final answer.
• Binary scoping was a real simplification, and it's named as one — dropping Neutral/Irrelevant made the problem tractable for a first pass, at the direct cost of not handling two of the four real classes in the data.

A 62%-accuracy binary baseline on 302 rows isn't an impressive number on its own — but reporting it accurately, alongside exactly what wasn't finished (validation evaluation, full-dataset training, 4-class handling), is more useful than a polished-sounding page that collapses under the first follow-up question about test coverage or validation accuracy. The next real step here is finishing the validation evaluation that's already half-built in the repo.