Live demo: https://umer-78.github.io/sentiment-analyzer/
A Naive Bayes text classifier written from scratch — the counting, the add-alpha smoothing, the log-space arithmetic — plus a tokeniser built for reviews, an explanation of every prediction, and a benchmark against scikit-learn on the same split.
$ sentiment predict --explain "The screen is lovely but the battery is hopeless"
positive 76.8% The screen is lovely but the battery is hopeless
hopeless -1.13 → negative
lovely +1.07 → positive
is +0.48 → positive
but +0.21 → positive
screen +0.17 → positive
the -0.07 → negativeThe interesting part of sentiment analysis is not the classifier. It is the text:
| Problem | Handled by |
|---|---|
not good must not look like good |
negation tagging until the clause ends (not_good) |
loooove, LOVE |
repeated letters squeezed, capitals kept as a _caps marker |
:) 👍 |
emoticons and emoji survive tokenisation |
! carries sentiment, , does not |
punctuation filtered, not blanket-stripped |
| unseen words at prediction time | ignored, rather than allowed to zero a class |
| long documents | log-space sums, so no underflow at 5,000 words |
2,000 reviews, 75/25 split:
1,500 training reviews, 500 test reviews
model accuracy precision recall f1 seconds
------------------------------------------------------------------------------
from scratch (unigrams) 0.968 0.969 0.969 0.969 0.08
from scratch (+ bigrams) 0.972 0.970 0.977 0.973 0.11
from scratch (no negation handling) 0.964 0.962 0.969 0.966 0.09
scikit-learn MultinomialNB 0.968 0.969 0.969 0.969 0.10
scikit-learn TF-IDF + logistic 0.972 0.970 0.977 0.973 0.11
The from-scratch model matches scikit-learn's MultinomialNB to three decimals,
which is the point: it is the same algorithm, written out. TF-IDF with logistic
regression is slightly ahead, as expected.
Most informative words it learned:
most positive most negative
recommended 4.01 zero -3.36
highly 3.99 refund -3.40
worth 3.40 asked -3.40
rupee 3.40 experience -3.53
so 3.37 not_again -3.64
anyone 3.35 not_recommend -3.84
pleased 3.26 money -3.90
without 3.26 not_buy -4.27
git clone https://github.com/umer-78/sentiment-analyzer.git
cd sentiment-analyzer
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
sentiment generate # build data/reviews.csv
sentiment train # train and save the model
sentiment benchmark # compare with scikit-learn
sentiment predict "not bad at all, no complaints"
sentiment words # strongest positive and negative words
echo "the app crashes constantly" | sentiment predict --explainAs a library:
from sentiment import NaiveBayes
model = NaiveBayes(alpha=1.0).fit(texts, labels)
model.predict_proba("works exactly as described :)") # {'positive': 0.99, 'negative': 0.01}
model.explain("slow and overpriced") # [('slow', -1.8), ('overpriced', -1.6)]sentiment generate builds the corpus from templates: positive and negative
openers, bodies and closers, mixed reviews, hard cases ("not bad at all", "looks
great, works badly"), filler sentences, typos, shouting, emoji — and 3% of
labels flipped, because no hand-labelled corpus is perfect.
This is synthetic data. It shows the pipeline works; it does not prove the model
would hit 97% on real reviews, and it will not know words the templates never
used. To use it for real, point sentiment train at a CSV with text,label
columns from your own data.
Contrast is not weighted. In "the screen is lovely, but the battery is hopeless" a person reads the clause after "but" as the verdict; the model just adds up evidence and leans positive. Fixing it properly needs clause-level parsing or a model with word order, which is where this simple approach stops and transformers begin.
ruff check .
python -m pytest -q # 18 testsThey check the tokeniser's negation scope and emphasis handling, that the smoothing matches the textbook formula, that likelihoods sum to 1, that a 5,000-word document does not underflow, that accuracy clears the majority baseline by a wide margin, and the whole CLI.