Sentiment Analysis — Approaches & Techniques

Sentiment analysis (opinion mining) determines the emotional tone/polarity of a piece of text — typically positive, negative, or neutral — and is one of the most widely deployed real-world NLP applications (product reviews, social media monitoring, brand tracking).

Levels of Sentiment Analysis

Approach 1 — Lexicon-Based (Rule-Based)

Uses a pre-built dictionary of words tagged with sentiment scores (e.g. "excellent" = +3, "terrible" = -3), then sums/averages scores across the text.

"The battery life is excellent but the camera is terrible"
   excellent (+3)                        terrible (-3)
Overall score: +3 + (-3) = 0  -> roughly neutral/mixed

Handling negation is critical for lexicon methods (recall Unit 1's stop-word caveat about "not"):

"not good"  -> naive lexicon sees "good" (+2) and ignores "not" -> WRONG (+2)
"not good"  -> negation-aware: flip sign of following sentiment word -> CORRECT (-2)

Approach 2 — Machine Learning-Based (Supervised)

Treats sentiment analysis as text classification (exactly the Naïve Bayes pipeline from the previous lesson) — train on labeled examples (review text → positive/negative label), extract TF-IDF features, and classify.

Approach 3 — Deep Learning / Transformer-Based (Preview of Unit 4)

Modern systems use contextual embeddings (BERT-style models) fine-tuned on sentiment-labeled data, which handle negation, sarcasm, and context far better than lexicon or bag-of-words approaches, since the representation of each word already depends on its surrounding sentence.

Comparing the Three Approaches

Challenges Specific to Sentiment Analysis

1. Negation: "not bad" is actually mildly positive, not negative.
2. Sarcasm/irony: "Oh fantastic, my flight got delayed again" — positive words, negative sentiment.
3. Comparative sentences: "Phone A is better than Phone B" — sentiment is relative, not absolute.
4. Domain dependence: "unpredictable" is negative for a car ("unpredictable brakes") but positive for a movie plot ("unpredictable twist").
5. Mixed/aspect-level sentiment: a review can be positive about one aspect and negative about another simultaneously.

We build a hands-on sentiment classifier — both lexicon-based and ML-based — in the next lesson.