Sentiment Analysis

Overview Sentiment Analysis (Opinion Mining) is the use of NLP to systematically identify, extract, and quantify affective states and subjective information. It determines if a piece of writing is positive, negative, or neutral.

Levels of Analysis

1. Document Level Classifying the whole document.

• "This review is Positive."

2. Sentence Level Classifying each sentence.

• "The screen is great." (Positive)
• "But the battery sucks." (Negative)

3. Aspect-Based Sentiment Analysis (ABSA) linking sentiment to specific attributes (Aspects).

• Entity: iPhone
• Aspect: Battery → Sentiment: Negative
• Aspect: Screen → Sentiment: Positive

Approaches

Rule-Based (VADER) Uses a dictionary of word scores ("Good" = +1.9, "Bad" = -1.5) and rules for amplifiers ("Very good" > "Good") and negations ("Not good" is negative).

• Good for social media (handles emojis/slang).
• Fast/Cheap.

Machine Learning (BERT/RoBERTa) Deep learning models trained to "read" the full context.

• "The movie was not unpredictable." → Positive (Double negative).
• SOTA accuracy.

Use Cases

• Brand Monitoring: Tracking Twitter/X mentions during a PR crisis.
• Stock Trading: Analyzing news headlines to predict market movement.
• Customer Support: Prioritizing angry tickets.
• Product Analysis: Aggregating pros/cons from Amazon reviews.

Challenges

• Sarcasm: "Great, my phone died again." (Machine sees "Great", human sees sarcasm).
• Nuance: "The movie was shorter than I expected." (Good or Bad?).

Sentiment analysis turns unstructured voice-of-customer data into tracked metrics.