Sentiment Analysis - Opinion Mining

What is Sentiment Analysis?

Sentiment analysis (opinion mining) is the computational task of identifying and extracting subjective information from text—determining whether a piece of writing expresses a positive, negative, or neutral attitude toward a subject. It goes beyond simple keyword matching to understand the emotional tone, opinion strength, and target of sentiments expressed in natural language.

Sentiment analysis is commercially vital: companies monitor brand perception across millions of social media posts, product reviews drive purchasing decisions, and financial analysts gauge market sentiment from news. Understanding automated sentiment extraction is essential for modern NLP practitioners.

Levels of Sentiment Analysis

Document-Level

Sentence-Level

Aspect-Based Sentiment Analysis (ABSA)

Approaches to Sentiment Analysis

Lexicon-Based (Rule-Based)

Machine Learning Approach

Deep Learning Approach

Worked Example: Sentiment Classification

Using VADER (Valence Aware Dictionary)

Challenges in Sentiment Analysis

Negation Handling

Sarcasm and Irony

Domain Adaptation

Comparative Sentiment

Evaluation Metrics

Applications

Interview Questions

Q: How do you handle sarcasm in sentiment analysis? A: Sarcasm detection is an open problem. Approaches include: (1) Context modeling—if surrounding text contradicts the statement, likely sarcastic, (2) Incongruity detection—positive words in negative situations, (3) Author history—some users are frequently sarcastic, (4) Multimodal cues—in speech, tone and facial expressions help. Current best: fine-tuned transformers with sarcasm-labeled data achieve ~75% detection accuracy.

Q: Why is aspect-based sentiment analysis harder than document-level? A: It requires identifying relevant aspects, determining which sentiment words apply to which aspects, and handling mixed sentiments. "Great food, terrible service" has no single document sentiment. The model must parse relationships between opinion words and their targets, sometimes across sentence boundaries.

Q: How does VADER differ from machine learning approaches? A: VADER is rule-based with a human-validated sentiment lexicon. It handles social media well (emoticons, slang, punctuation emphasis) without training data. ML approaches learn from data and can capture domain-specific patterns but need labeled examples. VADER is a good baseline; fine-tuned BERT is state-of-the-art.