Sentiment Analysis

What is sentiment analysis?

Sentiment analysis (also called opinion mining) is the natural-language-processing technique for automatically scoring text as positive, neutral, or negative — and in modern systems, with finer-grained emotional labels (joy, anger, sadness, disgust, fear, surprise). Applied to social media, sentiment analysis processes thousands or millions of posts per day to produce aggregate brand-health scores, competitor comparisons, and trend dashboards. The output is typically a numeric score (e.g., -1 to +1) per post, then aggregated by mention, hashtag, brand, or campaign.

Sentiment analysis went through three technical generations. (1) Lexicon-based (2008-2015) — count positive vs negative words from a hand-built dictionary. Worked OK for English, badly for sarcasm and irony, badly for non-English. (2) Classic ML (2015-2020) — supervised classifiers (Naive Bayes, SVM) trained on labeled data. Improvement over lexicons but still struggled with context. (3) Transformer-based (2020-present) — fine-tuned language models (BERT, RoBERTa, GPT) score sentiment with much richer context understanding. State-of-the-art for English; rapidly improving for other languages.

How brands use sentiment analysis

Three concrete use cases. (1) Brand health monitoring — aggregate sentiment over time produces a 'brand sentiment score' that PR and marketing teams track quarterly. A drift from +0.45 to +0.20 over six months signals brand erosion before it becomes obvious in revenue or NPS. (2) Crisis detection — real-time sentiment dashboards alert on sudden spikes of negative mentions. The classic case: 12pm noon, brand sentiment was +0.30; 12:45pm it's -0.50; that 45-minute swing is worth investigating immediately, often before the crisis is even visible to the brand's own social team. (3) Competitive benchmarking — comparing your brand's sentiment to competitors' produces a 'sentiment lead/lag' metric. A brand with consistent +0.40 sentiment vs competitors averaging +0.20 enjoys a perceptual moat that's worth marketing budget.

The other use case worth mentioning: campaign attribution. After a marketing campaign launch, watching the sentiment shift on brand mentions tells you whether the campaign moved perception (positive sentiment delta) or backfired (negative sentiment delta). Combined with reach and conversion metrics, sentiment closes the marketing-attribution loop.

Sentiment analysis limitations and pitfalls

Three known weaknesses. (1) Sarcasm — 'Wow this product is AMAZING' said sarcastically reads as positive to most sentiment models. Modern transformer models partially handle sarcasm via context, but it's still the largest source of false-positive readings. (2) Domain shift — sentiment models trained on news data often misclassify gaming culture, internet slang, in-group humor. Models need to be fine-tuned on the relevant domain. (3) Multilingual quality — English sentiment is solved at near-human level; Spanish, Portuguese, and Mandarin are good but not as good; smaller languages (Turkish, Vietnamese, Hindi) often have notably worse sentiment models. Brands operating in multiple languages should validate per-language quality before trusting aggregate dashboards.

A fourth pitfall is 'sentiment without context.' A brand can have +0.50 sentiment in a category where competitors are at +0.70 (so it's actually losing) or -0.10 in a category where competitors are at -0.30 (so it's winning). Always benchmark sentiment against competitors and category baseline; absolute sentiment scores in isolation are misleading.

Sentiment analysis vendors and tools

Frequently asked questions