Sentiment analysis goes beyond positive/negative/neutral classification. Production systems need aspect-based sentiment (positive about the product, negative about shipping), intensity scoring (mildly annoyed vs. furious), and temporal tracking to detect shifts.

Sentiment Dimensions

Polarity - The basic positive/negative/neutral classification. Useful for high-level dashboards but too coarse for actionable insights. A product with 60% positive and 40% negative sentiment needs to know what is driving the negative to take action.

Intensity - How strong is the sentiment? A scale from -1.0 to +1.0 captures intensity. “Slightly disappointed” and “absolutely furious” are both negative but require very different responses. Intensity scoring enables prioritization of the most strongly negative items.

Aspect-based sentiment - Sentiment tied to specific aspects of the product or service. A restaurant review might be positive about food quality but negative about wait times. Aspect-based sentiment identifies what specifically is driving overall sentiment, enabling targeted improvements.

Emotion classification - Beyond polarity, classify the specific emotion: frustrated, confused, grateful, excited, anxious. Different emotions suggest different response strategies. A frustrated customer needs a different response than a confused one.

Pipeline Architecture

Pre-processing - Clean and normalize text before analysis. Remove HTML tags, decode emojis and emoticons (they carry strong sentiment signals and should be preserved as text), normalize slang and abbreviations, and handle multilingual content.

Analysis tier - For high-volume streams, use a lightweight classifier for initial polarity scoring. Escalate items above a complexity threshold (mixed sentiment, sarcasm indicators, long text) to an LLM for nuanced analysis. This tiered approach balances cost and quality.

Aggregation - Individual sentiment scores are useful for routing and response. Aggregated sentiment over time, by topic, and by customer segment is useful for strategy. Build both real-time individual scoring and periodic aggregate analysis.

Handling Difficult Cases

Sarcasm - “Great, another update that breaks everything” is negative despite positive words. LLMs handle sarcasm better than traditional classifiers, but it remains the most common source of sentiment misclassification. Flag potential sarcasm for human review when stakes are high.

Mixed sentiment - “The product is amazing but the support is terrible” contains both positive and negative sentiment. Aspect-based analysis handles this naturally by scoring each aspect independently rather than producing a single overall score.

Implicit sentiment - “I have been waiting three weeks for a response” expresses no explicit sentiment but implies strong dissatisfaction. LLMs capture implicit sentiment better than keyword-based approaches, but prompts should explicitly ask the model to consider implied sentiment.

Scale Considerations

Streaming analysis - For real-time monitoring (social media, support tickets), process items as they arrive. Use a fast classification model for immediate scoring and queue items for deeper LLM analysis when needed.

Batch analysis - For historical analysis (survey responses, review datasets), process in batch for cost efficiency. Use batch API pricing when available.

Cost management - At high volume, sending every text to an LLM for sentiment analysis is expensive. Use a cascade: rule-based classification for clear cases (5-star reviews are positive), a small model for moderate cases, and an LLM for ambiguous cases. This reduces cost by 60-80% while maintaining quality on difficult items.