Sentiment analysis
Sentiment analysis is the automated classification of a piece of text as positive, negative, or neutral toward a named entity. In 2026, modern systems use a single LLM pass instead of the lexicon-and-regex stacks that dominated until 2022, lifting accuracy on sarcasm and contextual statements from roughly 70% to the high 80s. The number matters because every downstream alert, share-of-voice cut, and crisis trigger inherits its error rate from the sentiment layer beneath it.
What it is
Sentiment analysis takes a piece of text · a tweet, a comment, a review, a news headline · and outputs a label saying whether the writer feels positive, negative, or neutral about a target entity. Modern systems extend the label set to include emotion (anger, joy, fear) and intent (complaint, recommendation, question). The output is usually a label plus a confidence score.
The hard cases are sarcasm, mixed sentiment in a single sentence, and cultural context. "Great, another launch event" carries opposite meaning depending on the surrounding thread. Lexicon-based classifiers miss this almost every time. LLM-based classifiers miss it less often but still miss it.
Why it matters
Every brand-monitoring metric downstream of sentiment inherits its error rate. A 15% sentiment error rate becomes a 15% error rate on the negative-mention alert, on the crisis trigger, and on the share-of-voice split between favourable and unfavourable mentions. Sentiment is where the bias enters the system.
It also drives what the marketing team sees first. If 8% of legitimately negative mentions are misclassified as neutral, those mentions never surface · and the team misses them.
How Sentia uses it
Sentia runs a single Gemini Flash-Lite pass per event that classifies relevance, sentiment, virality, and crisis risk in one call. One model decision per mention, no chained classifiers, no regex pre-filters. The single-pass design avoids the cascading-error problem of stacked rule-based systems and keeps cost predictable at roughly $0.0007 per post.
When sentiment is low-confidence, the mention surfaces as neutral and gets flagged for human review on the monitor inbox. We would rather under-classify than mislabel a sarcastic post as praise.
Related concepts
Related Reading
Start with one monitor. Free.
Add a brand, paste a couple of competitor handles, and see your first calibrated readout in under five minutes.