Sentiment Analysis

General definition

Sentiment Analysis is the automatic classification of text by the attitude it expresses. The simplest form assigns a polarity label (positive, negative, neutral) or a score on a scale; richer forms detect discrete emotions (anger, joy, fear), rate urgency, or perform aspect-based analysis that attaches a sentiment to each thing mentioned (“the app is fast but support is slow” is positive about speed and negative about support). It is one of the oldest applied tasks in natural language processing and one of the most widely deployed.

• Lexicon methods: word lists with polarity weights; cheap, transparent, weak on sarcasm and context
• Trained classifiers: models such as fine-tuned BERT variants trained on labelled examples; fast, accurate within the domain they were trained on
• LLM-based: a large language model prompted to rate tone, explain why and handle nuance; flexible and multilingual, slower and more expensive per message
• Hybrid: a small classifier screens every message and an LLM reviews the borderline or high-stakes ones

In messaging products sentiment is an operational signal rather than a research output. Support platforms use it to prioritise queues, trigger escalation to a human when a conversation turns negative, and measure satisfaction continuously as a proxy for CSAT surveys that most customers never answer. Community and marketplace apps use it as one input to content moderation, catching hostility that keyword lists miss. Product teams aggregate it by feature or release to see what changed.

Its limits are well known. Sarcasm, mixed messages, domain jargon and short texts defeat simple models; cultural and linguistic differences shift what counts as negative; and a score on one message says little without the conversation around it. Treat sentiment as a trigger for attention, not a verdict, and keep a human in the loop for anything consequential.

In the Ethora ecosystem

In Ethora, sentiment scoring is something an AI agent does rather than a separate module. An agent attached to a room or channel in the AI SDK sees the messages, can be prompted to rate tone on each one or across the thread, and can act through tool calling: tag the conversation, notify a supervisor, or post a summary to another room. Because the model is a per-agent choice, that scoring can run on a self-hosted model on a dedicated deployment when transcripts must not leave the customer’s environment.

For moderation the Trust & Safety system is the enforcement layer: stop words, regular expressions and AI auto-moderation with block, censor, quarantine-and-report or report-only outcomes, and reports delivered to your own systems over a message bus, API or email. A sentiment signal from an agent and a policy hit from TnS can both land in the same dashboard, which is how a conversational AI deployment stays accountable.