Sentiment analysis

Sentiment analysis estimates whether a customer is satisfied, angry or neutral in a call or a message. In contact centers it is usually run on the call transcript, and some systems also take acoustic cues such as tone of voice, speaking rate and interruptions into account. The result is typically a single label for the whole conversation or a curve that changes as the call goes on. It is an estimate: irony, cultural differences in how people express themselves and very short answers can easily throw it off.

How it works

Early methods relied on lists of positive and negative words. Today, classification models and large language models read the context of a sentence before deciding. The conversation is split into segments, each segment is scored, and the customer's and the agent's sides are assessed separately. Turkish has its own traps, such as positive statements built with a negative ("hiç fena değil", literally "not bad at all") and regional turns of phrase, so a model needs to be tested on Turkish calls. The analysis is only as good as the call transcript underneath it.

Why it matters

Nobody can listen to every call, and sentiment analysis helps supervisors decide which ones to review first. A call that is going badly can be tied to live monitoring or to a handover rule; a caller getting upset is also a common trigger for handing a call from AI to a person. A sentiment score does not replace a CSAT survey, though: one is an estimate, the other is what the customer actually said. Rather than using the score on its own to assess agents, check it against a sample of calls you have listened to.