Sentiment analysis is the use of software to estimate the feeling expressed in text or speech, usually as positive, negative or neutral, and sometimes as a score or as specific emotions such as frustration. In contact centers and customer experience programs, it is applied to calls, chats, emails, surveys and reviews to see how customers feel across far more conversations than people could read or listen to.
At a glance
- Sentiment analysis estimates feeling from words, and in some voice products from tone; it does not read minds.
- It is commonly a feature of speech analytics, text analytics, CX or agent assist products.
- Results are scores or labels per conversation, per segment of a conversation, or trended over time.
- It is most useful for finding conversations and topics worth a human look, not as a precise measure of any one customer.
- Accuracy varies by language, channel and industry, so testing on your own data matters.
What problem it solves
Surveys capture how a small share of customers feel, usually after the fact, and often skewed toward those who were very happy or very unhappy. In many contact centers, supervisors can review only a handful of calls per agent each month. That leaves most conversations unexamined, and an angry customer or a growing problem with a product can go unnoticed until it shows up as churn or complaints.
Sentiment analysis can give a rough read on every conversation it is applied to. Managers can sort conversations by negative sentiment to review the ones that went badly, see which contact reasons or products generate the most frustration, and compare how sentiment changes during a conversation, for example whether customers end calmer than they started. In real time, it can alert a supervisor to a conversation going wrong while there is still time to help.
How it works
Getting text. Chats, emails and survey comments are already text. Calls are transcribed first, typically by the same engine used for speech analytics. Transcription errors carry through to the sentiment result.
Scoring. Language models, a form of artificial intelligence (AI), assess the words and phrasing to assign a sentiment label or score. Older approaches relied on lists of positive and negative words; newer ones consider context, which helps with phrases like “not bad”. Some voice products also weigh acoustic signals such as raised volume, fast speech or overtalk.
Breaking it down. Many products score sentiment separately for the customer and the agent, at points through the conversation, and by topic, so you can see that a customer was positive about the product but negative about billing.
Using the results. Scores feed dashboards and trend reports, quality management (QM) workflows that choose which conversations to review, real-time supervisor alerts and, in some products, agent assist prompts. Results can also be compared with survey scores such as customer satisfaction (CSAT) to check whether they line up.
When it matters for buyers
- When only a small share of customers answer surveys and you want a view of the rest.
- When choosing which calls quality teams should review, instead of sampling at random.
- When comparing speech analytics or CX analytics products, where sentiment is often a headline feature with very different quality behind it.
- When operating in several languages, since support and accuracy vary by language.
- When considering decisions about individual agents or customers based on scores, which calls for human review and, in some places, legal advice.
Our contact center as a service page covers where analytics fits in a platform evaluation.
Questions to ask vendors
- Which channels and languages do you support for sentiment, and are accuracy results available for each?
- Do you analyze words only, or also tone of voice?
- Is sentiment scored separately for customer and agent, and at points within the conversation?
- Can we test the product on a sample of our own conversations before signing?
- Is sentiment included in the base product, or a separately priced analytics tier?
- Where is conversation data processed, and is it used to train models outside our account?
- How do supervisors see and act on sentiment in real time?
How it differs from speech analytics
Speech analytics is the broader practice of transcribing and analyzing calls for topics, phrases, silence, compliance language and trends. Sentiment analysis is one type of analysis that a speech analytics product may perform, and it also applies to text channels that have nothing to do with speech. If a vendor sells “sentiment” on its own, ask what else you would need to act on it, such as topic detection, search and quality workflows. Both feed into a wider view of customer experience (CX).