Sentiment Analysis Explained: How to Find the Meaning in Mixed Reviews

Your support team gets a standing ovation. Your refund process gets booed. Both reviews arrive in the same sentence: "The agent was brilliant, but waiting three weeks for my refund was frustrating."

File that under "positive" and the refund problem disappears. Choose "negative" and the agent's good work gets lost. Call it "neutral" and you've somehow turned two strong opinions into a shrug.

Sentiment analysis helps make sense of feedback like this. For customer and employee insights teams, its real value comes from understanding what people are evaluating and where their experiences differ.

So, what is sentiment analysis?

Sentiment analysis is a natural language processing (NLP) technique that identifies positive, negative, or neutral opinions expressed in text. It can evaluate an entire comment or the sentiment toward individual topics within it.

Applied to surveys, reviews, social media support conversations, and employee feedback, it helps turn written opinions into patterns you can actually understand.

Our example from earlier contains mixed sentiment: praise for the agent and criticism of the refund speed. That difference is important here. Neutral feedback expresses no clear positive or negative evaluation. Mixed feedback contains both, or neither. People can appreciate one part of an experience while being thoroughly unimpressed by another.

Capturing these nuances has taken years of development. Researchers were classifying sentiment with machine learning in the early 2000s. Since then, advances in language models have improved how systems interpret context, helping distinguish different opinions within a single comment instead of relying on positive and negative words alone.

What sentiment reveals beyond a survey score

A three-star rating gives you a result. The comment can give you the drivers behind it.

Two customers might award the same score for different reasons. One appreciates the support but finds the price too high. The other considers the price fair but struggles to get help. Matching scores, different problems. Employee feedback has the same layers. Someone can value their colleagues while finding their workload unsustainable. An overall engagement score leaves those details bundled together.

Research from Wiley, combining sentiment analysis with ratings in 2,493 airline reviews, distinguished responses that shared a numerical rating but expressed different sentiment. The combined approach helped identify service dimensions needing attention.

💡 Tip: running a driver analysis explains the dimensions impacting your rating (up or down), and topic sentiment makes the driver analysis much more effective.

Sentiment analysis adds that structured view of people's evaluations alongside NPS, CSAT, or star ratings. It helps you explore what the score leaves unexplained. The comments still need interpretation. They offer evidence about an experience, rather than proof that a particular issue caused the rating.

Why topic-level sentiment makes feedback more useful

Whole-comment sentiment works well when the message is focused: "The delivery was excellent." But customers often have more than one thing to say. A single comment can praise the product, question the price, and complain about delivery.

An internal Caplena analysis of roughly 1.5 million online-review entries and 7.4 million survey-feedback entries found that online reviews averaged 209 characters, compared with 107 for survey feedback. That's almost twice the length. But when you bring these sources together, the challenge goes beyond how much people write: you need to distinguish what they like from what they don't, even within the same comment.

Topic-level sentiment analysis, also called aspect-based sentiment analysis, separates those evaluations.

Now the interpretation becomes more useful. Across hundreds of comments, repeated praise for agents alongside criticism of refund delays suggests investigating the refund process before assuming service friendliness is the problem.

The same principle applies to "I love working with my team, but our workload is unsustainable." Keeping those evaluations separate helps you recognize a strength while investigating a concern. In Caplena, topics and sentiment are separate parts of the analysis. One topic can contain both positive and negative feedback, making differences easier to see.

How does sentiment analysis work?

Automated sentiment analysis uses language patterns to classify expressed opinions. Two broad approaches underpin the work.

1. Rule-based methods use dictionaries and predefined rules, such as recognizing positive words and adjusting for negation. They can work well within a defined scope, although unfamiliar expressions and context require additional rules.

2. Machine-learning methods, including modern language models, learn patterns from data and use surrounding language to interpret a statement.

Context does a lot of heavy lifting. "Unpredictable" might praise a film's plot while criticizing a delivery service. Nobody wants their pizza to arrive with a surprise ending.

In Caplena, large language models assign topics and sentiment, considering project context, question wording, and topic descriptions. It also assigns topics based on optional reviews done by humans. Praise for support and criticism of refund speed stay separate. You can fine tune assignments, refine descriptions, and update the analysis. It can pick up on irony and sarcasm too, e.g. "Brilliant. Another three weeks for a refund." Context and cues like exaggeration, emojis, and capitalization help the AI recognize criticism behind seemingly positive words.

Sentiment versus emotion: what does each one tell you?

Sentiment describes a positive, negative, or neutral evaluation. Emotion analysis adds more specific labels, such as frustration, confusion, relief, or enthusiasm. Think of Inside Out: Joy and Anger offer more detail than a simple thumbs-up or thumbs-down. Real feedback, however, rarely gives us such neatly labeled characters.

Our refund customer explicitly expresses frustration. Another might be confused about the instructions. Both could evaluate the process negatively, but their comments suggest different investigations: does it need to move faster or become easier to understand?

When that detail is important, Caplena's Smart Columns can add emotion labels to feedback alongside topic-level sentiment. A whole-comment emotion doesn't necessarily apply to every topic mentioned; our frustrated customer still praised the agent. Emotion labels describe what the language supports, with room for ambiguity where the evidence is less clear.

💡 Tip: Give negative feedback a closer look. Use Smart Columns to group comments by emotions such as frustration or disappointment, then compare the topics mentioned negatively in each group. This helps you focus your investigation and prioritize the areas triggering customers the most. .

How to read sentiment scores without losing the detail

Once feedback is classified, you can compare positive, negative, and neutral shares by topic, group, or time period.

A net sentiment score summarizes the balance.
Net sentiment = percentage of positive comments − percentage of negative comments

Suppose 100 comments are classified as 60 positive, 25 negative, and 15 neutral. Net sentiment is +35, on a scale from −100 to +100. It's a useful summary, especially when feedback has no accompanying rating. But the headline number has a talent for leaving things out.

A score of zero could mean every comment is neutral. It could also mean half are positive and half negative. The average looks calm; the feedback may be anything but. Keep the distribution, topics, comment count, and representative examples nearby. Ten comments and a consistent pattern across thousands deserve different levels of confidence.

No rating? You can still calculate net sentiment from the text alone. That makes it useful for feedback such as social media posts and support messages, where customers describe their experience without giving it a score.

When ratings do exist, net sentiment adds another perspective. If NPS or star ratings differ across countries, compare sentiment and topics alongside them: do customers describe similar experiences despite giving different scores, or do their comments reveal specific service problems? Compare similar feedback sources and customer groups, and keep the measures distinct: net sentiment reflects written opinions; NPS measures willingness to recommend.

How reliable is sentiment analysis anyway?

Reliability depends on the language, task, and context. A single accuracy percentage rarely tells the whole story. Take "Very little." After "What did you enjoy?" It suggests dissatisfaction. After "What needs improving?" it could suggest the opposite. Two words, quite a big difference.

To check how well the analysis handles your feedback, build in three habits.

• Preserve context. Keep questions and relevant background attached to the feedback.

• Review varied examples. Include positive, negative, neutral, and ambiguous comments.

• Investigate surprises. Read the underlying feedback before presenting an unexpected finding.

Review whether the analysis captures the distinctions your team needs, including less frequent issues. That gives you a stronger basis for decisions than an overall accuracy figure alone.

From sentiment patterns to better business questions

Suppose negative feedback about refund speed increases. Before announcing a company-wide refund crisis, check where the change comes from. Is it concentrated in one market, payment method, or period? What do the comments describe? Share that evidence with the team responsible for the process.

Frequency and negativity help focus the investigation, but they aren't a complete priority list. A common inconvenience might matter less than a rarer, more serious failure. Consider reach, consequences, and relevant business measures. Also check whether your feedback mix changed. More support tickets and fewer public reviews could lower overall sentiment even while results within each channel remain steady.

"

When a backlash on social media sparked around our new product pricing, I could show within 15 minutes that sentiment among current customers was still positive. Without multi-source analysis, I'd have had to check each survey separately.

Monika Wirtz

CX Platform Owner

Researchers from the University of Illinois used nearly 800,000 observations across Amazon electronics products and found that sentiment-related information improved forecasts of future ratings. Text can carry useful signals, although every change shouldn't be treated as a prediction of churn or revenue.

Caplena's Insights Radar and Tasks connect pattern monitoring with assigned follow-up, giving findings an owner and a next step. This goes beyond flaggin issues. What sparks delight deserve attention too: Insights Radar also shows you what to repeat.

Start with a question your feedback can answer

What do customers criticize about refunds? What do employees appreciate about onboarding? Choose one question, identify the relevant topics, and inspect the sentiment and comments behind the patterns. Add ratings or emotions where they help explain the experience.

With Caplena, you can analyze topics and sentiment together, refine the analysis, and explore customer and employee feedback with the evidence close at hand. That gives you a clearer next step. Your brilliant support agent gets the credit. Your refund process gets the attention.