How Thematic Sentiment Analysis works
Thematic Sentiment Analysis uses our proprietary algorithm. This algorithm adds a sentiment score to comments and themes. Our solution has been trained specifically on feedback (survey responses, reviews, support chats) and has been optimized to predict with a high accuracy on this data.
1. What Thematic Sentiment Analysis does
Each comment is split into segments, each segment is tagged with themes and a sentiment score. This means that a single sentence might have both positive and negative sentiment.
For easy visualization, we split sentiment into three categories: positive, neutral and negative. (The example above does not feature neutral.)
Thematic also aggregates sentiment scores across all themes and comments.
For example, in Thematic you can view volume of themes by sentiment. (Grey shows neutral.)
You can export an image that captures sentiment across several themes:
2. How accurate is Thematic Sentiment Analysis?
We constantly update our sentiment model by both adding training data and replacing the technology behind it. Our most recent test on 7000 test examples has shown accuracy of 99.9%, when it comes to differentiating between positive and negative sentiment.
This means that for every 1000 examples, our model made only 1 error.
Neutral vs positive or neutral vs negative is more tricky. This is where even people disagree with each other often. Our accuracy on all 3 classes is 93% and is in line with human agreement.
3. Why is Thematic Sentiment Analysis different?
You can easily build your own sentiment analysis model or get an LLM to guess the sentiment. Most LLMs are highly accurate, and they can even deal with sarcasm. But Thematic Sentiment Analysis is purpose-built for business feedback — combining human-tuned precision, intensity scoring, and scalable performance for consistent, real-world results.
✅ Human-Guided AI for Better Accuracy
Our model combines AI speed with human insight. It learns from human-reviewed examples, which helps it handle tricky language and reduces common AI biases — so the results align more closely with how real people would judge sentiment.
✅ Smarter Training Through Multi-Model Consensus
Instead of learning from one single source, the model uses input from multiple AI systems to “vote” on the correct sentiment. This creates higher-quality training data, which leads to more reliable and balanced insights for your team.
✅ Sentiment Intensity, Not Just Positive/Negative
Our model doesn’t just tell you whether a comment is positive, neutral, or negative — it also measures how strong the emotion is. This lets you spot not just what customers/employees feel, but how intensely they feel it.
✅ Fast, Scalable, and Enterprise-Ready
Designed for real-world use, our model can process millions of feedback comments in real time — without requiring expensive infrastructure. You get fast insights at scale, whether for daily reports, live dashboards or triggered actions such as personalization.
✅ Always Learning, Always Up-to-Date
The model is built for continuous improvement. As new data becomes available, it retrains automatically — no manual effort needed. This means your sentiment insights stay fresh, accurate, and relevant as customer language evolves.