How AI Assigns Topics (Legacy AI)

Understand how AI labeling works, how it’s trained, and how your data is handled.

Open-ended responses often contain the richest insights, but analyzing them at scale is rarely straightforward. To help with this, Caplena’s AI automatically tags each response with the most relevant topics, turning free text into structured data you can actually work with.

Below, we’ll walk through how the AI works, how it’s trained, and how your data is handled.

How the AI Works

Caplena’s AI doesn’t rely on basic keyword matching. It understands the meaning behind the words, including synonyms, sentiment, and even negations.

For example:

• It knows that “overpriced” and “too expensive” mean the same thing — no need to manually define synonyms.
• It detects sentiment automatically. For sentiment-enabled topics, only one sentiment (positive, negative, or neutral) will be applied per response. So a comment won’t be tagged as both “Service (Positive)” and “Service (Negative).”

Want to go deeper? Download our whitepaper here.

AI Training Process

Caplena’s AI gets smarter in three stages:

Pretraining

The AI is first trained on a massive set of unstructured text — essentially a large portion of the internet. This gives it broad language understanding, including synonyms, antonyms, negations, and more.

Training on Caplena’s Data

Next, we train it on millions of real, hand-reviewed responses across many industries. This helps the AI develop a deep understanding of how real people express themselves in feedback.

Fine-Tuning (Optional)

Finally, you can fine-tune the AI on your own data by reviewing and correcting a small sample of responses. This step personalizes the AI to your company’s tone, terminology, and unique context.

Use Focus Mode to review the rows where the AI is least confident. That’s where your feedback has the biggest impact.

Model Architecture

This is part of our secret sauce and can’t be shared in detail, but here’s what we can say:

• We use a heavily modified Transformer architecture, similar to the models behind tools like ChatGPT.
• Technically, the AI performs multi-label classification — meaning each comment can get zero, one, or many topics assigned.

Want to dig into the tech? Here you can find a great intro to Transformers.

Is My Data Shared?

No. Your data is never shared with other users. Here’s how it works:

• The AI learns general patterns from tagged data — e.g., that “The package arrived late” maps to “Slow delivery.”
• These learnings are stored as mathematical weights, not as readable text.
• No identifiable or original text is ever exposed, not even internally.

Only the results of topic assignments are visible to you, and the model is never public.