If you think labeling text spans with LLMs is easy, you probably have not tried it yourself (we have! 🙃).

Any method you can think of – be it tagging, matching, or indexing – has flaws.

In our new preprint, we tested them all 💪We also proposed how to improve one of them.

arxiv.org/abs/2601.16946

I look forward to field-testing the app on high schoolers and NLP students!

The app contains step-by-step animations explaining the mechanics of LLMs – all pre-computed using real models.

It targets the sweet spot between "too technical" and "too handwavy" 👌

As a bonus, the supports multiple languages (including 🇨🇿)!

Dal jsem si to jako challenge za covidu, když byly uzavřené hranice okresů. Jen jsem si to rozdělil na dva dny.

Západní okraj mi přišel mnohem hezčí, než východní, ale možná už mě jen druhý den moc bolely nohy 😄

When a model output gets a BLEU score of 12.34: 😶
When you can *see* the errors in the output: 🤯

Check out factgenie, our new error span annotation tool:
github.com/ufal/factgenie

(Just matured into v1.0.1! 🎂)

#nlp #nlg