"Hallo" has CEFR of C2

[hieuristic]

Not sure why the simple "Hallo" has the CEFR of C2 and ZIX of -10.0

to reproduce:

from zix.understandability import get_zix, get_cefr

w = "Hallo"
z = get_zix(w)
l = get_cefr(z)
print(f"{w} | ZIX: {z} | level: {l}")

[rnckp]

Thanks! Let me investigate.

[rnckp]

Thanks for flagging this — good catch.

The understandability index is trained on paragraph-level text and is not intended to assess CEFR levels for single words, very short sentences, or text fragments. That said, I see that this behavior might be surprising for users.

As a short-term mitigation, I’ve added a user warning when the package is used on inputs of five words or fewer. I have also added a note to the README and a comment in the demo notebook.

In addition to that I ran a quick experiment by augmenting the training data with very short sentences and single-word samples, which does improve estimates for these edge cases. However, I need to evaluate potential downstream effects before changing the training data composition.

For now, we plan to keep the primary use case focused on paragraph-level text. Estimating the difficulty of individual words is likely better handled via word-frequency lists or official CEFR vocabulary resources (e.g., Goethe-Institut lists).

I’ll keep this issue open while I explore the available options further. Thanks again for raising this and for your interest in the project.