language:

This BERTic model is designed to annotate incorrectly spelled words in text. It utilizes the following labels:

Imagine we have the following Croatian text:

If we convert input data to format acceptable by BERTic model:

The model might return the following predictions (note: predictions chosen for demonstration/explanation, not reproducibility!):

We can observe that in the input sentence, the word prepoznije and nalazaju are spelled incorrectly, so the model marks them with the token (1).

Testing model with generated test sets provides following result:

Precision: 0.9954 Recall: 0.8764 F1 Score: 0.9321 F0.5 Score: 0.9691

Testing the model with test sets constructed using the Croatian corpus of non-professional written language by typical speakers and speakers with language disorders RAPUT 1.0 dataset provides the following results:

Precision: 0.8213
Recall: 0.3921
F1 Score: 0.5308
F0.5 Score: 0.6738

The authors acknowledge the financial support from the Slovenian Research and Innovation Agency - research core funding No. P6-0411: Language Resources and Technologies for Slovene and research project No. J7-3159: Empirical foundations for digitally-supported development of writing skills.

Thanks to Martin Božič, Marko Robnik-Šikonja and Špela Arhar Holdt for developing this model.