--

Nice article !

Actually, CharacterBERT can be very useful in many other domains, as Text to Code issues.

However, I think you miss an interessting comparison with BPE technique, which is also a technique to avoid vocabulary dependencies and use a character embeding for tokens.

--

--

Abdelkader Rhouati

Data scientist & Ph.D. researcher on AI. My area of expertise is around Deep Learning, NLP, and XAI