Preprint
Automatic punctuation restoration with BERT models
Attila Nagy et al., arXiv, 2021
Read the originalRead on October 16, 2026
- What it is
- A 2021 preprint that restores punctuation in text with BERT models for English (TED Talks) and Hungarian. It reports one macro-averaged F1 score per language.
- How much weight it carries
- Preprint that reports a macro average across punctuation marks, so it does not show how commas or periods performed separately. It works on transcripts with the punctuation removed, not on live speech recognition output.
Facts taken from this source
- On English TED Talks transcripts, the best BERT models reached a macro-averaged F1 of 79.8 for punctuation restoration.
“Our best models achieve a macro-averaged ... score of 79.8 in English” (Abstract)