FullStop: punctuation and segmentation prediction for Dutch with transformers

Abstract

When applying automated speech recognition (ASR) for Belgian Dutch, the output consists of an unsegmented stream of words, without any punctuation. A next step is to perform segmentation and insert punctuation, making the ASR output more readable and easy to manually correct. We present the first (as far as we know) publicly available punctuation insertion system for Dutch that functions at a usable level and that is publicly available. The model we present here is an extension of the approach of Guhr et al. (In: Swiss Text Analytics Conference. Shared task on Sentence End and Punctuation Prediction in NLG Text, 2021) for Dutch: we finetuned the Dutch language model RobBERT on a punctuation prediction sequence classification task. The model was finetuned on two datasets: the Dutch side of Europarl and the SoNaR corpus. For every word in the input sequence, the model predicts a punctuation marker that follows the word. In cases where the language is unknown or where code switching applies, we have extended an existing multilingual model with Dutch. Previous work showed that such a multilingual model, based on “xlm-roberta-base” performs on par or sometimes even better than the monolingual cases. The system was evaluated on in-domain data as a classifier and on out-of-domain data as a sentence segmentation system through full stop prediction. The evaluations on sentence segmentation on out of domain data show that models finetuned on SoNaR show the best results, which can be attributed to SoNaR being a reference corpus containing different language registers. The multilingual models show an even better precision (at the cost of a lower recall) compared to the monolingual models.