Punctuation restoration in Swedish through fine-tuned KB-BERT

TL;DR

This paper introduces prestoBERT, a fine-tuned Swedish BERT model for automatic punctuation restoration, achieving competitive F1-scores and demonstrating potential benefits for NLP tasks like speech-to-text and automated text processing.

Contribution

The study presents a novel fine-tuning approach for Swedish BERT to restore punctuation, with a detailed evaluation against human performance and international models.

Findings

prestoBERT achieved an F1-score of 78.9

Comparable performance to Hungarian and Chinese models

Human evaluators scored 81.7 but struggled with punctuation consistency

Abstract

Presented here is a method for automatic punctuation restoration in Swedish using a BERT model. The method is based on KB-BERT, a publicly available, neural network language model pre-trained on a Swedish corpus by National Library of Sweden. This model has then been fine-tuned for this specific task using a corpus of government texts. With a lower-case and unpunctuated Swedish text as input, the model is supposed to return a grammatically correct punctuated copy of the text as output. A successful solution to this problem brings benefits for an array of NLP domains, such as speech-to-text and automated text. Only the punctuation marks period, comma and question marks were considered for the project, due to a lack of data for more rare marks such as semicolon. Additionally, some marks are somewhat interchangeable with the more common, such as exclamation points and periods. Thus, the data set had all exclamation points replaced with periods. The fine-tuned Swedish BERT model, dubbed prestoBERT, achieved an overall F1-score of 78.9. The proposed model scored similarly to international counterparts, with Hungarian and Chinese models obtaining F1-scores of 82.2 and 75.6 respectively. As further comparison, a human evaluation case study was carried out. The human test group achieved an overall F1-score of 81.7, but scored substantially worse than prestoBERT on both period and comma. Inspecting output sentences from the model and humans show satisfactory results, despite the difference in F1-score. The disconnect seems to stem from an unnecessary focus on replicating the exact same punctuation used in the test set, rather than providing any of the number of correct interpretations. If the loss function could be rewritten to reward all grammatically correct outputs, rather than only the one original example, the performance could improve significantly for both prestoBERT and the human group.