Improved models for automatic punctuation prediction for spoken and written text

This paper presents improved models for the automatic prediction of punctuation marks in written or spoken text. Various kinds of textual features are combined using Conditional Random Fields. These features include language model scores, token n-grams, sentence length, and syntactic information extracted from parse trees. The resulting models are evaluated on several different tasks, ranging from formal newspaper text to informal, dictated messages and documents, and from written text to spoken text. The newly developed models outperform a hidden-event language model by up to 26% relative in F-score. Evaluation of punctuation prediction on erroneous ASR output as well as evaluation against multiple references is not straightforward. We propose modifications of existing evaluation methods to handle these cases.