Prosodic phrase break prediction: problems in the evaluation of models against a gold standard

The goal of automatic phrase break prediction is to identify prosodic-syntactic boundaries in text which correspond to the way a native speaker might process or chunk that same text as speech. This is treated as a classification task in machine learning and output predictions from language models are evaluated against a ‘gold standard’: human-labelled prosodic phrase break annotations in transcriptions of recorded speech - the speech corpus. Despite the introduction of rigorous metrics such as precision and recall, the evaluation of phrase break models is still problematic because prosody is inherently variable; morphosyntactic analysis and prosodic annotations for a given text are not representative of the range of parsing and phrasing strategies available to, and exhibited by, native speakers. This article recommends creating automatically-generated POS tagged and prosodically annotated variants of a text to enrich the gold standard and enable more robust ‘noise-tolerant’ evaluation of language models.