Search

There have been some suggestions in linguistics and cognitive science that humans process continuous speech by routinely chunking it up into smaller units. The nature of the process is open to debate, which is complicated by the apparent existence of two entirely different chunking processes, both of which seem to be warranted by the limitations of working memory. To overcome them, humans seem to both combine items into larger units for future retrieval (usage-based chunking), and partition incoming streams into temporal groups (perceptual chunking). To determine linguistic properties and cognitive constraints of perceptual chunking, most previous research has employed short-constructed stimuli modeled on written language. In contrast, we presented linguistically naïve listeners with excerpts of natural speech from corpora and collected their intuitive perceptions of chunk boundaries. We then used mixed-effects logistic regression models to find out to what extent pauses, prosody, syntax, chunk duration, and surprisal predict chunk boundary perception. The results showed that all cues were important, suggesting cue degeneracy, but with substantial variation across listeners and speech excerpts. Chunk duration had a strong effect, supporting the cognitive constraint hypothesis. The direction of the surprisal effect supported the distinction between perceptual and usage-based chunking.

Human perception relies on chunking up an incoming information stream into smaller units to make sense of it. Evidence of chunking has been found across different domains, including visual events, music, and dance movement. It is largely uncontested that language processing must also proceed in smaller chunks of some kind. What these online chunks consist in is much less understood. In this paper, we propose that cognitively relevant chunks can be identified by crowdsourcing listener perceptions of chunk boundaries in real-time speech, even if the listeners are non-native speakers of the language. We present a paradigm in which experiment participants simultaneously listen to short extracts of authentic speech and mark chunk boundaries using a custom-built tablet application. We then test the internal validity of the method by measuring the extent to which fluent L2 listeners agree on chunk boundaries. To do this, we use three datasets collected within the paradigm and a suite of different statistical methods. The external validity of the method is studied in a separate paper and is briefly discussed at the end.