Sleep-driven computations in speech processing.

Frost, R. L. A. & Monaghan, P. (2015, June). Poster presented at the New Directions in Implicit and Explicit Language Learning symposium, Lancaster, UK.

One of the primary challenges facing language learners is identifying words and learning the grammatical rules that apply between these wordsPast research into statistical learning has provided conflicting accounts regarding the separability of processing for these tasks. One possibility is that words are first identified and then relations between words are computed (Peña, Bonatti, Nespor & Mehler, 2002), meaning different processes may apply to these tasks. Alternatively, it may be the case that these tasks are resolved in combination during language learning; instead of distinct operations, the same type of learning process may apply to both word identification and grammar acquisition.   A further alternative to signal-driven changes in computation is that learning specific instances and abstraction of structure can be supported by separation of these processes during sleep (Kumaran & McClelland, 2012). Sleep has a profound influence on abstraction and generalisation across a range of tasks (e.g. Fenn, Nusbaum & Margoliash, 2003), including the acquisition and generalisation of language structure (Nieuwenhuis, Folia, Forkstam, Jensen & Petersson, 2013; Gómez, Bootzin & Nadel, 2006). However, in these previous studies, segmentation and generalisation were not simultaneously required of the learner. In this study, we tested directly whether sleep-driven changes in computation could be observed for segmentation and generalisation tasks without changing the speech signal itself. We trained participants on an artificial language that comprised nonadjacent dependencies, then tested their ability to complete tasks of segmentation and generalisation 12- and 24- hours later. Our results show that sleep-related computations lead to improvements for both segmentation and generalisation, but have distinct signatures on learning: Findings indicate the possibility for an enduring benefit of sleep for segmentation, regardless of whether that sleep is immediate or delayed, however findings show a short-term benefit of immediate sleep for generalisation. Such a pattern of results is consistent with a view of word learning and grammar learning as distinct declarative and procedural tasks (Ullman, 2004), which are underwritten by different sleep-based mechanisms.