29 november 2007

Lezing door prof. dr. R.W.J. Kager
Titel: Inducing abstract phonotactic constraints from connected speech 
Plaats: Symposiumzaal, Meertens Instituut
Aanvang: 16.00 uur 

Inducing abstract phonotactic constraints from connected speech
René Kager (Utrecht Institute of Linguistics)

The issue of whether linguistic knowledge is abstract or statistical in nature remains controversial. Proponents of purely statistical models of language claim that results of acquisitional studies and computational interpretations imply a radical reduction of the role of abstract grammatical knowledge. Proponents of abstractness, on the other hand, have argued that although statistical learning may be a major learning mechanism, it pre-supposes a basic representational alphabet, while purely statistical models are not able to account for cases of genuine rule learning. Here I will discuss some recent work from our group suggesting that low-level statistical properties of the language are represented by native listeners in the form of abstract constraints, which indirectly provides evidence for an abstract constraint induction mechanism in the learner. I will focus on a set of constraints which govern the co-occurrence of consonants with identical place of articulation in Dutch. These affect the processing of non-words (Kager & Shatzman 2007), as well as speech segmentation of artificial languages (Boll & Kager in preparation). Finally, I will present preliminary results of a computational model (Adriaans & Kager in preparation) which induces abstract phonotactic constraints from connected speech. This model collects statistics on low-level segment distributions, proposes constraints on the co-occurrence of individual segments, and generalizes from these constraints into more abstract ones. The model will be compared with other models for speech segmentation based on statistics only.