Introduction to Exemplar-Based Models of Language Acquisition and Use Rens Bod and David Cochran School of Computer Science University of St Andrews St Andrews, Scotland, UK {rb,davec}@cs.st-and.ac.uk 1 The notion of exemplar Exemplar-based models have become part and parcel of linguistics: they are used in phonetics and phonology (Johnson 1997; Pierrehumbert 2001), in morphology (Skousen 1989; Krott et al. 2002), syntax (Bod 1998; Tomasello 2003; Reali and Christiansen, in press) and semantics (Batali 2002). The appeal of exemplar models is that they can explain phenomena that are problematic for rule-based models. These phenomena include frequency effects in comprehension and prediction, gradual language change, and the dynamics of language learning. Exemplar models aim to capture the detailed episodic memory of linguistic events that humans retain, by storing exemplars over time and comparing new input against them. But what exactly are exemplars? A quick look at the literature suggests that there is no agreement on what an exemplar is and what it does. One goal of this workshop is to clarify the various notions of exemplar and to investigate the possibility of a unifying framework. Following Hay and Bresnan (2006), there are at least two strands of exemplar-based models that were developed rather independently of one another, and to which they refer as “phonetic exemplar theory” and “syntactic exemplar theory”. 2 Phonetic Exemplar Theory Exemplars were introduced as a psychological model for perception and categorization by Hintzman (1986), Nosofsky (1986) and others. These exemplar models have been investigated in speech perception and production where the representation of lexical items are conceptualized as underlying, abstract forms (Johnson 1997). Every time we encounter a particular word, we store the phonetic memory of that word together with non-linguistic detail. When a new word is encountered, it is classified according to its similarity to the exemplars already stored. Its similarity to any single stored exemplar can be computed as its ‘distance’ from the exemplar in the parameter space. The most probable labelling given the labelling of the exemplars in the neighborhood is computed on the basis of stored exemplars. Pierrehumbert (2001) views exemplar theory as providing a way to formalize the detailed phonetic knowledge that speakers have about the categories of their language. The acquisition of this knowledge can be understood in terms of the acquisition of a large number of memory traces of experiences. This exemplar-based view appears in many different frameworks. It has been used by Skousen (1989), Krott et al. (2002) and Daelemans and van den Bosch (2005) to model analogical processes in various subfields. The differences between the frameworks lie in the definitions of similarity and distance metrics and the functions that are maximized. But the common underlying idea is that all previous linguistic experiences are stored as exemplars and linguistic behavior is predicted by the most similar stored exemplars. 3 Syntactic Exemplar Theory According to the exemplar-based conception in syntax, there are no explicit rules of grammar. Instead, a ‘grammar’ arises out of analogical generalizations over stored chunks of previous language experiences. Language experiences can