Sequential Structure Suffices to Solve Nativist Puzzles

The debate between hierarchical versus sequential structure in language acquisition has recently flared up again (cf. Frank, Bod & Christiansen 2012; Pesetsky 2013). Roughly, the nativist view on language endorses that human language acquisition is guided by innate rules that operate on hierarchical structures. The empiricist view assumes that language acquisition is the product of abstractions from empirical input but leaves it as an open question whether sequential or hierarchical structure is needed. Some empirical models use sequential structure (e.g. Reali & Christiansen 2005) while other models are based on hierarchical structure (Bod 2009; Bod & Smets 2012). Much work in empirical language acquisition has focused on a relatively small set of phenomena such as auxiliary fronting. For example, Reali & Christiansen (2005) argued that auxiliary fronting could be learned by linear models based on sequential structure, though Kam et al. (2008) showed that the success of these models depend on accidental English facts. Other empiricist approaches have taken the notion of structural dependency together with a combination operation as minimal requirements (e.g. Bod 2009), which overcomes the problems raised by Kam et al. (2008). In Bod and Smets (2012) it was shown that a much larger set of phenomena can be learned by an empiricist computational model. These phenomena are well-known in the generativist literature (Ross 1967; Adger 2003) and are related to wh-questions, relative clause formation, topicalization, extraposition and left dislocation. It turned out that these hard cases can be learned by an unsupervised tree-substitution grammar induction algorithm that returns the sentence with the best-ranked derivation for a particular phenomenon, using only a very small fraction of the input a child receives. However, Bod and Smets (2012) also observed that these nativist cases were learned by using relatively shallow structures with little or no hierarchy. This raised the question as to how much structure is actually needed to learn these syntactic constraints. In the current paper, we present a very simple model that reduces all syntactic structuring to concatenations of substrings without any hierarchy. We show that almost all results obtained by the hierarchical grammar in Bod & Smets (2012) can also be learned by means of a sequential grammar using substringconcatenation only. It should be stressed that the essence of the debate between nativism and empiricism lies often in the relative contribution of prior knowledge and linguistic experience (cf. Lidz et al. 2003; Ambridge & Lieven 2011; Clark and Lappin 2011). Following the nativist view, the linguistic evidence is so hopelessly underdetermined that innate components are necessary. This Argument from the Poverty of the Stimulus can be phrased as follows (see Pullum & Scholz 2002 for a detailed discussion):