Language acquisition in the absence of explicit negative evidence: can simple recurrent networks obviate the need for domain-specific learning devices?

Rohde & Plaut, 1999 argue that their work with Elman' s simple recurrent ne twork (henceforth, SRN) "suggests that learning the structure of natural la nguage may be possible despite a lack of explicit negative feedback.,, in t he absence of detailed innate language-acquisition mechanisms". They furthe r argue that "a key factor in overcoming the 'logical problem' of language acquisition (Baker&McCarthy, 1981) is the use of implicit negative evidence ." (Implicit negative evidence is information about something that does not appear when it was predicted to appear.) R&P are surely correct that some versions of the simple recurrent network d o not rely on negative evidence and that such networks are able in some cas es to utilize implicit negative evidence.(1) But R&P do not show that these models avoid the kinds of errors that children make, do not show that thes e models derive the same generalizations as children do, and do not show th at these models use indirect negative evidence in ways that would obviate t he need for innate, domain-specific learning devices. All that they offer i s a simulation of a tiny fragment of a simplified version of English; they do not fit the model's data against any data derived from children. Their s ystem does not provide any sort of syntactic or semantic representation of the sentences that it is exposed to, and it does not make a principled dist inction between infrequent and ungrammatical sentences. This is not enough to establish the adequacy of the model, and more careful inspection reveals a serious, principled limitation that stems directly from its treatment of implicit negative evidence.