Paper presented at the 4th International Conference on the Evolution of Language. Cambridge, MA.
We first review empirical studies that directly compare SL in humans and primates across three categories (Conway & Christiansen, 2001): the learning of arbitrary, fixed sequences; statistical learning; and the learning of hierarchical structure. Although there is considerable overlap between primates and humans on these tasks, primates nonetheless appear rather limited in their ability to learn the structure of complex hierarchical sequences. Because language fundamentally involves hierarchical structure as the basis for unbounded productivity, we hypothesize that computational limitations on hierarchical learning is an important factor in explaining why primates have not evolved complex syntactic communication.
A series of neural network simulations provide support for this hypothesis. Simple recurrent networks (SRNs) representing humans and primates were trained on three SL tasks and one language learning task. The "primate" SRNs differed from the "human" SRNs only by the addition of noise in the feedback connections. The rationale behind the simulations was to match the performance of each group of networks to the relevant behavioral SL data, and then to test their ability to acquire linguistic structure. In the first task, the SRNs were required to learn five-element, fixed sequences. Both SRN groups achieved a comparable level of performance, roughly equal to that of chimpanzees and preschool children (Kawai & Matsuzawa, 2000). In the second task, the SRNs were tested on their ability to extract statistical information from a speech-like stream (Hauser et al., 2001); again, the networksO performances mirrored that of the empirical data, with no species differences evident. In the third task, the SRNs were tested on their ability to hierarchically combine nesting cups (Johnson-Pynn et al., 1999). The networks showed the same pattern of results as the empirical data: the primate, but not the human, SRNs were limited with respect to their hierarchical learning. Finally, we trained the networks on a simple language involving limited recursion in the form of relative clauses. We found that the primate SRNs had problems acquiring the recursive structure of the language, whereas the human SRNs were successful.
Together, the experimental data and our simulation results suggest that
differences in SL - specifically, in hierarchical learning - provide a partial
explanation for why humans but not other primate species have developed
language with complex syntactic structure. Thus, the ability to learn and
process hierarchically organized temporal sequences, which is limited in
non-human primates, appears to be a crucial piece of the language evolution
puzzle.
Please email suggestions/errors to mhc27@cornell.edu |