Proceedings of the Annual Meeting of the Cognitive Science Society

Determining the correct form of a verb in context requires anunderstanding of the syntactic structure of the sentence. Re-current neural networks have been shown to perform this taskwith an error rate comparable to humans, despite the fact thatthey are not designed with explicit syntactic representations.To examine the extent to which the syntactic representationsof these networks are similar to those used by humans whenprocessing sentences, we compare the detailed pattern of er-rors that RNNs and humans make on this task. Despite signif-icant similarities (attraction errors, asymmetry between singu-lar and plural subjects), the error patterns differed in importantways. In particular, in complex sentences with relative clauseserror rates increased in RNNs but decreased in humans. Fur-thermore, RNNs showed a cumulative effect of attractors buthumans did not. We conclude that at least in some respects thesyntactic representations acquired by RNNs are fundamentallydifferent from those used by humans.