Running head: Acquisition of Anaphora by SRNs The Acquisition of Anaphora by Simple Recurrent Networks

This paper describes experiments in applying the Simple Recurrent Network (SRN) architecture (Elman 1991, 1993) to the task of learning the grammatical dependencies involved in re exive and pronominal anaphora. This task requires more re ned sensitivity to grammatical structure than others that have been studied in previous explorations of the language learning capacity of SRNs. When evaluated quantitatively, SRNs performed very well in assigning reference to pronouns and re exives. However, the manner in which SRNs achieved such performance and the generalizations that the networks learned diverged in certain key respects from those of the target grammar: (i) linear N-V-re exive and N-V-pronoun sequences which do not stand in a relevant structural relation nonetheless signi cantly a ect the network's interpretations, and lead to errors unlike those made by humans; (ii) the network represents di erent sentence types in distinct fashions, inhibiting systematic generalization from one to the other; (iii) the network's knowledge of the conditions on anaphoric dependencies is item speci c and does not generalize to novel lexical items. These results have important consequences not only for the viability of SRNs are models of language learning but also for questions of systematicity of generalization in neural networks (Hadley 1994; Marcus 1998). Acquisition of Anaphora by SRNs 2