Experience and Generalization in a Connectionist Model of Mandarin Chinese Relative Clause Processing

Sentences containing relative clauses are well known to be difficult to comprehend, and they have long been an arena in which to investigate the role of working memory in language comprehension. However, recent work has suggested that relative clause processing is better described by ambiguity resolution processes than by limits on extrinsic working memory. We investigated these alternative views with a Simple Recurrent Network (SRN) model of relative clause processing in Mandarin Chinese, which has a unique pattern of word order across main and relative clauses and which has yielded mixed results in human comprehension studies. To assess the model’s ability to generalize from similar sentence structures, and to observe effects of ambiguity through the sentence, we trained the model on several different sentence types, based on a detailed corpus analysis of Mandarin relative clauses and simple sentences, coded to include patterns of noun animacy in the various structures. The model was evaluated on16 different relative clause subtypes. Its performance corresponded well to human reading times, including effects previously attributed to working memory overflow. The model’s performance across a wide variety of sentence types suggested that the seemingly inconsistent results in some prior empirical studies stemmed from failures to consider the full range of sentence types in empirical studies. Crucially, sentence difficulty for the model was not simply a reflection of sentence frequency in the training set; the model generalized from similar sentences and showed high error rates at points of ambiguity. The results suggest that SRNs are a powerful tool to examine the complicated constraint-satisfaction process of sentence comprehension, and that understanding comprehension of specific structures must include consideration of experiences with other similar structures in the language