Chang, F. (2002). Symbolically speaking: A connectionist model of sentence production. Cognitive Science, 26(5), 609–651 pdf.
The ability to combine words into novel sentences has been used to argue that humans have symbolic language production abilities. Critiques of connectionist models of language often center on the inability of these models to generalize symbolically (Fodor & Pylyshyn, 1988; Marcus, 1998). To address these issues, a connectionist model of sentence production was developed. The model had variables (role-concept bindings) that were inspired by spatial representations ( Landau & Jackendoff, 1993). In order to take advantage of these variables, a novel dual-pathway architecture with event semantics is proposed and shown to be better at symbolic generalization than several variants. This architecture has one pathway for mapping message content to words and a separate pathway that enforces sequencing constraints. Analysis of the model’s hidden units demonstrated that the model learned different types of information in each pathway, and that the model’s compositional behavior arose from the combination of these two pathways. The model’s ability to balance symbolic and statistical behavior in syntax acquisition and to model aphasic double dissociations provided independent support for the dual-pathway architecture.
Erratum:
The actual grammar used to generate sentences that the models were trained on deviated slightly from that reported in the article. Specifically, on benefactive datves with non-light verbs, the grammar generated sentences always without goal arguments. The following sections of the article are impacted:
pg. 612 in Table 1; pg 647, Appendix A: change "bake" to "make" in example sentences.
pg. 613: change "(e.g. make, bake)" to "(e.g. make)".
pg. 617;pg 621: remove "for the cafe" in example sentences.
pg. 638. change "only in transitive and benefactive dative frames" to "only in transitive frames."
pg. 628-629. About half of the dog-goal test items used the benefactive datives structures with DOG in the goal role. Because this meaning was associated with a two argument transitive structure (without the phrase that included the word "dog"), this test probably overestimated the ability of the models to produce this novel prepositional phrases with the word "dog". Another test set (the overt-dog-goal test set) was created made up of sentences with DOG in the goal slot, and a structure where the word "dog" was overtly produced. When tested in same manner as the original dog-goal test, the results are on the whole similar (Prod-SRN 0%, No-event-semantics model 18%, Linked-path model 36%, Dual-path model 60%). Model type is significant [F(3,9) = 12.0, p < 0.002]. Pairwise comparisons demonstrate that the Dual-path model is superior to the Prod-SRN and No-event-semantics models [Fs(1,9) > 16.3, ps < 0.003] and marginally superior to the Linked-path model [F(1,9) = 5.1, p < 0.06]. These results, in concert with the original dog-goal results, suggest that the Dual-path model is better at placing words into novel sentence positions.
None of the substantive conclusions of the study are affected by these changes.