Abstract

Transient blend states and discrete agreement-driven errors in sentence production

---

Abstract

Errors in subject-verb agreement are common in everyday language production. This has been studied using a preamble completion task in which a participant hears or reads a preamble containing inflected nouns and forms a complete English sentence (“The key to the cabinets” could be completed as "The key to the cabinets is gold.") Existing work has focused on errors arising in selecting the correct verb form for production in the presence of a more ‘local’ noun with different number features (The key to the cabinets are gold). However, the same paradigm elicits substantial numbers of preamble errors ("The key to the cabinets" repeated as "The key to the cabinet") that existing theories have largely failed to address.

\n

We propose a Gradient Symbolic Computation (GSC) account of agreement and preamble errors. Sentence processing is modeled as a continuous-time, continuous-state stochastic dynamical system. Within this continuous representational space, a subset of states reflect discrete symbolic structures. The remainder are blend states where multiple symbols are simultaneously partially active. Initial phases of computation prefer blend states; an additional dynamic control parameter, commitment strength, pushes the model to discrete structures. This process, combined with stochastic gradient ascent dynamics respecting grammatical constraints on syntactic structures, yields discrete sentence outputs. We propose that transient blend states allow portions of target and non-target syntactic structures to interact, yielding both verb and preamble errors.

Keywords: sentence production, dynamical systems. neural networks, Gradient Symbolic Computation

How to Cite:

Goldrick, M., Brehm, L., Cho, P. & Smolensky, P., (2019) “Transient blend states and discrete agreement-driven errors in sentence production”, Society for Computation in Linguistics 2(1), 375-376. doi: https://doi.org/10.7275/n0b2-5305