The Unification Space implemented as a localist neural net: predictions and error-tolerance in a constraint-based parser

Abstract

We introduce a novel computer implementation of the Unification-Space parser (Vosse and Kempen in Cognition 75:105-143, 2000) in the form of a localist neural network whose dynamics is based on interactive activation and inhibition. The wiring of the network is determined by Performance Grammar (Kempen and Harbusch in Verb constructions in German and Dutch. Benjamins, Amsterdam, 2003), a lexicalist formalism with feature unification as binding operation. While the network is processing input word strings incrementally, the evolving shape of parse trees is represented in the form of changing patterns of activation in nodes that code for syntactic properties of words and phrases, and for the grammatical functions they fulfill. The system is capable, at least qualitatively and rudimentarily, of simulating several important dynamic aspects of human syntactic parsing, including garden-path phenomena and reanalysis, effects of complexity (various types of clause embeddings), faulttolerance in case of unification failures and unknown words, and predictive parsing (expectation-based analysis, surprisal effects). English is the target language of the parser described.

Syntactic priming during sentence comprehension: Evidence for the lexical boost.

“…The results do not cleanly discriminate between an account under which each lexical item has its own pool of syntactic structure representations and an account under which the ‘lexical boost’ is driven by associations between lexical entries and a pool of shared syntactic structure representations. However, the results are compatible with syntactic structure-building processes that are influenced by the lexical properties of critical words (e.g., MacDonald et al, 1994; Vosse & Kempen, 2000, 2009). …”

“…In fact, lexicalized syntax accounts such as unification grammar (Vosse & Kempen, 2000, 2009) view parsing as a process of selecting lexical representations that have the appropriate “tags” that allow them to combine with other words in a sentence. In the case of the reduced relative, the initial lexical choice for the subject noun would normally be the version that has the noun at its core and an “S” node at the “top”.…”

“…See Jackendoff, 2003; Vosse & Kempen, 2000, 2009, for more complete, linguistically motivated descriptions of “vertically integrated” representations. …”

Effective connectivity of cortical and subcortical regions during unification of sentence structure

“…The lexical-syntactic information is most likely not 'copied' from the area necessary for its retrieval (LpMTG) to the area necessary for unification (LIFG). Instead, the sustained activation of lexical-syntactic information could be triggered by feedback from the LIFG to the LpMTG (see Vosse and Kempen (2009) for a computational implementation hereof). The amount and/or duration of lexicalsyntactic activation is a function of the unification load imposed by the combinatorial operations necessary for unification.…”

Action and Language Mechanisms in the Brain: Data, Models and Neuroinformatics

“…In particular, most current models assign some role to competition between items (MacWhinney 1987; Vosse and Kempen 2000; Vosse and Kempen 2009) on the various linguistic levels (but see Clifton and Staub 2008 for explicit empirical counterarguments). All current models assume that processing occurs incrementally, trying to make role assignments and attachments as soon as possible (Kempen and Hoenkamp 1987) based on the relative strength of surface cues to interpretation (MacWhinney 2011).…”