Abstract

CCG parsing effort and surprisal jointly predict RT but underpredict garden-path effects

---

Abstract

A prominent approach to explaining sentence processing difficulty is surprisal theory (Hale 2001).  In recent years, surprisal has often been estimated using large language models that do not have explicit representations of syntactic structures, let alone structure-building operations; even so, it predicts word-level difficulty in incremental processing (Oh et al. 2022).  While surprisal theory has been prominent in the study of garden path effects as a model of ambiguity resolution, it has been shown to under-predict the magnitude of such effects in self-paced reading (van Schijndel & Linzen 2021, Arehalli et al. 2022).  On the other hand, models incorporating complexity metrics that take incremental structure-building operations explicitly into account have been shown to improve fit to eye-tracking (Demberg & Keller 2008, Demberg et al. 2013) and neuroimaging data (Brennan et al. 2016, Stanojevi? et al. 2023).

Building on these lines of work, we first ask (Q1) whether a structure building-based complexity metric derived from a CCG (Combinatory Categorial Grammar) parser improves model fit to reading time data beyond surprisal estimates.  We then explore (Q2) the extent to which this complexity metric can predict processing effort related to the recovery of temporally ambiguous sentences.  While our metrics do not straightforwardly predict garden path effects, they predict processing effort in unambiguous sentences.

Keywords: cognitive modeling, computational linguistics, syntactic processing

How to Cite:

Ozaki, S., De Santo, A., Linzen, T. & Dillon, B., (2024) “CCG parsing effort and surprisal jointly predict RT but underpredict garden-path effects”, Society for Computation in Linguistics 7(1), 362–364. doi: https://doi.org/10.7275/scil.2229