Bott, O. 1 , Koornneef, A. 2 & Solstad, T. 3
1 University of Tuebingen
2 Leiden University
3 Leibniz-Centre General Linguistics (ZAS), Berlin
Implicit Causality (IC) verbs constitute a central topic in research on prediction in natural language processing. Selecting for two animate arguments, IC verbs display a strong preference for an explanation focusing on one argument, as shown in numerous sentence continuation experiments:
//
(1) JOHN annoyed Mary because ? HE was noisy.
(2) John admired MARY because ? SHE was clever.
//
The predictive nature of IC verbs is still insufficiently understood, however. Key questions include (i) what is predicted: a word/pronoun (e.g. HE/SHE in 1/2) or a type of explanation (e.g. a property of John's in 1, and of Mary's in 2), (ii) what triggers the prediction: lexical semantics (annoy/admire in 1/2) or world knowledge, and (iii) what the processing profile (timing of the connective and the pronoun measured within different experimental paradigms) of the prediction should be.
//
Based on a formal-semantic theory of IC (Bott/Solstad:2014), results from previous experimental research (e.g., Koornneef/van Berkum:2006;Featherstone/Sturt:2011) and recent insights into the nature of predictive processing in general (e.g. Kuperberg/Jaeger:2016), we propose a comprehensive framework for predictive processing of IC. Crucially, we consider in detail the relation between what is predicted (the predictee; cf. Kamide:2008) and the properties of linguistic expressions such as pronouns (see 1/2 above) that may be taken to (in)validate predictions (before integration is completed). Rooted in previous experimental research on IC, we evaluate the range of top-down and bottom-up processes: Which linguistic levels are involved and how do they interact? Our framework further allows for assessing general and paradigm-specific processing profiles in terms of facilitation/violation costs. Last, but not least, we argue that focusing on the relation between predictees and what we will term (in)validators also for other phenomena and experimental manipulations may contribute towards a better understanding and more precise models of language-based prediction in general.