A “bag-of-arguments” mechanism for initial verb predictions

ABSTRACT

Previous studies have shown that comprehenders use rich contextual information to anticipate upcoming input on the fly, but less is known about how comprehenders integrate different sources of information to generate predictions in real time. The current study examines the time course with which the lexical meaning and structural roles of preverbal arguments impact comprehenders’ lexical semantic predictions about an upcoming verb in two event-related potential (ERP) experiments that use the N400 amplitude as a measure of online predictability. Experiment 1 showed that the N400 was sensitive to predictability when the verb's cloze probability was reduced by substituting one of the arguments (e.g. “The superintendent overheard which tenant/realtor the landlord had evicted … ”), but not when the verb's cloze probability was reduced by simply swapping the roles of the arguments (e.g. “The restaurant owner forgot which customer/waitress the waitress/customer had served … ”). Experiment 2 showed that argument substitution elicited an N400 effect even when the substituted argument appeared elsewhere in the sentence, indicating that verb predictions are specifically driven by the arguments in the same clause as the verb, rather than by a simple “bag-of-words” mechanism. We propose that verb predictions initially rely on a “bag-of-arguments” mechanism, which specifically relies on the lexical meaning, but not the structural roles, of the arguments in a clause.

KEYWORDS:

• Language comprehension
• prediction
• thematic relations
• event-related potentials
• N400

Acknowledgements

We thank Glynis MacMillan, Shefali Shah, and Erin Mahoney for invaluable help in materials creation and data collection.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. We acknowledge that there is a distinction between being an argument of a verb vs. being in the same clause as the verb. However, the present study was not designed to address this distinction. We will return to this in the General Discussion.

2. The space is constructed by creating a table of the log-entropy normed frequencies with which stemmed words occur in a collection of documents, performing singular value decomposition (SVD) on this table, and reducing the dimensionality of the SVD matrices. For a more detailed introduction to LSA, including a discussion of its limitations as a measure of semantic similarity, please see Landauer, Foltz, and Laham (1998).

Additional information

Funding

This work was supported by the National Science Foundation grant to Colin Phillips [grant number BCS-0848554].