Supporting materials for Arnold (2024) “What categories matter for tracking discourse statistics?”

What categories matter for tracking discourse statistics?

Jennifer E. Arnold

funded by NSF #1917840

data and SAS code available at: https://osf.io/f2vta/?view_only=f05b694ec3f64e7885612b5c8da3c685

Manuscript under review; can be cited as:

Arnold, J. E. (under review). What categories matter for tracking discourse statistics? Ms., UNC Chapel Hill.

Related conference presentations:

Arnold, J. E. (2024). Learning discourse patterns through exposure: Mixed input helps identify informative categories. Poster (expected); Experimental approaches to Linguistic Meaning, June 2024, U. Pennsylvania.

Arnold, J. E. (2024). Learning discourse patterns through exposure: Mixed input helps identify informative categories. Poster (expected); Human Sentence Processing Conference, May 2024, U. Michigan.

Abstract

The language processing system includes knowledge about which patterns are more frequent. People even track reference relations, for example the fact that subjects are more likely to be re-mentioned than other entities in a sentence, such that “Ana ate with Will and she…” is more frequent than “Ana ate with Will and he…”, or that semantic categories like goals are more likely to be re-mentioned than sources (Langlois & Arnold, 2020; Johnson & Arnold, 2023). But a puzzle stems from the fact that each input contains multiple possible representations that might be tracked. How do people know which ones to encode? In goal-source transfer verbs (e.g. “Ana got a book from Will and he…”), the subject is always a goal. Do people encode this example as an instance of subject reference or goal reference? Using an adaptation paradigm, we test which patterns influence subsequent pronoun interpretation. In Experiment 1, where the exposure items only use goal-source transfer verbs, people showed an adaptation effect only for test items with goal-source and not source-goal test items. In experiment 2, where 75% of the exposure items used goal-source transfer verbs and 25% used agent-comitative verbs, people showed adaptation for both goal-source and source-goal verbs. We conclude that experiment 1 supported learning about both thematic and syntactic roles, which canceled each other out for the mismatching verbs. In experiment 2, variability in thematic roles signaled that the informative property about antecedent identity was syntactic role. Results show that referential relations are encoded and pooled across multiple inputs.