The time course of the subject advantage in Mandarin: an experimental and computational investigation
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Many languages show a “subject advantage” in sentence processing, where relative clauses (RCs) with subject dependencies are processed more easily than those with object dependencies. This observation has led to theories about the mechanisms involved in processing long-distance dependencies including Dependency Locality Theory (Gibson, 1998), the hypotheses that the subject advantage is cross-linguistically universal (e.g., O’Grady et al., 2003), as well as the probabilistic expectation-based comprehension and surprisal (e.g., Hale, 2001; Levy 2008). ☐ In this dissertation, I re-examine RCs in Mandarin, a language for which findings are mixed: some prior studies report an SRC advantage, and others an ORC advantage. Two original experiments which use nominal classifiers to provide a syntactic cue during online parsing were presented, which allows disambiguating a sentence-initial RC from a matrix clause. Based on the results, I find that Mandarin has a temporary ORC advantage (at the point of interpreting the filler into the structural representation), but that this then gives way to an SRC advantage. The former is governed by dependency locality (Gibson, 1998), and the latter is governed by the “universality” of SRC advantage. To explore the role of expectation-based theories in processing linguistic dependencies, I adopt a computational approach: employing one transformer-based large language model (LLM) - GPT2-Chinese to analyze the effect of surprisal. This allows us to assess how well the surprisal explains reading time for relative clause processing in Mandarin. The results of surprisal analysis reveal that the language model doesn’t capture syntactic and semantic relationships in a way that mirrors human processing. ☐ I propose a two-step model for processing dependencies in RCs. The first step is integration of the filler with the gap site, and I maintain that whether a language shows an SRC advantage or an ORC advantage for this step is governed by dependency locality. Following integration, the second step requires establishing a thematic interpretation for the now-integrated wh pronoun (either a null wh operator or an overt wh pronoun like “who”). As in other languages, Mandarin has an SRC advantage at this step, and I hypothesize that this component of processing is universally subject-biased. The expectation-driven parsing mechanisms doesn’t fit into our two-step model for Mandarin RC processing.
Description
Keywords
Classifiers, Filler-gap dependencies, Mandarin, Relative clauses, Subject advantage