Not All Phrases Are Equally Attractive: Experimental Evidence for Selective Agreement Attraction Effects

Abstract

Research on memory retrieval during sentence comprehension suggests that similarity-based interference is mediated by the grammatical function of the distractor. For instance, Van Dyke and McElree (2011) observed interference during retrieval for subject-verb thematic binding when the distractor occurred as an oblique argument inside a prepositional phrase (PP), but not when it occurred as a core argument in direct object position. This contrast motivated the proposal that constituent encodings vary in the distinctiveness of their memory representations based on an argument hierarchy, which makes them differentially susceptible to interference. However, this hypothesis has not been explicitly tested. The present study uses an interference paradigm involving agreement attraction (e.g., Wagers et al., 2009) to test whether the argument status of the distractor determines susceptibility to interference. Results from two self-paced reading experiments show a clear contrast: agreement attraction is observed for oblique arguments (e.g., PP distractors), but attraction is nullified for core arguments (i.e., direct object and subject distractors). A follow-up experiment showed that this contrast cannot be reduced to the syntactic position of the distractor, favoring an account based on the semantic properties of the distractor. These findings support the proposal that interference is mediated by the argument status of the distractor and extend previous results by showing that the effect generalizes to a broader set of syntactic contexts and a wider range of syntactic dependencies. More generally, these results motivate a more nuanced account of real-time agreement processing that depends on both retrieval and encoding mechanisms.

Gender Agreement Attraction in Greek Comprehension

“…Overall, the data cleaning did not affect more than 15% of the data (Ratcliff, 1993) in Experiment 1 (Supplementary Material S3). The regions analyzed were R3 (the region of the Head/pre-critical region), R4 (the region of the Attractor/first critical region), R5 (the region of the agreement target/second critical region), and the post-critical regions, namely R6 and R7 (for potential spillover effects), similar to previous attraction studies (e.g., Wagers et al, 2009;Lago et al, 2015;Slioussar and Malko, 2016;Parker and An, 2018). RTs were log-transformed in all models.…”

Neural Language Models Capture Some, But Not All, Agreement Attraction Effects

“…This is the measure of the attraction effect we aim to simulate. Parker and An (2018) found that this effect was smaller when the attractor appeared in a core argument like in Ex. 8 than when it appeared in an oblique argument as in Ex.…”

“…Core arguments, due to their importance to interpretation, have been argued to be encoded more carefully than their less-critical oblique counterparts (Van Dyke & McElree, 2011). Parker and An (2018) predict that this would lead to stronger attraction effects from attractors in oblique arguments, as the poor encoding may cause the attractor to be more easily confused with the subject. Parker and An (2018) tested this hypothesis in a self-paced reading study using materials such as the following: Figure 6: Human and simulation results for Parker and An (2018).…”

“…Parker and An (2018) predict that this would lead to stronger attraction effects from attractors in oblique arguments, as the poor encoding may cause the attractor to be more easily confused with the subject. Parker and An (2018) tested this hypothesis in a self-paced reading study using materials such as the following: Figure 6: Human and simulation results for Parker and An (2018). Dots represent individual model means; triangles represent averages over 5 models.…”

“…In this paper, we will use LSTM language models to simulate results from six experiments in the agreement attraction literature: (1) Attractors in prepositional phrases generate a stronger attraction effect than those in relative clauses (Bock & Cutting, 1992); (2) Attractors closer to the verb exert a stronger attraction effect, measured in both syntactic (Franck et al, 2002) and 3linear (Haskell & Macdonald, 2005) distance; (4) Collective subjects with distributive readings having higher rates of plural agreement than those with collective readings (Humphreys & Bock, 2005); (5) Attractors in oblique arguments create a larger attraction effect than those in core arguments (Parker & An, 2018); (6) Attractors outside of the clause where agreement is computed cause attraction effects, and attraction effects make ungrammatical sentences seem grammatical, but do not make grammatical sentences seem ungrammatical (Wagers et al, 2009). We find that LSTMs are able to capture the critical human behavior in three of them, indicating that (1) some agreement attraction phenomena can be captured by a simple sequence model without any built-in language-specific mechanisms, but (2) capturing the other phenomena may require models with such mechanisms.…”