Sentences unfold over time in both production and comprehension. The apparent ease of forming syntactic and semantic dependencies between words that are not temporally adjacent suggests that language users have a fairly efficient working memory where partially complete linguistic representations can be stored and updated in real-time (Lewis et al., 2006, McElree, 2006, Wagers and McElree, 2013). In recent years, verbal agreement has offered an important model case for understanding how language users make use of such a working memory to encode syntactic dependencies (Badecker and Kuminiak, 2007, Bhatia and Dillon, 2022, Dempsey et al., 2022, Franck and Wagers, 2020, Lago et al., 2021, Lakretz et al., 2021, Schlueter et al., 2018, Villata et al., 2018, Wagers et al., 2009). Since verbal agreement is a canonically structure-dependent relationship between the morphological features of one constituent in the sentence (canonically, the subject) and the features on the verb (Adger and Harbour, 2007, Bjorkman and Zeijlstra, 2019, Chomsky, 1995, Chomsky, 2000, Chomsky, 2001, Pesetsky and Torrego, 2007, Preminger, 2014, a.m.o.), it has provided a useful probe into how language users can track hierarchically structured linguistic dependencies in real time.

Language users show characteristic errors in computing agreement dependencies: Many languages show number agreement attraction errors in the production of subject-verb agreement. Roughly speaking, agreement attraction occurs when the verb erroneously agrees with a distractor noun instead of agreeing with the syntactically appropriate head noun, as in (1):

The key to the cabinets are on the table. (Bock & Miller, 1991).

The distribution and severity of these errors have yielded important insights into the memory mechanisms used to encode and track syntactic dependencies in comprehension. It is tempting to attribute the error in (1) to a ‘proximity concord’ error (Francis, 1986, Quirk et al., 1985), since ‘the cabinets are’ forms a locally coherent substring in (1) (Tabor, Galantucci, & Richardson, 2004). However, despite the appeal of this idea, there is ample evidence that agreement attraction reflects a somewhat more abstract encoding or retrieval error. For example, plural distractors embedded in a prepositional phrase modifier lead to more attraction than those inside a relative clause although the distance between the verb and the local noun is the same in both contexts (Bock and Cutting, 1992, Solomon and Pearlmutter, 2004), non-local distractors can cause agreement attraction (Bock and Miller, 1991, Franck and Wagers, 2020, Wagers et al., 2009), and distractors that have more features that correlate with subjecthood create more attraction than those that do not (Badecker and Kuminiak, 2007, Slioussar, 2018, Thornton and MacDonald, 2003; a.o.—but see Avetisyan et al., 2020).

The present paper investigates the claim that intervenors with more subject-like features create higher error rates. Specifically, we ask whether number agreement attraction can be modulated by language-specific distributional properties that correlate with subjecthood. To do this, we investigate agreement attraction in Romanian, a Romance language spoken by more than 20 million native speakers worldwide. As we detail below, language-specific distributional constraints make bare nouns less likely subjects in Romanian. In four speeded forced choice sentence continuation experiments, we test for the presence of agreement attraction effects in Romanian with different types of distractors. We find that bare noun distractors cause less agreement attraction than other types of distractors in Romanian. On balance, we argue that this supports the predictions of a cue-based retrieval account of agreement attraction errors, according to which agreement attraction reflects an error that arises when a verb initiates a cue-based retrieval for a subject in working memory and is more likely to activate a full determiner phrase (DP) than a bare noun in Romanian.

Our paper is structured as follows: after a brief introduction, Section “Agreement attraction and similarity-based interference” presents the phenomenon of agreement attraction in general. Section “Background on Romanian” presents background on some of the properties specific to Romanian which are relevant for agreement attraction. In Sections “Experiments 1–3” and “Experiment 4 (Within-subjects replication)”, we present four experiments using the speeded forced choice sentence continuation task to examine how much agreement attraction different types of distractors cause in Romanian. Section “General discussion” proposes an explanation in terms of cue-based retrieval, while Section “Conclusion” draws the final conclusions of our investigation.