The grammaticality asymmetry in agreement attraction reflects response bias: Experimental and modeling evidence

Transitivity and non-uniform subjecthood in agreement attraction

Research on human language converges on a view in which a grammatical "subject" is the most saliently encoded entity in mental representation. However, subjecthood is not a syntactically uniform phenomenon. Notably, many languages encode morphological distinctions between subjects of transitive verbs (i.e., verbs that require an object) and subjects of intransitive verbs. We ask how this typological pattern manifests in a language like English (which does not morphologically signal it) by examining the "distinctiveness" of transitive versus intransitive subjects in memory during online sentence processing. We conducted a self-paced reading experiment that tested for "attraction" effects (Dillon et al., Journal of Memory and Language, 69(2), 85-103, 2013; Wagers et al., Journal of Memory and Language, 61, 206-237, 2009) in the processing of subject-verb number agreement. We find that transitive subjects trigger attraction effects, but that these effects are mitigated for intransitive subject attractors (independently of the number of other noun phrases present in the intervening clause). We interpret this as indicating that transitive subjects are less distinctive and therefore less representationally salient than intransitive subjects: This is because a transitive subject must compete with another clause-mate core argument (i.e., a direct object), which draws on resources from the same pool of memory resources. On the other hand, an intransitive subject minimally only competes with a non-core argument (i.e., an oblique noun phrase); this consumes fewer memory resources, leaving the subject to enjoy greater spoils.

Expectations modulate retrieval interference during ellipsis resolution

Memory operations during language comprehension are subject to interference: retrieval is harder when items are linguistically similar to each other. We test how such interference effects might be modulated by linguistic expectations. Theories differ in how these factors might interact; we consider three possibilities: (i) predictability determines the need for retrieval, (ii) predictability affects cue-preference during retrieval, or (iii) word predictability moderates the effect of noise in memory during retrieval. We first demonstrate that expectations for a target word modulate retrieval interference in Mandarin noun-phrase ellipsis in an electroencephalography (EEG) experiment. This result obtains in globally ungrammatical sentences - termed "facilitatory interference." Such a pattern is inconsistent with theories that focus only on the need for retrieval. To tease apart cue-preferences from noisy-memory representations, we operationalize the latter using a Transformer neural network language model. Confronting the model with our stimuli reveals an interference effect, consistent with prior work, but that effect does not interact with predictability in contrast to human EEG results. Together, these data are most consistent with the hypothesis that the predictability of target items affects cue-preferences during retrieval.

Number Agreement Attraction in Czech Comprehension: Negligible Facilitation Effects

Number agreement attraction in comprehension has been extensively studied in various languages and it has been claimed that attraction effects are generally present across languages. In this paper, four experiments on Czech are presented, each examining a different structure. The Bayesian hierarchical models and Bayes factor analysis pointed towards no agreement attraction effects in three of the experiments. Only in one experiment an effect interpretable as signaling agreement attraction was observed. Its size, however, was so small that it did not translate into a clear preference for models with agreement attraction. The data from the four experiments were further compared to available data from several other languages (English, Armenian, Arabic, and Spanish). The emerging picture is that in Czech, agreement attraction effects are negligible in size if they appear at all. This presents a serious challenge to current theoretical explanations of agreement attraction effects.

Determiner-Number Specification and Non-Local Agreement Computation in L1 and L2 Processing

The present study employed a self-paced reading task in conjunction with concurrent acceptability judgements to examine how similar or different English natives and Chinese learners of English are when processing non-local agreement. We also tested how determiner-number specification modulates number agreement computation in both native and non-native processing by manipulating number marking with demonstrative determiners (the versus that/these). Results suggest both groups were sensitive to non-local agreement violations, indexed by longer reading times for sentences containing number violations. Furthermore, we found determiner-number specification facilitated processing of number violations in both native and non-native groups in an acceptability judgement task only, with stronger sensitivity to violations with demonstrative determiners than those with bare determiners. Contrary to some theories that predict qualitative differences between native and non-native processing, we did not find any significant differences between native and non-native speakers, despite the fact that the Chinese speakers of English had to process a novel linguistic feature absent in their native language.

The Role of Case Syncretism in Agreement Attraction: A Comprehension Study

Many production and comprehension experiments have studied attraction errors in agreement, primarily in number (e.g., “The key to the cabinets were rusty”). Studies on gender agreement attraction are still sparse, especially in comprehension. We present two self-paced reading experiments on Russian focusing on the role of syncretism in this phenomenon. Russian nouns are inflected for case and number, and some forms have the same inflections (are syncretic). In several experiments on Slovak, it was shown that both head and attractor syncretism play a role for gender agreement in production. We demonstrate for the first time that this is also the case in comprehension. The role of head noun syncretism has not been analyzed in any previous comprehension studies, also for number agreement. We conclude that syncretic forms create uncertainty, which is crucial for agreement disruption. These results are better compatible with retrieval approaches to agreement attraction. We discuss the implications of our findings for the nature of the retrieval cues used to establish morphosyntactic dependencies. The question whether case marking modulates agreement attraction in comprehension has also been addressed in a study on Armenian, and it found no evidence of such influence. We offer an explanation of the conflicting findings from several studies based on the syntactic constructions they used as materials.

Semantic Attraction in Sentence Comprehension

Agreement attraction is a cross-linguistic phenomenon where a verb occasionally agrees not with its subject, as required by grammar, but instead with an unrelated noun ("The key to the cabinets were…"). Despite the clear violation of grammatical rules, comprehenders often rate these sentences as acceptable. Contenders for explaining agreement attraction fall into two broad classes: Morphosyntactic accounts specifically designed to explain agreement attraction, and more general sentence processing models, such as the Lewis and Vasishth model, which explain attraction as a consequence of how linguistic structure is stored and accessed in content-addressable memory. In the present research, we disambiguate between these two classes by testing a surprising prediction made by the Lewis and Vasishth model but not by the morphosyntactic accounts, namely, that attraction should not be limited to morphosyntax, but that semantic features of unrelated nouns equally induce attraction. A recent study by Cunnings and Sturt provided initial evidence that this may be the case. Here, we report three single-trial experiments in English that compared semantic and agreement attraction and tested whether and how the two interact. All three experiments showed strong semantically induced attraction effects closely mirroring agreement attraction effects. We complement these results with computational simulations which confirmed that the Lewis and Vasishth model can faithfully reproduce the observed results. In sum, our findings suggest that attraction is a more general phenomenon than is commonly believed, and therefore favor more general sentence processing models, such as the Lewis and Vasishth model.

PIPS: A Parallel Planning Model of Sentence Production

Subject-verb agreement errors are common in sentence production. Many studies have used experimental paradigms targeting the production of subject-verb agreement from a sentence preamble (The key to the cabinets) and eliciting verb errors (… *were shiny). Through reanalysis of previous data (50 experiments; 102,369 observations), we show that this paradigm also results in many errors in preamble repetition, particularly of local noun number (The key to the *cabinet). We explore the mechanisms of both errors in parallelism in producing syntax (PIPS), a model in the Gradient Symbolic Computation framework. PIPS models sentence production using a continuous-state stochastic dynamical system that optimizes grammatical constraints (shaped by previous experience) over vector representations of symbolic structures. At intermediate stages in the computation, grammatical constraints allow multiple competing parses to be partially activated, resulting in stable but transient conjunctive blend states. In the context of the preamble completion task, memory constraints reduce the strength of the target structure, allowing for co-activation of non-target parses where the local noun controls the verb (notional agreement and locally agreeing relative clauses) and non-target parses that include structural constituents with contrasting number specifications (e.g., plural instead of singular local noun). Simulations of the preamble completion task reveal that these partially activated non-target parses, as well the need to balance accurate encoding of lexical and syntactic aspects of the prompt, result in errors. In other words: Because sentence processing is embedded in a processor with finite memory and prior experience with production, interference from non-target production plans causes errors.

The Reading Signatures of Agreement Attraction

The comprehension of subject-verb agreement shows "attraction effects," which reveal that number computations can be derailed by nouns that are grammatically unlicensed to control agreement with a verb. However, previous results are mixed regarding whether attraction affects the processing of grammatical and ungrammatical sentences alike. In a large-sample eye-tracking replication of Lago et al. (2015), we support this "grammaticality asymmetry" by showing that the reading profiles associated with attraction depend on sentence grammaticality. In ungrammatical sentences, attraction affected both fixation durations and regressive eye-movements at the critical disagreeing verb. Meanwhile, both grammatical and ungrammatical sentences showed effects of the attractor noun number prior to the verb, in the first- and second-pass reading of the subject phrase. This contrast suggests that attraction effects in comprehension have at least two different sources: the first reflects verb-triggered processes that operate mainly in ungrammatical sentences. The second source reflects difficulties in the encoding of the subject phrase, which disturb comprehension in both grammatical and ungrammatical sentences.

A transposed-word effect across space and time: Evidence from Chinese

A compelling account of the reading process holds that words must be encoded serially, and so recognized strictly one at a time in the order they are encountered. However, this view has been challenged recently, based on evidence showing that readers sometimes fail to notice when adjacent words appear in ungrammatical order. This is argued to show that words are actually encoded in parallel, so that multiple words are processed simultaneously and therefore might be recognized out of order. We tested this account in an experiment in Chinese with 112 skilled readers, employing methods used previously to demonstrate flexible word order processing, and display techniques that allowed or disallowed the parallel encoding of words. The results provided evidence for flexible word order processing even when words must be encoded serially. Accordingly, while word order can be processed flexibly during reading, this need not entail that words are encoded in parallel.

Modeling Misretrieval and Feature Substitution in Agreement Attraction: A Computational Evaluation

We present computational modeling results based on a self-paced reading study investigating number attraction effects in Eastern Armenian. We implement three novel computational models of agreement attraction in a Bayesian framework and compare their predictive fit to the data using k-fold cross-validation. We find that our data are better accounted for by an encoding-based model of agreement attraction, compared to a retrieval-based model. A novel methodological contribution of our study is the use of comprehension questions with open-ended responses, so that both misinterpretation of the number feature of the subject phrase and misassignment of the thematic subject role of the verb can be investigated at the same time. We find evidence for both types of misinterpretation in our study, sometimes in the same trial. However, the specific error patterns in our data are not fully consistent with any previously proposed model.

Using eye tracking to investigate failure to notice word transpositions in reading

Previous research (Mirault, Snell, & Grainger, 2018) has demonstrated that subjects sometimes incorrectly judge an ungrammatical sentence as grammatical when it is created by the transposition of two words in a grammatical sentence (e.g., The white was cat big). Here we present two eye-tracking experiments designed to assess the prevalence of this phenomenon in a more natural reading task, and to explore theoretical explanations. Readers failed to notice transpositions at about the same rate as in Mirault et al. (2018). Failure to notice the transposition was more common when both words were short, and when readers' eyes skipped, rather than directly fixated, one of the two words. The status of the transposed words as open- or closed-class did not have a reliable effect. The transposed words caused disruption in the eye movement record only on trials when participants ultimately judged the sentence to be ungrammatical, not when they judged the sentence to be grammatical. We argue that the results are not entirely consistent with the account offered by Mirault et al. (2018), which attributes failure to notice transpositions to parallel processing of adjacent words, or with a late, post-perceptual rational inference account (Gibson, Bergen, & Piantadosi, 2013). We propose that word recognition is serial, but post-lexical integration of each word into its context may not be perfectly incremental.

Attraction Effects for Verbal Gender and Number Are Similar but Not Identical: Self-Paced Reading Evidence From Modern Standard Arabic

Previous work on the comprehension of agreement has shown that incorrectly inflected verbs do not trigger responses typically seen with fully ungrammatical verbs when the preceding sentential context furnishes a possibly matching distractor noun (i.e., agreement attraction). We report eight studies, three being direct replications, designed to assess the degree of similarity of these errors in the comprehension of subject-verb agreement along the dimensions of grammatical gender and number in Modern Standard Arabic. A meta-analysis of the results demonstrate the presence of agreement attraction effects in reading comprehension for gender and number on verbs. Moreover, the meta-analysis demonstrates that these two features do not behave identically: gender effects are larger and occur later relative to number attraction effects. These results challenge models of agreement that predict agreement features to be equipotent and show that real-time models of agreement require modifications in the form of cue-weighting in order to account for these differential results.

Gender Agreement Attraction in Greek Comprehension

This work explores gender agreement attraction in comprehension. Attraction occurs when an agreement error (such as, "the key to the cabinets are rusty") goes unnoticed, leading to the illusion of grammaticality due to a mismatch between the value of the head and the value of a local intervening phase (attractor). According to retrieval accounts, these errors occur during cue retrieval from memory and predict illusions of grammaticality. Alternatively, representational accounts predict that the errors occur due to the faulty representation of certain features, thus, illusions of ungrammaticality are also expected. In four experiments we explore: (a) whether gender agreement attraction occurs in Greek and the strategy/-ies employed, (b) the role of the agreement target, (c) the timing of gender agreement attraction, (d) the role of phonological matching between the nominal inflectional morphemes of the attractor and the agreement target, and (e) participants' sensitivity to agreement when there is no conflict from the attractor. In all four experiments, the grammaticality of the sentence and the attractor value (match or mismatch with the head) and also the phonological matching between the attractor and the agreement target in ungrammatical sentences were manipulated. The agreement target was either an adjectival predicate or an object-clitic and the gender value of the head was feminine or neuter. Attraction was found in all measures during the time-course of adjectival predicates (Experiment 1) and object-clitics (Experiment 2), and in timed (Experiment 3), and untimed (Experiment 4) judgments. Even more, both gender values showed attraction and the results mainly suggest that participants experience illusions of grammaticality, confirming retrieval accounts. Phonological matching did not modulate attraction in any of the experiments, suggesting that the similarity in the morphophonological realization between the agreement target and the attractor does not increase attraction. Furthermore, participants were sensitive to gender agreement violations in the absence of gender mismatch between the head and the attractor, suggesting that they respect agreement rules and have both neuter and feminine available in their feature content repertoire, although with some tendency in favor of neuter in feminine agreement contexts. The impact of these findings is discussed within the concept of attraction and sensitivity to agreement violations.

Error-Driven Retrieval in Agreement Attraction Rarely Leads to Misinterpretation

Previous work on agreement computation in sentence comprehension motivates a model in which the parser predicts the verb's number and engages in retrieval of the agreement controller only when it detects a mismatch between the prediction and the bottom-up input. It is the error-driven second stage of this process that is prone to similarity-based interference and can result in the illusory licensing of a subject-verb number agreement violation in the presence of a structurally irrelevant noun matching the number marking on the verb ('The bed by the lamps were…'), giving rise to an effect known as 'agreement attraction'. Here we ask to what extent the error-driven retrieval process underlying the illusory licensing alters the structural and thematic representation of the sentence. We use a novel dual-task paradigm that combines self-paced reading with a speeded forced choice task to investigate whether agreement attraction leads comprehenders to erroneously interpret the attractor as the thematic subject, which would indicate structural reanalysis. Participants read sentence fragments ('The bed by the lamp/lamps was/were undoubtedly quite') and completed the sentences by choosing between two adjectives ('comfortable'/'bright') which were either compatible with the subject's head noun or with the attractor. We found the expected agreement attraction profile in the self-paced reading data but the interpretive error occurs on only a small subset of attraction trials, suggesting that in agreement attraction agreement checking rarely matches the thematic relation. We propose that illusory licensing of an agreement violation often reflects a low-level rechecking process that is only concerned with number and does not have an impact on the structural representation of the sentence. Interestingly, this suggests that error-driven repair processes can result in a globally inconsistent final sentence representation with a persistent mismatch between the subject and the verb.