To extract meaning from novel sentences, comprehenders need to form associations between words on the fly. These associations are governed by grammatical rules. For example, in the sentence: The professor always forgets himself, both the verb forgets and the pronoun himself have to match some features of the subject (number and/or gender). Comprehenders quickly and accurately apply such complex grammatical rules in most cases. Yet, fallibilities have been identified (Wagers et al., 2009, Parker and Phillips, 2017, for a review see Jäger, Engelmann, & Vasishth, 2017). These fallibilities are informative with regard to the memory architecture that mediates linguistic dependency formation.

One outstanding question in this area is whether memory representations are fixed. Some memory models assume that items are encoded accurately and kept intact during processing. In these models, interference arises only during access to memory representations. Others models trace fallibilities to inaccuracy of the memory representations themselves. These models assume that the representation of items in memory can be distorted during processing. In the current study, we examine the interaction between multiple related dependencies (Molinaro et al., 2008), to target the question of shifts in memory representation. We investigate how the formation of one dependency affects memory representations that are later used in subsequent dependencies. We focus on the interaction between subject-verb agreement and antecedent-reflexive dependencies, as a test case. This can offer novel insights into the memory architecture underlying sentence processing.

The hallmark of memory fallibility in linguistic dependency formation is agreement attraction. A structurally irrelevant element, a so-called distractor, can trigger verbal agreement errors. In these cases, agreement is formed with the distractor rather than with the subject. Such errors arise in both production (e.g. Bock and Miller, 1991, Eberhard et al., 2005, Kandel et al., 2022) and comprehension (Pearlmutter et al., 1999, Wagers et al., 2009, Hammerly et al., 2019). These are termed agreement attraction errors.

For example, speakers produce ungrammatical plural verbs (‘are’) more often in configurations like (1b), which include a plural distractor (‘cabinets’), relative to configurations with a singular distractor (1a). Similarly, in comprehension, an ungrammatical verb is sometimes perceived as well-formed in the presence of a distractor matching its features, as in (1b) (“grammaticality illusion”). In those environments, ungrammatical verbs are also read faster (1b relative to 1a) (Pearlmutter et al., 1999, Wagers et al., 2009).

(1) a. The key to the cabinet are rusty.

b. The key to the cabinets are rusty.

Attraction effects in comprehension are ubiquitous; they have been observed in different languages (e.g. Slovak: Badecker & Kuminiak, 2007; Spanish: Lago, Shalom, Sigman, Lau, & Phillips, 2015; Arabic: Tucker et al., 2015, Tucker et al., 2021; Russian: Slioussar, 2018; Turkish: Lago, Gračanin-Yuksek, Şafak, Demir, Kırkıcı, & Felser, 2019); and are found in various experimental measures (e.g. Pearlmutter et al., 1999, Staub, 2009, Wagers et al., 2009, Dillon et al., 2013, Tanner et al., 2014).

Broadly speaking, attraction effects in comprehension are thought to reflect memory limitations. The most prominent account has been couched within the Cue-Based Retrieval (CBR) model (Lewis & Vasishth, 2005). In this model, attraction arises during the retrieval process initiated at the verb (Wagers et al., 2009; see Jäger, Engelmann, & Vasishth, 2017 for a review). This process attempts to access the subject using cues which are derived from the verb, including agreement features and structural features (i.e. a cue specifying that the target needs to be in subject position). These cues include agreement features and structural features (i.e. a cue specifying that the target needs to be in a subject position). The cues activate memory items which match them, and activated candidates race for retrieval (Vasishth, Nicenboim, Engelmann, & Burchert, 2019). When verbal agreement does not match the subject, but rather the distractor (as in 1b), the result is multiple partial matches: the subject (‘key’) matches the structural cue, and the distractor (‘cabinets’) matches the agreement cue (“plural”). A race involving two partial matches is more balanced, and therefore faster to terminate on average, than a race between a partial match and a non-matching item (“statistical facilitation”, Raab, 1962). Thus, a speed-up in reading times arises in sentences like (1b) relative to ones like (1a). In addition, the distractor will occasionally be retrieved instead of the subject in these partial match configurations.

Other accounts of attraction focus on representational distortion. In these models, attraction reflects an error in the memory representation of the subject’s features, rather than an error in accessing the subject head. In particular, representational accounts, such as Feature Percolation or Marking & Morphing. suggest that the plurality of the distractor disrupts the singular feature of the subject’s representation (Bock and Eberhard, 1993, Franck et al., 2002, Eberhard et al., 2005). Thus, in (1b), the subject could be erroneously represented as ‘keys’ rather than ‘key’. This leads to compatibility with a plural verb.

Representational models of agreement attraction have been originally designed to account for production errors, and have received relatively little attention in the comprehension literature (but see Hammerly, Staub, & Dillon, 2019). However, they could naturally be extended to comprehension. Such an extension would be based on incremental memory encoding. For example, in (1), the subject head would be parsed first. At this point, the subject would accurately be encoded as singular. Then the distractor (cabinets) would be encoded to memory. The plurality of the distractor could spread to the memory representation of the subject and distort its number marking. This process would yield equivocal or incorrect (plural) number marking for the subject phrase. Thus, grammaticality illusions may arise at a subsequent ungrammatical verb. Such a mechanism requires that memory representations are not fixed - they are affected by subsequent material in the sentence. This is a crucial point of contention, as CBR assumes fixed memory representations.

A different line of research provides evidence that the memory representation of the subject is not fixed, and can shift based on verbal agreement. In an ERP study, Molinaro and colleagues (2008) found that reflexives that followed a subject-verb mismatch (as in 3) elicited a P600 effect (usually associated with detection of ungrammaticality) when they mismatched the verb (‘herself’ in 3). Reflexives which mismatched the subject (‘themselves’), did not elicit this effect. Molinaro and colleagues (2008) proposed that comprehenders coerced the features of the subject to match those of the verb. Such a process, whereby features of a previous constituent are edited in memory, is at odds with the assumption of the CBR model that representations are fixed. The preference for a match between the reflexive and the verb calls for a model in which representations can be edited or distorted during processing.

(3) The famous dancer were nervously preparing to face the crowd.

It is unclear whether the findings can be directly explained in terms of Feature Percolation or a Marking & Morphing. Features could in principle “migrate” from the verb to the subject. Representational interference models often license effects of different sources of morphosyntactic marking (Eberhard, 1997, Eberhard et al., 2005, Bleotu and Dillon, 2022) and should in principle allow the verb to act as a “distractor”. However, effects of verbal agreement on the subject’s representation have not been discussed in this literature, to the best of our knowledge. In these models a crucial determinant of the ability to obscure the representation of the subject is the constituent’s distance from the subject. In the absence of previous literature in this area it is not clear whether a verb should be considered close enough to drastically affect the subject’s representation.

Here we propose an extension of representational distortion accounts - the agreement updating hypothesis. We build on the idea that distractors can create distorted representations, and add a rational updating function that can counteract this interference. We suggest that editing of memory representations, as observed in Molinero’s study, is not an accidental distortion/interference (as in Feature Percolation or Marking and Morphing). Instead, we propose that such edits reflect rational inference about likely features of the subject.

Rational inference approaches suggest that comprehenders consider the possibility that the speaker/writer’s message was corrupted by production and perception errors. Thus, comprehenders consider non-veridical interpretations. Comprehenders may ultimately adopt these interpretations based on (Bayesian) inference about likely errors and probable messages (e.g. Levy, 2008a, Levy, 2008, Gibson et al., 2013, Keshev and Meltzer-Asscher, 2021). Such an inference can also consider possible memory errors (Futrell, Gibson, & Levy, 2020).

Inference that adjusts representations for memory errors allows updating an item’s features based on information in the sentence. When reading the verb in (3), comprehenders may conclude that the subject’s form in the input or its representation in memory was distorted. If comprehenders consider production errors, they may reinterpret the subject in light of the verb. A reader is likely to assume that the subject was erroneously produced as singular since this involves a very plausible error - an omission of the plural morpheme ‘s’. The other way to restore grammaticality would be positing that the writer substituted ‘was’ with ‘were'. This type of distortion would involve more letter edits and thus would make a less likely typing error. If comprehenders maintain uncertainty about their own memory representations, they are again more likely to question the features of the subject, which are no longer perceptually available. Thus, either way the subject’s representation would be modified to agree with the verb. Editing features of the subject at the verb entails that a subsequent reflexive dependency would use the new features (derived from the verb) rather than those of the veridical subject. This straightforwardly accounts for the findings of Molinaro et al. (2008): A reflexive anaphor that matches the verb would fit in better than one that matches the veridical subject.

Note that under this approach, features of the subject are not distorted due to accidental feature migration. Instead they are updated based on grammatical knowledge about subject-verb agreement. For this reason, the priors of rational inference dictates that the verb will affect the representation of the subject but not those of the distractor.1 In addition, updating should change the number feature of the subject but not lead to an interpretation where the distractor is the subject since agreement errors are likely (involve few edits and arise in native and non-native speech). Interestingly, this mechanism can force grammatical relations after memory interference distorted representation, when combined with a representational interference model. Therefore, we propose that agreement updating can reduce vulnerability to interference coming from other elements in the sentence.

A shallow processing strategy might be a possible explanation of Molinaro’s findings. In Good Enough Processing approaches (Ferreira et al., 2002, Ferreira and Patson, 2007), comprehenders do not necessarily build a complete, globally grammatical, representation of the input. Thus, in (3) comprehenders might rely on shallow local verb-reflexive match to determine the reflexive’s acceptability. Patson and Husband (2016) similarly proposed that agreement attraction reflects the effect of a local distractor on the verb within a processing stream which is not committed to syntactic structure. However, attraction is readily observed also in configurations where the distractor precedes the target (Wagers et al., 2009, Staub, 2010, Tucker et al., 2015, Tucker et al., 2021, Parker and Phillips, 2017; a.o.). This suggests that linear order is not a major contributor to agreement dependencies. Therefore, we consider shallow processing less likely as an account of Molinaro’s findings.

The current study investigates whether memory representations are fixed or whether they can be updated, pitting CBR against the agreement updating hypothesis. To do so, we move the topic of investigation into the heart of the memory interference arena - attraction effects. We thus examine the interaction between reflexive attraction and verbal agreement in Hebrew.

Attraction has been observed in reflexive anaphors (Parker and Phillips, 2017, Sloggett, 2017, Jäger et al., 2020), though reflexives may be less susceptible to attraction than verbs (Dillon et al., 2013, Sturt, 2003, Cunnings and Sturt, 2014, Parker and Phillips, 2017). To date, reflexive attraction studies have examined the reflexive-antecedent dependency in isolation, when the preceding verb does not bear the relevant agreement features. This was achieved either by relying on the anaphor’s gender features, which are not reflected in English verbal agreement (Sturt, 2003, Cunnings and Felser, 2013, Cunnings and Sturt, 2014, Parker and Phillips, 2017, Sloggett, 2017) or by using verb forms which do not carry number information, such as the English past tense (Dillon et al., 2013, Jäger et al., 2020).

Hebrew’s rich agreement system manifests gender marking on both verbs and reflexive anaphors. This offers an ideal environment for testing how the two dependencies interact. We test the vulnerability of an ungrammatical reflexive anaphor (mismatching the subject’s gender) to the gender of a structurally irrelevant distractor (matching or mismatching the subject). We test this in three different cases: (i) following a verb that does not carry agreement information (as in most previous studies of reflexive attraction); (ii) when the verb carries the grammatical gender feature, matching the subject; and (iii) when the verb manifests an ungrammatical gender feature, mismatching the subject. Fig. 1 illustrates the full design. In the following subsection we delineate predictions of the alternative approaches, focusing on prediction of reflexive attraction contrasts.

In CBR, recent retrieval of an item increases its activation levels (Vasishth & Lewis, 2006). Higher activation, in turn, can facilitate access to this item later on. Therefore, accessing the distractor at the verb should make it more accessible at later retrieval sites (e.g. at the reflexive). This suggests that reflexive attraction would be most prominent following ungrammatical verbs (see Fig. 2A). When the verb is ungrammatical, accidental access to a distractor is likely due to agreement attraction at the verb. In contrast, when the verb is grammatically inflected, or non-inflected, the distractor is less likely to be accessed: Any verbal cue that matches the distractor would also match the subject, but only the subject matches the verb’s structural cue. Thus, the distractor’s activation would be boosted prior to the reflexive only in ungrammatical verb conditions, leading to increased rates of retrieval errors and a processing speedup.

Differences between grammatically inflected verbs and non-inflected verbs are more nuanced, in CBR. The latency contrast between match and mismatch configurations is predicted to be larger following grammatically inflected verbs relative to non-inflected verbs (Fig. 2A), under CBR. Let us see why. A grammatically inflected verb matches both the subject and the distractor, in the match configuration. The dual match reduces the activation of the subject (due to the so-called “fan effect”). Overall low activation levels make later retrieval slower (at the reflexive), contributing to the contrast between match and mismatch configurations. The reflexive attraction is a speed-up in distractor mismatch relative to distractor match conditions. Thus, this fan effect contributes to our measure of attraction, even though it doesn’t necessarily make the distractor more attractive. With non-inflected verbs, on the other hand, the agreement features are not probed and therefore the distractor-target match does not affect retrieval latency. On the whole, the CBR model predicts a graded pattern of reflexive attraction effects. It predicts facilitatory interference in Fig. 1′s (b), (d), and (f) relative to (a), (c), and (e). However, the effect sizes would be such that the (f)-(e) contrast is the most prominent, and the (a)-(b) contrast would be the least prominent. For additional information on the simulations, see Appendix A.

In contrast, the agreement updating hypothesis predicts the most prominent reflexive attraction following non-inflected verbs (Fig. 2B), that is in Fig. 1′s (a), (b) contrast. Under this approach, the distractor might distort the representation of the subject, but subsequent verbal agreement would reduce the reliance on this distorted memory. When the verb agrees with the subject (grammatical conditions, (c-d) in Fig. 1), an updating strategy would undo the distractor’s interference and restore the subject’s correct features. When the verb does not agree with the subject (ungrammatical conditions, (e-f) in Fig. 1), the verb will shift the subject’s representation, regardless of the existence of a matching distractor. This would again nullify the effect of the distractor. Therefore, aligning the memory representation of the subject with (grammatical or ungrammatical) verbal features would reduce reflexive attraction. When the verb is non-inflected (Fig. 1′s (a, b)), reflexive attraction would be free to arise since nothing blocks the distractor from distorting the representation of the subject.

The predictions of the agreement updating hypothesis above assume that the comprehender edits the subject due to their uncertainty about representation of previous constituents. As explained in 1.2, rational inference could also be based on inference of production, and mostly typing, errors. In that case, editing is determined based on likelihood of letter insertion/deletion. We focus on general memory uncertainty rather than letter edits, for two reasons. First, it is not clear that comprehenders consider letter edits in cases of subject-verb mismatch. For example, Keshev & Meltzer-Asscher (2020) did not find an effect of letter edits on noisy-channel inference about Hebrew subject-verb agreement. More recently, Qian and Levy (2023) found that comprehenders tend to edit verbal agreement despite larger letter level distance, presumably due to the higher likelihood of production error. Thus comprehenders might consider the likelihood of whole morpheme errors rather than likelihood of letter edits in the context of subject-verb agreement (but see Poliak et al., 2023 for other evidence from Russian). In addition, we would like to note that it is not trivial to derive predictions of rational inference with regard to letter edits given other distortions to memory. It is easier to delineate rational letter edits if all corruption processes (agreement attraction and agreement updating) are assumed to be a form of letter level rational inference. However, we are currently hesitant to frame distortions originating from the distractor as rational inference. See Appendix D for additional details about a rational letter edit account.

To sum up, CBR predicts that reflexive attraction is less prominent when the verb bears no agreement cues, certainly relative to ungrammatically inflected verbs and possibly also relative to grammatically inflected ones. The agreement updating hypothesis, on the other hand, predicts the opposite interaction, by which reflexive attraction is more prominent when the verb bears no agreement cues (relative to both grammatical and ungrammatical verbs).