Glossa Psycholinguistics

1.1 Comprehension, production, and shared representations

The extent to which language production and comprehension draw on a shared set of representations is a long-standing question (e.g., Bresnan & Kaplan, 1984; Chater, McCauley & Christiansen, 2016; Garrett, 1980; Levelt, 1989; see Meyer et al., 2016, for review). On one hand, shared representations are a parsimonious explanation for overlap between language users’ production and comprehension repertoires, and for the fact that children apparently learn to produce by comprehending (Bresnan & Kaplan, 1984; Chater et al., 2016). On the other hand, comprehension and production unfold differently in real time, with different types of uncertainty and different goals, and each process may be best served by different types of representations (Clark & Malt, 1984; Indefrey, 2018; Levelt, 1989). Within the larger debate, evidence such as robust priming of abstract syntactic structures from comprehension to production (e.g., Bock et al., 2007) has led to a relatively broad assumption that syntactic representations are shared, even if phonological or phonotactic representations may not be (e.g., Kittredge & Dell, 2016).

If production and comprehension reference the same body of shared syntactic knowledge, then how do mismatches between them arise? First, there are reasons other than absence of representation for something to be absent from production. For instance, people who swear little or not at all avoid the relevant words and phrases because they are taboo, not because they cannot produce them. Second, people may appear to comprehend a structure (or even comprehend it quite successfully), despite lacking the relevant representation. This is one possible explanation for instances like Till death us do part, where a comprehender understands the gist of a structure despite being unable to produce novel instances. In such cases, comprehension might succeed by repurposing other knowledge through analogy, a bit like successfully turning a screw with a coin rather than a screwdriver, or by gathering meaning from context rather than building it compositionally.

On the other hand, it may be that, despite evidence for syntactic priming from comprehension to production and children’s apparent ability to learn to produce by comprehending, representations are not shared. Evidence for priming from comprehension to production comes from structures that participants both produce and comprehend (e.g., passives; Bock et al., 2007), and demonstrates robust connections between modalities. This evidence is therefore straightforwardly compatible with shared representations, but can also be modeled by a system with separate but linked representations (e.g., Clark & Malt, 1984). If representations are not shared, then an additional source of comprehension-production mismatches becomes possible: language users may have representations in their comprehension repertoires that are not in their productive repertoires.

Demonstrating the existence of structures that people comprehend without analogy or reliance on context but which they cannot effectively produce would present a strong challenge to models with fully shared representations. Conversely, evidence that even the most accurate comprehension of unproduced structures is achieved through analogy or context would favor models with shared representations. One promising place to test these possibilities is in closely related dialects, which often have both important structural differences and large areas of overlap.

1.2 Evidence from comprehension of unproduced structures

Several threads of prior work have explored people’s comprehension of structures they do not produce, but which are attested in closely related dialects. The findings suggest a range of patterns in how comprehenders handle such structures: a lack of representation for unfamiliar structures but a role for analogy in interpreting them, existing representations for at least some familiar unproduced structures, and cases where people achieve accurate comprehension of unproduced structures in context. We outline some of the evidence for these patterns here.

A variety of studies have explored people’s adaptation to unfamiliar, unproduced structures (e.g., US Midland needs Ved; Maher & Wood, 2011) over the course of an experiment (e.g., Kaschak & Glenberg, 2004). Participants are typically slow to read such structures early in an experiment but speed up with exposure. This has been proposed to reflect an initial absence of representation, followed by learning (Fine et al., 2013; Fraundorf & Jaeger, 2016; Kaschak & Glenberg, 2004). However, studies show that participants also speed up on superficially similar, unattested structures after exposure to needs Ved (e.g., The lawn needs mow, Boland et al., 2023; see also The copier will recycled, Fraundorf & Jaeger, 2016; also compare Weissler & Brennan, 2020), which suggests that any representations participants build over the course of the experiment are vague, at best. Intriguingly, readers unfamiliar with needs Ved show patterns of acceptance with different verbs (needs, wants, likes) that echo those of people who use the structure (Blanchette et al., 2026), suggesting a role for analogy or application of other linguistic knowledge in the absence of an existing representation.

Other studies show that people have existing representations for familiar unproduced structures. For instance, Squires (2014) found that Mainstream US English (MUSE) users were consistently faster to read a common but unproduced structure, singular don’t (e.g., the turtle don’t walk very fast), as compared to a superficially similar unattested one, plural doesn’t (e.g., the turtles doesn’t…). This result suggests that MUSE users began the experiment with an existing representation of the familiar structure, and that this representation held enough detail to differentiate it from the unattested one. Together, these findings suggest that: (i) exposure to a structure may encourage the development of a representation, (ii) in the absence of an existing representation, other linguistic knowledge may be recruited, and (iii) with longer (and perhaps richer) exposure, the representation becomes at least somewhat more detailed and specific.

Like Squires (2014), the current article explores a feature that, though absent from the production patterns of our participants, is common in many widely used varieties of English and therefore likely to be familiar: negative concord (e.g., They didn’t bake no cookies ‘They baked no cookies’). Like singular don’t, it is a structure that MUSE users have substantial exposure to, but do not typically produce. Unlike singular don’t, it is interpretively complex, allowing us to explore the comprehension process. Prior work has shown that when sentences like They didn’t bake no cookies are presented with preceding contextual support, MUSE users readily interpret them with a single semantic negation (i.e., ‘They baked no cookies’), despite the absence of this pattern in their production (Blanchette & Lukyanenko, 2019a; see 1.4). The current article explores how MUSE users arrive at this single negation interpretation, focusing on the role of context and analogy. To do this, we provide a following, rather than a preceding supportive context, and compare negative concord directly to a linguistically parallel, produced structure, namely, negative polarity item constructions (e.g., They didn’t bake any cookies).

Below, we outline key patterns in English negation and prior work addressing MUSE users’ comprehension of negative concord before providing more detail on the design and predictions for the current study.

1.3 Negative dependencies in English

To express a single semantic negation in a sentence with a negative auxiliary, English speakers have three primary options: they can use a bare plural, as in (1), a negative polarity item (NPI; e.g., anybody, anything, any X), as in (2), or a negative noun (e.g., nobody, nothing, no X) in a negative concord structure, as in (3).

Examples (1–3) have similar meanings (Childs, 2017), but they are not identical in terms of their structure. While the bare plural things in (1) can stand on its own, the NPI anything in (2) and the negative noun nothing in (3) are tied to the preceding negation. NPIs need a preceding element like negation to sound natural (*She said anything; Ladusaw, 1979), and in negative concord structures, the negative particle n’t and the negative noun work together to create a single semantic negation (Haegeman & Zanuttini, 1996). Thus, sentences like (2) and (3) involve negative dependencies, in which a noun phrase is grammatically dependent on a preceding negation. Given the choice between NPI constructions like (2) and negative concord sentences like (3), MUSE users typically choose NPIs (Robinson & Thoms, 2020; Wolfram & Fasold, 1974).

Interestingly, in antecedent clauses in conditionals, as in (4–6) below, the same three types of noun phrases are all possible, but result in two different meanings:

In (4) and (5), the surprising event is the person saying something, and in (6) it is them saying nothing. Thus, words like anything and nothing contribute the same meaning under negation, and opposite meanings in conditionals.

It is important to note that strings like (3) have two possible interpretations: In addition to the negative concord reading described above, it is also possible for each syntactic negation to contribute a separate semantic negation, resulting in what is known as a double negation interpretation (i.e., She didn’t say nothing, meaning ‘It is not the case that she said nothing’ or ‘She said something’). The double negation interpretation is considered prescriptively correct and occurs in MUSE users’ production, though infrequently, due to its heavy pragmatic conditioning (Horn, 2001). In contrast, there is a strong prescriptive bias against the negative concord use of such sentences, and despite its frequent occurrence in English more broadly, negative concord is generally absent from MUSE production (Wolfram & Fasold, 1974).

1.4 Previous evidence of MUSE users’ comprehension of negative concord

Prior studies have explored aspects of MUSE users’ comprehension of negative concord.¹ In one, Blanchette and Lukyanenko (2019a) used eye-tracking during reading to explore MUSE users’ comprehension of the negative concord and double negation interpretations of sentences like (3). The authors used preceding context sentences, such as those shown in (7), to encourage the negative concord or double negation interpretation of critical sentences with two negations, as in (8).

MUSE users spent less time reading critical sentences and less time rereading context sentences in negative concord contexts than in double negation contexts. That is, analyses showed that the negative concord interpretation was easier for MUSE users to reach than the double negation interpretation. This result provides an initial suggestion that MUSE users’ comprehension repertoires may include an existing representation for negative concord, despite its absence from production and the availability of another interpretation for the same string.

But how does MUSE users’ comprehension of negative concord compare to that of equivalent structures they produce? In a second study, Blanchette and Lukyanenko (2019b) compared MUSE users’ interpretation of NPIs, negative noun phrases, and bare plurals (Table 1), using offline judgments. In each sentence, the first clause was either negative or conditional and contained one of the three noun types, resulting in six combinations: negative-bare plural, negative-NPI, negative-negative, conditional-bare plural, conditional-NPI and conditional-negative. The second clause served as following context, clarifying and elaborating on the meaning of the first clause.

Participants rated the naturalness of the first clause and the extent to which the second clause made sense as a follow-up on a scale of 1–7. Second clauses after negative first clauses were designed to follow from a single negation interpretation, while second clauses after conditionals were designed to follow from a non-negative reading. Thus, second clauses should make sense following all first clauses except conditional-negative ones, in which the meaning reverses, as in (6) above. For example, in the last sentence shown in Table 1, it does not make sense that a backpack would be lighter if nothing had been left in a locker.

MUSE users’ naturalness ratings reflected their production patterns: They rated negative-negative first clauses low (median = 3) and all other first clauses high (medians = 6–7). Despite this, participants’ second clause ratings reflected that they understood all first clauses with the intended interpretation. In particular, they rated second clauses as similarly and highly sensible following both negative-negative and negative-NPI first clauses (medians = 7), and as nonsensical following conditional-negative sentences (median = 1),² meaning that they understood all three as having a single negation interpretation. As this was an offline task, it is unclear precisely how participants arrived at the single negation interpretation. While it is possible that the single negation interpretation was their first interpretation of negative-negative first clauses, it is also possible that they initially arrived at a double negation interpretation (compare, e.g., adults’ double negation interpretations in Thornton et al., 2016), which they revised upon encountering conflicting information in the second clause.

1.5 The current study

Negative concord represents an accurately comprehended, unproduced structure for MUSE users. The current study is designed to help clarify how MUSE users arrive at the intended single negation interpretation of negative concord sentences, with particular interest in whether doing so requires drawing on analogies or context. Evidence for required use of context or analogy would suggest that, despite the relative ease and accuracy of their comprehension (Blanchette & Lukyanenko, 2019a), MUSE users might not have an existing representation of the structure to draw on. In contrast, evidence for successful interpretation without reliance on context or analogy would argue for a rich existing representation of an unproduced structure. To investigate this, we used eye-tracking while reading to explore MUSE users’ comprehension processes as they read the stimuli from Blanchette and Lukyanenko (2019b). This allowed us to compare reading patterns for negative concord and negative-NPI constructions in sentences where the contextual support for the negative concord reading follows the relevant clause rather than precedes it.

If participants in Blanchette and Lukyanenko’s (2019b) study succeeded with negative concord items by revising an initial double negation interpretation upon encountering the context provided by the second clause, we predict slower reading of the second clause and more regressions to the first clause in negative-negative items than in negative-NPI items. In contrast, if participants interpret negative-negative items with a single negation without relying on context, we expect second clause reading patterns to be similar in negative-negative and negative-NPI items. This design also allows us to ask whether MUSE users rely on the simplest analogy – treating nothing as interchangeable with anything when this swap would resolve a seemingly incongruous meaning. If participants are willing to make this swap, they should be able to salvage the otherwise nonsensical non-negative continuations in conditional-negative items, and we would predict similar reading patterns in conditional-negative and negative-negative items.

While an answer in either direction will not resolve the question of whether syntactic representations are shared between comprehension and production, determining whether there is a key role for context or analogy in MUSE users’ comprehension of negative concord represents an important step in determining the existence or absence of structures that are deeply and accurately comprehended, but absent from the production repertoire.

2.1 Participants

Fifty-three participants (41 women, 12 men) were recruited in State College, Pennsylvania, USA. Participants were 18 to 66 years old (median = 23, mean = 28). All grew up in the US, most in the mid-Atlantic (n = 46), and a few elsewhere (Pacific Northwest, n = 2; Florida, n = 2; Midwest, n = 2; California, n = 1). All reported English as their first language, and 6 also reported early exposure to another language (Spanish, n = 3; Arabic, n = 1; ASL, n = 1; Italian, n = 1). Many participants were college students. Highest levels of education completed were therefore high school or some college (n = 28), 4-year college (n = 9), some post-graduate study or a post-graduate degree (n = 16).

In order to confirm that the participants were MUSE users, after the main task, we asked them to rate how likely they and their family and friends were to use negative concord or NPIs to express negative meanings on a scale from 1 (never) to 7 (always).³ As shown in Table 2, participants’ self-ratings were low for negative concord and high for NPIs. Participants were slightly more likely to ascribe negative concord use to family and friends than to themselves, perhaps because of social pressure or simply because social networks can be large and may include one or more negative concord users. Only one participant rated themself as more likely to use negative concord than NPIs (self, 5 vs. 4; family and friends, 5 vs. 3).⁴ An additional two participants indicated that they never hear NPIs from family and friends. However, these same participants rated themselves as quite likely to use NPIs to express the relevant meanings (6 and 7), and gave very low ratings for both NC use and exposure (all 1s). It is not fully clear what they intended to indicate with the low NPI exposure ratings, but, on the whole, these results suggest that our participants were not regular negative concord users.

2.2 Materials and design

The stimuli for the current study were the same as in Blanchette and Lukyanenko (2019b). The critical items were 48 two-clause sentences, each with 6 versions created by crossing first-clause type (negative, conditional) with first-clause object NP type (bare plural, NPI, negative NP), as shown in Table 3. The critical sentences were arranged into 6 lists using a Latin Square design. The 48 critical sentences in each list were interspersed with 112 two-clause fillers of similar complexity. See the supplementary materials for a full list of stimuli. To maximize the likelihood that negative concord would seem natural, sentences used features common in casual spoken English (e.g., gonna, contractions, there’s + plural).

Sentence order was randomized for each participant, with the constraint that critical sentences could not appear adjacent to one another. Twenty-four percent of the sentences were followed by true/false comprehension questions, and 15% (16 fillers, 8 critical) had second clauses that did not make sense following the first.

The experiment was run on an SR Research Eyelink 1000+ eye-tracker in head-stabilized mode. Stimuli were presented in black text (25-point Consolas) on a white background on a 24-inch monitor.

2.3 Procedure

The eye-tracking task took approximately 30 minutes and included 9-point calibration. The experiment began with 4 practice trials, half with sensical and half with nonsensical continuations, each followed by a true/false comprehension question that targeted details or direct inferences from one or both clauses (e.g., Sentence: The fluffy dog dragged mud all over the living room, so his owners are gonna reward him with a treat. Comprehension Question: The living room was dirty because of the dog. T/F?). During the practice phase, comprehension questions were followed by feedback on whether participants’ responses matched the expected response. Of the 160 test trials, 24% were followed by true/false comprehension questions without feedback. Comprehension questions were intended to encourage reading for comprehension and attention to content, but did not target the key negative or non-negative meanings. On each trial, the two clauses appeared on separate lines. Verification questions, when they occurred, appeared on a subsequent screen. Participants completed trials at their own pace and were prompted to take a break halfway through the experiment. Participants completed a demographic questionnaire immediately following the eye-tracking task. In total, participation took about 45 minutes.

2.4 Data preparation and exclusions

Fixations that were short (<80 ms) and close (<.5° of visual angle) to a temporally adjacent fixation were merged. Remaining fixations shorter than 40 ms or longer than 1200 ms were removed from the data. This resulted in the removal of 342 fixations (0.2%). Fixations outside the interest areas were also removed (1106 fixations, 2%).

At the beginning of data collection, five items inadvertently included typos in one or more of their forms. These typos were subsequently fixed. To maintain a reasonable balance between sentence types, all versions of the affected items were removed for participants before the fix (n = 135 trials, 5% of 2544 possible trials), and all were retained for participants afterward. Additionally, trials were eliminated from the dataset if they contained no data (n = 7), or if the participant did not fixate both clauses (n = 17). This left 2385 trials for analysis.

All participants met the inclusion criteria: all had at least 75% accuracy on the verification questions for sensical fillers, and no participant had more than half of their trials in any condition excluded.

3.1 First-clause go-past time

As shown in Figure 1, participants were slower to read negative-negative (i.e., negative concord) first clauses than any other type. Interestingly, conditional first clauses appeared somewhat slower than negative first clauses for bare plural and NPI items, perhaps due to the relatively complex logic involved in interpreting them (e.g., Fugard, et al., 2011). However, the large slow-down in go-past time for negative-negative first clauses caused this pattern to reverse for negative object NPs, such that negative sentences were read more slowly than conditional ones.

Figure 1: Mean go-past time for the key region of the first clause. Error bars show standard error of the mean.

The slow-down for negative-negative first clauses was confirmed in the linear mixed-effects model analysis, the results of which are shown in Table 4. There was a main effect of the NPI vs. negative NP contrast, and a significant interaction of sentence type and the NPI vs. negative NP contrast, such that the slow-down between reading an NPI object and a negative object was much more dramatic for negative sentences than conditional ones. This is consistent with participants being surprised to encounter negative concord in writing. A simple-effects analysis showed that the NPI vs. negative NP contrast is reliable for the negative sentences (B = 270.92, t = 6.45, p < .001), but not the conditional sentences (B = 44.12, t = 1.05, p = .29). There was also a reliable effect of trial order, as participants sped up over the course of the study.

3.2 Second-clause total time

Figure 2 shows that participants spent substantially more time reading the key region of the second clause in conditional-negative sentences, where that continuation did not make sense, than in negative-negative sentences. This pattern suggests that participants arrived at the intended negative concord interpretation of negative-negative sentences before reading the second clause. We also see an unpredicted difference between the conditional-NPI and negative-NPI sentences.

Figure 2: Mean total time for the key region of the second clause. Error bars show standard error of the mean.

These patterns were largely confirmed by the linear mixed-effects model (Table 5). Reliable main effects of sentence type and the NPI vs. negative NP contrast indicated that continuations were read more slowly in conditional items overall and in items with negative NPs, and a reliable effect of trial order indicated a tendency to speed up over the course of the study. The predicted interaction between sentence type and the NPI vs. negative NP contrast was not significant, likely due to the unexpectedly long reading times in conditional-NPI sentences. However, a simple effects analysis showed that the NPI vs. negative NP contrast was reliable for the conditional sentences (B = 89.58, t = 2.06, p = .04), but not for the negative sentences (B = 54.15, t = 1.24, p = .21). We return to this discrepancy in 3.4, below.

3.3 Rereading time

Figure 3 shows that, overall, participants spent more time rereading the key region of the first clause after entering the key region of the second clause in conditional sentences than in negative sentences. This difference was somewhat larger for sentences with negative NPs than for those with NPIs or bare plurals.

Figure 3: Mean rereading time for the key region of the first clause, given that the key region of the second clause has been fixated. Error bars show standard error of the mean.

The results of a linear mixed-effects model analysis of rereading time are shown in Table 6. There was again a reliable effect of trial order, sentence type, and of the NPI vs. negative NP contrast, but no reliable interaction between sentence type and the NPI vs. negative NP contrast. A simple main effects analysis showed that the increase in rereading time between NPI and negative NP was marginal for the conditional sentences (B = 97.59, t = 1.94, p = .05), and not reliable for the negative ones (B = 68.13, t = 1.35, p = .18).

3.4 Discussion of statistical patterns

While the predicted interaction between sentence type and the NPI vs. negative NP contrast did not emerge for second-clause total time or for first-clause rereading time, this appears to be for reasons unrelated to our central questions. In particular, it seems that the expected interaction failed to emerge not because of difficulty associated with the interpretation of negative-negative (i.e., negative concord) sentences, but rather because of an unpredicted pattern of longer reading times in conditional sentences, particularly conditional-NPI sentences. Despite the lack of a significant interaction, simple-effects analyses showed the predicted patterns within sentence types. Participants spent reliably more time reading second clauses and marginally more time revisiting first clauses in conditional-negative than in conditional-NPI sentences. They did not show reliable differences between negative-NPI sentences and negative-negative sentences on either measure.

There are likely multiple sources for the larger-than-expected differences between negative-NPI and conditional-NPI sentences for the revision measures. First, conditional sentences were read more slowly than expected, overall. Longer reading times began emerging for conditional sentences in first-clause go-past time (Figure 1). This, along with prior literature (e.g., Fugard, et al., 2011), suggests that there are complexities in processing conditionals that we did not account for when we designed our comparisons, and which are neither related to the fit between first and second clauses nor to the type of object NP. Second, there is a substantial increase in first-clause rereading time and second-clause total time for conditional-NPI sentences, relative to conditional-bare plural sentences. This may be due to the complexity added by the dependency between the conditional context and the NPI, but may also be due to the stimulus materials: Revisiting the results of Blanchette and Lukyanenko (2019b) with the current results in mind, we notice a wider spread of second-clause sensicality ratings for conditional-NPI sentences (1^st quartile = 5, median = 6, on a 7-point scale) than for negative-NPI sentences (1^st quartile = 7, median = 7). Along with comments from an anonymous reviewer and our own reflection, this pattern suggests that some of the non-negative continuations may not have followed as sensibly from conditional-NPI clauses as intended. Compare, for instance, the sentences below, repeated from Table 3.

It follows fairly directly that if nobody is warned, as in (9), people will think it’s safe to stay. It is less certain that if anybody is warned, as in (10), people in general will know that it is risky to stay. For instance, the news anchor warning the camera crew would satisfy the first clause of (10), but would be unlikely to result in the general awareness implied by the continuation. This pattern was not universal in the stimuli (see Table 1: If my older sister leaves anything in her locker, then her backpack will be a bit lighter during her walk home), but it may have contributed to the larger-than-expected differences between conditional-NPI and negative-NPI sentences.

In contrast to the unexpected patterns in conditional sentences, patterns for the negative sentences were in line with expectations: negative-bare plural and negative-NPI sentences showed relatively similar reading times on all three measures. It was only on the initial encounter with negative-negative first clauses (go-past time) that participants reliably slowed down. In particular, we note that second-clause total times are numerically comparable in negative-bare plural and negative-negative sentences. This suggests that unproduced negative concord structures are interpreted with similar ease to produced structures, both those with a negative dependency (negative-NPI) and those without (negative-bare plural).

On the whole, despite the lack of the predicted interaction between sentence type and the NPI vs. negative NP contrast, we take the statistical patterns to support the conclusion that MUSE users arrive at the negative concord reading of negative-negative sentences without engaging in a context-driven revision process.

4.1 Implications for shared representations

The current results rule out a reliance on context and the simplest analogies as explanations for MUSE users’ success with negative concord, and are consistent with MUSE users accessing an existing representation during comprehension, despite negative concord’s systematic absence from their production. Negative concord therefore remains a candidate mismatch between MUSE users’ comprehension and production repertoires and an interesting case to investigate further.

The current results are compatible with the possibility that comprehension and production draw on separate syntactic representations. There are two ways in which these results remain compatible with a shared representations approach, however. First, negative concord may be absent from the shared repertoire, and participants may succeed through a more complex, linguistically constrained analogy. If the complex analogy to negative-NPI constructions applies, then negative concord is absent from production because nothing prompts the analogy. Second, negative concord may be present in a shared repertoire. If this is the case, then negative concord structures are absent from production either because of intentional avoidance, due to prescriptive prohibitions on “double negatives” (Horn, 2010), or a more subtle effect of the low frequency of the structure in MUSE, as might be predicted on a number of models (e.g., Chater, et al., 2016; Robinson & Thoms, 2021).