Using Structural Priming to Study Scopal Representations and Operations

Emmanuel Chemla, Lewis BottLinguistic 2015 · DOI

TL;DR

This paper investigates how we understand ambiguous sentences, specifically those with quantifiers like "every" and "a." The researchers found that people are "primed" to interpret these sentences in a certain way after being exposed to similar structures, suggesting this priming happens at a representational level rather than through a specific processing operation.

What Problem Does This Paper Address?

Sentences can have multiple meanings depending on the order in which quantifiers are interpreted, a phenomenon known as "scope ambiguity." For example, "Every student read a book" could mean each student read a different book (universal-wide scope) or that there's one specific book that all students read (universal-narrow scope). Previous research showed that exposure to one interpretation can influence how people interpret subsequent ambiguous sentences. However, it was unclear whether this "priming" effect stemmed from people developing a specific mental representation of the sentence's meaning or from them using a particular processing operation (like a scope-reversing mechanism). This paper aims to disentangle these two possibilities.

Methods

The study employed a structural priming paradigm, similar to previous work. Participants were shown "prime" sentences paired with pictures that forced a specific interpretation (either wide or narrow scope). Then, they were shown "probe" sentences, which were also scopally ambiguous, and asked to choose the picture that best matched their interpretation. The key innovation was testing priming across different types of ambiguous sentences (Universal-Existential, Existential-Universal, and Universal-Negation) to see if priming occurred based on the interpretation itself or the operation used to derive that interpretation. Statistical analysis using logit mixed-effect models was used to analyze the response data.

Datasets

The study used custom-generated experimental stimuli. Participants were presented with sentences constructed from three frames:

  • Universal-Existential (U-E): "Every [Shape 1] is [above/below] a [Shape 2]"
  • Existential-Universal (E-U): "There is a [Shape 1] [above/below] every [Shape 2]"
  • Universal-Negation (U-neg): "Every [Shape 1] is not [Color]"

These sentences were paired with images designed to disambiguate their meaning. No pre-existing datasets were used for training or evaluation.

Key Insights and Ideas

  • Representation over Operation: The most significant finding is that priming appears to be driven by the activation of abstract representations of sentence meaning, rather than by the repeated use of a specific processing operation (like scope reversal).
  • Cross-Sentence Priming: Priming effects were observed not only within the same sentence structure but also between different types of scopally ambiguous sentences, supporting the idea of a shared representational level.
  • Impoverished Representations: The representations involved in this priming are "impoverished," meaning they focus on the core scope-taking elements (quantifiers, negation) and abstract away from other linguistic details.
  • U-neg Sentences are Different: Sentences involving negation (U-neg) did not show significant priming effects, suggesting they might be processed differently or that the priming mechanisms at play are specific to quantifier interactions.

Key Results

  • Participants were more likely to choose an interpretation for a probe sentence that matched the interpretation of the preceding prime sentence, demonstrating priming.
  • This priming effect was consistent across different sentence configurations (U-E and E-U), supporting the "representation account." If it were operation-based, priming would have been expected to go in opposite directions for U-E and E-U sentences.
  • Specifically, U-wide primes led to U-wide responses in probes, regardless of the probe sentence type, which aligns with a representational priming hypothesis.
  • No significant priming effects were found for the U-neg sentences, suggesting they might be an exception or require different experimental conditions.

Limitations and Future Work

  • Visual Priming: The authors acknowledge the potential for "visual priming" from the images themselves, though they argue their experimental design and comparison with previous studies make this unlikely to be the sole driver of their results.
  • U-neg Sentence Anomaly: The lack of priming for U-neg sentences is a notable limitation. The authors suggest this could be due to design choices, shallow processing, or a fundamental difference in how negative operators are processed. Future work could explore this further.
  • Scope Mechanisms: The paper suggests future research should investigate other "kinds of scopal mechanisms" and how they interact with different linguistic elements beyond quantifiers. Testing priming across sentences with entirely different quantifier pairs could also reveal more about the generalizability of these findings.

Summary updated

Abstract

Sentence (1) contains a universal quantifier every and an indefinite, existential quantifier a. Depending on which of these two elements takes scope over the other, we obtain two possible interpretations, as paraphrased unambiguously in (2).The universal interpretation in (2a) is surface scope: the scopal relation between the universal quantifier every and the existential quantifier a matches the order and hierarchy in which they appear in (1). The interpretation in (2b) is reverse scope: the scopal relation is reversed between the sentence and the interpretive level.Our point of departure is an interesting finding by Raffray and Pickering (2010), who demonstrated that people can be primed to derive particular scopal interpretations (see details below). In this squib, we emphasize that such priming results are useful either to characterize realistic layers of linguistic representation or to confirm the existence of linguistic operations of certain kinds. We do so by discussing two possible interpretations of Raffray and Pickering’s results and testing them in an experiment.Raffray and Pickering’s (2010) study involved a structural priming paradigm (e.g., Bock 1986, Branigan, Pickering, and McLean 2005, Pickering and Branigan 1998, Thothathiri and Snedeker 2008; for a review, see Pickering and Ferreira 2008). Participants completed a sentence-picture verification task that involved matching one of two pictures to a scopally ambiguous sentence, such as (1). There were prime trials and probe trials. In prime trials, participants were forced to interpret the sentence with universal wide-scope (or universal narrowscope) by the nature of the target picture, which contained a widescope image (or a narrow-scope image), and the foil picture, which contained an image that was inconsistent with the sentence. In probe trials, which immediately followed prime trials, one of the pictures corresponded to the universal wide-scope interpretation of the sentence and the other to the universal narrow-scope interpretation. Participants were therefore ‘‘free’’ to choose either interpretation in the probe trials. Across two experiments, Raffray and Pickering found that participants were more likely to select the picture matching the wide-scope interpretation following a wide-scope prime trial than following a narrowscope prime trial. In other words, participants were primed to derive sentence interpretations with particular scopal relations. From these results, Raffray and Pickering concluded that participants formed disambiguated abstract representations that specify quantifier scope relations.Concretely, the claim is that there exists a level of representation that looks like the following semiabstract patterns for the two possible interpretations of (1):These representations abstract away from some information—for example, the content of the lexical material filling in the . . . parts. If these representations exist, we can understand why the activation of one of these patterns for a given sentence (the prime) can strengthen the activation of the same pattern for a subsequent sentence.Priming effects can reveal a level of representation at which prime and probe are made equivalent, as discussed above. But they can also argue for the existence of some operation that applies equally to both prime and probe. In our concrete case, there might be an operation that transforms an (a)-type interpretation into a (b)-type interpretation. This operation would correspond to a scope-reversing operation, call it OSR. (An alternative plausible candidate for such an operation when indefinites are involved would be a domain-narrowing operation; we come back to this in the discussion section (section 3)). Such OSR operations would traditionally fall under the label of ‘‘movement’’ operations in the syntactic literature (e.g., Quantifier Raising; see, e.g., Fox 2000). Formal and experimental results argue for the existence of such operations. For instance, developmental inquiries show that children under a certain age mostly access surface scope interpretations for a variety of configurations and only later the whole set of their corresponding reverse scope interpretations. This suggests that scope reversal is indeed a single piece in the system, a coherent operation that children do not use early on, but that they then, in one step, master across the board (see, e.g., Conroy, Lidz, and Musolino 2009 for more detailed discussion).In terms of this OSR operation, the (a)-type interpretation would follow from a representation in which hierarchical order and linear order have been aligned. The OSR operation could apply to this representation and reverse the scopal relations. In representational terms: (b) = OSR(a).We can now describe an alternative interpretation of the priming effects mentioned above, as evidence for this OSR operation. Indeed, if the OSR operation has been applied once to the prime, it may apply again to the probe, more so than if the operation was not applied to the prime. Hence, priming results may be seen as evidence for layers of representation (section 1.2) or for linguistic operations involved in these representations (the current section 1.3).We will provide new evidence in favor of the linguistic layer interpretation, which was Raffray and Pickering’s (2010) original interpretation of their data and which may not be the most naturally available from a linguistic standpoint. Our strategy will be to study priming effects between sentences of different profiles. Specifically, we will study sentences such as (4).Such sentences have the same two types of interpretations as the original example (1), the universal wide-scope interpretation and the universal narrow-scope interpretation. Crucially, however, the order of the quantifiers is reversed in (4): the indefinite a now occurs before the universal quantifier every. Hence, the two interpretations are obtained differently in (1) and (4) in terms of the OSR operation. Specifically, the OSR operation is involved to obtain the existential interpretation in (1) and the universal interpretation in (4). Hence, the surface scope interpretation of one sentence is the reverse scope interpretation of the other.If the interpretation account is correct, then we should find priming of interpretation across sentences of types (1) and (4): the universal interpretation of one should prime the universal interpretation of the other. If the operation account is correct, then we should find the opposite type of priming across these sentence types, such that the surface scope of one primes the surface scope of the other—that is, a universal wide-scope interpretation of one sentence primes a universal narrow-scope interpretation of the other.1The operation interpretation based on OSR makes broader predictions. First, priming may occur across sentences using completely different pairs of quantifiers: sentences of the form Q1 V Q2 and Q3 V Q4 (where Q1, Q2, Q3, and Q4 are quantified DPs involving four different quantifiers) could prime each other’s surface and reverse scope interpretations, even though their impoverished representations as in (3) have nothing relevant in common. Second, radically different types of sentences require scope resolutions that may involve the OSR operation. For instance, negation may interact with a quantifier in roughly the same way in which two quantifiers (every and a) may interact. Consider sentence (6), which may receive the two interpretations paraphrased in (7) depending on whether negation takes scope over or below the universal quantifier every.Descriptively, one may recognize in (7) a surface scope interpretation and a reverse scope interpretation, obtained by way of an operation that could have much in common with the OSR operation. If this is correct and if Raffray and Pickering’s (2010) priming effect is of the second type (priming of operation), then we predict priming to occur within negative sentences such as (6) and between sentences of this type and those with two quantifiers, such as (1). If the priming effect is based on interpretation and representations of the type in (3), then no such priming is expected.2Our experiment tested whether scopal priming occurs because a representation is primed or because an operation is primed. Participants completed a sentence-picture verification task, similar to Raffray and Pickering’s (2010). Each trial consisted of a sentence and two pictures, and participants clicked on the picture that best matched the meaning of the sentence. The experimental trials all involved scopally ambiguous sentences. There were prime trials, in which the pictures dictated that only one meaning was acceptable, and probe trials, in which either meaning was acceptable (see figure 1). Following Raffray and Pickering, we expected that the (forced) meaning of the sentence in the prime trials would influence the (free) meaning of the sentence in the probe trials. Across trials, sentences could share the same abstract interpretation (e.g., universal wide-scope) or the presence of the OSR operation (e.g., reverse scope). Of interest was whether scope operation or interpretation would predict priming direction.We used three types of configurations (sentences): universal-existential (U-E), as in (1); existential-universal (E-U), as in (4); and universal- negation (U-neg), as in (6). Each configuration could result in two interpretations: the universal wide-scope (U-wide) interpretation, as described in the (a) examples of (2), (5), and (7), or the universal narrow-scope (U-narrow) interpretation, as in the corresponding (b) examples (see the Interpretation column of table 1). The interpretations could also be classified with respect to the direction of the scopal relation, surface scope or reverse scope; but, crucially, the mapping between interpretation and the presence of a scope operation needed to obtain reverse scope varied across sentences (see the Operation column in table 1). The dissociation between interpretation and operation meant that the priming accounts could be differentiated. Priming was tested between all configurations. Thus, there were within-configuration trials, in which prime-probe pairs used the same configuration (e.g., E-U primes followed by E-U probes), and between-configuration trials, in which prime-probe pairs used different configurations (e.g., E-U primes followed by U-neg probes). For within-configuration trials, the interpretation and operation accounts make the same predictions. For between-configuration U-E/E-U trials, however, the interpretation and operation accounts make different predictions. The interpretation account predicts that between-configuration priming will show U-wide primes leading to U-wide responses, whereas the operation account predicts that U-wide primes should lead to U-narrow responses. Predictions for U-neg sentences can be extracted from table 1.Eighty participants were recruited online using Amazon Turk. Of these, 2 were removed because their native language was not English, and 5 more were removed because their accuracy on the prime sentences was less than 90%.Experimental sentences were constructed according to one of three frames.An important remark is in order. The E-U sentences used an expletive construction instead of a plain indefinite, which is degraded in subject position (see discussion and corpus studies in Michaelis and Hartwell 2007 and references therein). One may wonder whether this could block the possibility of a reverse scope interpretation, no matter how it comes about, but (jumping ahead a bit) table 2 shows that even though there is a bias toward a U-narrow reading, our participants also accessed the U-wide reading, which thus leaves room for us to investigate possible priming effects.3Shapes were hearts, squares, dots, triangles, stars, or notes, and colors were black, blue, green, or red. Examples are shown in figure 1. The list of stimuli was obtained by inserting colors, shapes, and above/below randomly into the appropriate sentence frame. This meant that different primes had different combinations of shapes and colors. However, there was only one list of stimuli, used for all participants.For each sentence, we constructed three types of images: a foil F consistent with none of the interpretations, a target image T1 consistent only with the first interpretation, and a target image T2 consistent only with the second interpretation.4Each trial consisted of a sentence and two of its three associated images, as follows:We included control trials using the same sentence frame, Every [Shape] is [Color], with one and only one of the two images making the sentence true.Primes and probes formed pairs of experimental trials. Prime trials were formed from one of the three sentence configurations and expressed as one of two interpretations (U-wide or U-narrow). Probe trials consisted of one of three configurations. A complete experimental set therefore consisted of (3 × 2) × 3 = 18 prime-probe pairs. We counterbalanced the position of the target image for the prime trials (left or right) and whether the target probe response was on the same side as the target prime response, to obtain 4 replications of 18 experimental pairs = 72 pairs. There were also 32 control trials, making the total number of trials 72 × 2 + 32 = 176. Experimental pairs and control trials were presented in a different random order for each participant.We analyzed our data by modeling response-type likelihood using logit mixed-effect models ( Jaeger 2008). The random effects structure was maximal in the sense recommended by Barr et al. (2013); that is, random intercepts and slopes were included for all repeated-measures factors (all factors where subjects were random effects), but only intercepts were included for between-element factors (where items were used as random effects). In addition to the factors reported for the individual analyses below, we included prime-probe position consis tency as a fixed factor, that is, whether the correct response to the prime was on the same side as the subsequent probe.5 We report p-values derived from the χ2 likelihood ratio test comparing the more complex models to simpler models, and the Wald Z p-values resulting from the best-fitting model. Treatment coding was used throughout.Trials with incorrect responses to the primes were discarded, as in Raffray and Pickering 2010, and were also extremely rare (E-U primes: 1.3%; U-E primes: 1.4%; U-neg primes: 1.4%). Control trials were answered accurately (2.7% error rate) and were also discarded.Consider first the results from the E-U and U-E sentences (those tested by Raffray and Pickering (2010)). The left panel of figure 2 illustrates responses when primes and probes involved the same configuration, that is, within-configuration trials (E-U → E-U and U-E → U-E trials; raw counts are shown in table 2). When the probe is preceded by a Uwide prime, there are more U-wide responses to the probes, regardless of the probe type. Both the operation and the interpretation accounts predict this pattern. Model analyses involving prime interpretation (Uwide vs. U-narrow) and probe type (E-U vs. U-E) revealed a significant effect of prime interpretation (see table 3), χ2(1) = 6.51, p = .010, but no interaction, χ2 = 1.82, p = .18. Simple effects analysis revealed that there was a robust priming effect for E-U sentences, χ2(1) = 6.79, p = .019, but not for U-E sentences, χ2(1) < 1. Thus, while the priming effects we observe seem weaker than those found by Raffray and Pickering, we obtained substantial within-configuration priming with the same sorts of sentences.When primes and probes involve different configurations, the interpretation and operation accounts make different predictions. The interpretation account predicts that between-configuration priming will show effects similar to those of the within-configuration sentences, but the operation account predicts that priming for the E-U sentences will occur in the direction opposite to that of priming for the U-E sentences. The right panel of figure 2 shows the between-configuration U-E and E-U probe responses. We observe substantial priming effects for these sentences, consistent with an interpretation account. More specifically, there was a main effect of the prime interpretation (U-wide vs. U-narrow), χ2 = 11.97, p < .001, and no interaction between the probe type (E-U vs. U-E) and the prime interpretation, χ2 < 1. Simple effects analysis revealed that prime interpretation had a significant effect when the prime was E-U, χ2(1) = 5.38, p = .020, and when it was U-E, χ2(1) = 9.38, p =.0022. It is also possible to encode the prime interpretation as surface or reverse scope (see table 1), in which case reverse scope priming leads to more U-wide responses for the U-E sentences but fewer U-wide responses for the E-U sentences—that is, an interaction, χ2(1) = 13.19, p < .001. Both analyses indicate that interpretation is a good predictor of priming whereas order is a bad predictor.6There was no significant effect of any of the prime sentences on the U-neg sentences, regardless of whether the prime types were considered separately, U-neg N U-neg, χ2(1) = 1, E-U → U-neg, χ2(1)_1.50, p_.22, or U-E → U-neg, χ2(1) <1, or combined, with prime interpretation (U-wide vs. U-narrow) and prime type (Uneg, E-U, U-E) as factors, χ2(2) < 1; nor did the U-neg sentences prime either U-E or E-U probes, χ2(1)’s < 1. Although these results are consistent with an interpretation account, the lack of withinconfiguration priming makes the lack of between-configuration priming difficult to interpret, as we discuss below.The interpretation of scopally ambiguous sentences can be primed. Our experiment pitted priming of representations against priming of a scopal ‘‘movement’’ operation. We found evidence for the former, representation-based priming. (The latter, operation-based type of priming may still exist, but it is at best hidden by the former.)Before discussing their implication, we would like to discuss the possibility that our findings could be driven by mere ‘‘visual priming’’ between images. In one may wonder whether participants are not patterns in the images, such that clicked on a U-wide picture, they would be by a U-wide picture again (e.g., because they both two of three In some of its this is to from priming at the for a particular pattern in an image is at to for a particular meaning a mere of a There may be to the two One would be to the images so that the U-wide is hidden as much as but that is a matter of of that the relevant pattern by it would not be possible to whether the corresponding interpretation of the is Raffray and Pickering (2010) obtained results similar to and their images were in such a way that the pattern would have to be abstract to account for the whole alternative control for priming may in the same images, but the sentences in such a way that linguistic elements to any priming effect example, the scopally ambiguous quantifiers in the prime sentence and these with such as There are below the The of significant priming when there was no sentence would indicate that priming effects were not to priming. Indeed, Raffray and Pickering such a study and did not observe significant priming. as our study is similar to we would that our effects were also not by results confirm Raffray and Pickering’s (2010) original interpretation in terms of the existence of a certain level of This level of representation is It to elements such as universal and existential quantifiers, to the of content the between such as and and other elements such as of representation have also been on (see, e.g., Fox and to mere or from plain In the it would be interesting to how the two can be and how the two can each we find for Raffray and Pickering’s (2010) representational priming than an priming account, our with respect to the form of primed Raffray and Pickering tested whether sentences similar to those in (4) primed sentences like (1), as we have (see 1), but they did not observe between-configuration priming we thus concluded that to scope to quantified than the more representation we describe in (3) and in the like the account to by Raffray and There are between our study and such as the of U-neg sentences in our study and the of the pictures, and we which of these is for the in results, or indeed whether Raffray and Pickering a studies may be to the We the important however, is not between different of representational but between representational and which has been the of this study and in which there is no between our this one may wonder the of the U-neg sentences, for which we to find priming can be First, we to find priming because of some of the such as our of For instance, a mentioned that these sentences were by the presence of colors in our such that participants could be by these colors to the they would be and therefore on a of these sentences that would not the level of representation or to priming. If this is the then studies may be more in these Second, U-neg sentences might priming in a more the people have with negative (e.g., and that the scope representation across trials. Our data between these but evidence from the developmental literature suggests that negative scope can indeed be primed under the right the Lidz, and Musolino (2010) that the U-narrow interpretation of U-neg sentences could be primed with children using a sentence would to every is Although this is a different type of with a different it is with the that U-neg priming at an impoverished level of interpretation of the U-neg findings may also be useful in a of the priming The two possible interpretations in an and this could the priming a interpretation may prime a interpretation. This could be as a bias toward a to a sentence is a way to be by the to make it and in an experimental such as this toward may that may be with priming This would predict priming within and from U-neg sentences as which we did not However, this is based on the interpretation of a which may be in other as discussed above. in priming of how linguistic elements quantifiers interact would be of interest not only to this but also to whether or not similar scopal apply to different of scopal results priming effects at a at which some linguistic elements are (the between universal and existential quantifiers is and at which are For however, it would be interesting to whether other of scopal could the current For instance, one may that the to a may be the of wide-scope interpretation of indefinites (see such that it is the of the that is than the interpretation it leads to to and for understand we would to more For instance, one may whether priming is to the presence of an indefinite or some other that a in the to quantifier scope More one could test priming across sentences involving different pairs of quantifiers so that the possible interpretations are not at any Priming may or priming of scope operations may then
BibTeX
@inproceedings{chemla2015using,
  author = {Emmanuel Chemla and Lewis Bott},
  title = {{Using Structural Priming to Study Scopal Representations and Operations}},
  booktitle = {Linguistic},
  year = {2015},
  doi = {10.1162/ling\_a\_00178},
}
Figure