Dylan Hardenbergh ’16, Laine Stranahan, & Jesse Snedeker


Contrastive inferences are a robust class of language processing inferences at the semantics-pragmatics interface.  Prior work has demonstrated that when a subject is presented with a display of objects and auditory directions to select one of the objects, the presence of a contrastive pair item to the target object facilitates his or her convergence onto the target with the use of prenominal scalar modifiers.  Put in more concrete terms, if a tall glass and a short glass are present in an array of objects, a subject will assume the speaker is referring to the taller of the two glasses upon hearing the word “tall” in the command “pick up the tall glass,” as a rational speaker would include this adjective only if it were necessary to differentiate between two objects of the same type.  Little is known about the cognitive mechanisms underlying contrastive inferences and the extent to which they mirror other forms of psycholinguistic inference-making.  To explore this issue, we conducted a dual-task paradigm experiment that required participants to memorize letter sequences while completing a visual-world eye-tracking activity to assess contrastive inference processing.  The purpose of the letter memorization task was to burden some subjects with a high degree of cognitive load; previous work has suggested that cognitive load disrupts pragmatic inferences but leaves semantic elements of meaning unaffected, thus helping to disentangle these two aspects of language processing.  Our results demonstrate that cognitive load inhibits contrastive inference-making, which indicates that language processing facilitated by contrastive inferences is characterized by an initial semantic-only phase followed by the onset of pragmatic cues.  This finding suggests that contrastive inference-making imitates other modes of language processing at the semantics-pragmatics interface.


  1. An Overview of the Semantics-Pragmatics Interface

The process by which humans attribute meaning to language represents a complex interplay between the linguistic domains of semantics and pragmatics.  Semantics is concerned with the truth-conditional content of utterances and is not dependent on the context in which language arises.  Pragmatics, on the other hand, represents the elements of meaning not directly encoded in language; the perceived goals of a speaker, the environmental context, and shared communication norms between individuals all have the potential to contribute to the meaning of language at the pragmatic level of processing.  An incredible breadth of human language relies on pragmatics to convey meaningful ideas from one speaker to another.  A few examples of such utterances are:

  • The boy ate some of the cookies.
  • She hasn’t had anything to eat.
  • I’m so hungry I could eat a horse.

(1) relies on a type of pragmatic inference known as scalar implicature.  A rational listener would generally assume that this sentence means the boy ate some of the cookies but did not eat all of them, even though this conclusion isn’t a strict logical consequence of the language used.  (2) requires the listener to apply some understood timeframe to derive meaning from the sentence.  Presumably the person in question has eaten at previous points in her life, but it is implied that she has not had anything to eat within a specific timeframe of interest (perhaps this particular day or this particular mealtime).  The speaker in (3) does not intend to express literal meaning with this utterance, but rather assumes that his or her audience understands the meaning associated with this figure of speech (namely, “I’m very hungry”).

While pragmatics is clearly a vital component of human language, fairly little is known about the cognitive mechanisms underlying this diverse set of inferences and assumptions.  Furthermore, the division between semantics and pragmatics is often unclear and many instances of language processing do not fall cleanly within one realm or the other.  One notable distinction between semantics and pragmatics is that pragmatic inferences are cancelable and may disappear in certain contexts, while semantic meaning may not (Horn, 1972).  Indeed, if the speaker in (1) had instead stated, “The boy ate some of the cookies—in fact, he ate all of them,” the scalar implicature associated with this utterance would disappear.  Previous research on this interface has suggested that semantic levels of representation are generated first, and it is only after additional processing, much of which may be time-consuming and cognitively laborious, that pragmatic cues are incorporated into meaning (Huang & Snedeker, 2009).  This reality gives psycholinguists the capacity to differentiate between elements of meaning that are semantic (instantaneous) and pragmatic (delayed).

  1. Contrastive Inferences

Contrastive inferences are a well-attested phenomenon of the semantics-pragmatics interface.  These inferences are predicated on a shared understanding between individuals that speech ought to be neither over-informative nor under-informative (Carston, 1999).  All humans possess these assumptions about language, and we use them to generate predictions about the speech we hear in real time.  The examples below illustrate the basic logic underlying contrastive inferences:

  • Give me the book.
  • Give me the large book.

(4) would be an appropriate utterance to use if there is only one book in the physical context or if a specific book has been previously mentioned in the discourse and is thus contextually salient to both the speaker and the listener.  Additional descriptive information about the book in situations like these would be extraneous and unnecessary.  (5) would be a suitable utterance if there are two or more books in the present context and if those books differ with regard to their sizes.  Omitting the adjective “large” would lead to an under-informative sentence and would create a scenario that is ambiguous for the listener.  These examples illustrate the systematic tendency of humans to maximize the efficiency of their utterances while still including adequate detail for their language to be meaningful to others.  A contrastive inference might occur if sentence (5) is uttered in a context with two books that differ in size.  Upon hearing the prenominal adjective “large” but before hearing the word “book,” a listener could deduce that the speaker would only include this detail if it were necessary to differentiate between two items of the same type; thus, the speaker is likely referring to the larger of the two books.

Sedivy et al. (1999) laid the groundwork for experimentally verifying the usage of contrastive inferences in language processing via a visual-world paradigm eye-tracking study.  Subjects were shown a series of visual displays containing four objects.  Some of these displays contained objects that formed a contrastive pair, which consisted of two objects of the same type that differ in a particular way (in this case, the size of the objects).  Audio commands directed the subjects to pick up certain objects in each display, and in some cases these commands contained prenominal scalar adjectives (tall/short, fat/thin, large/small, etc.).  Sedivy and colleagues found that when commands with prenominal adjectives were used, the presence of an intact contrastive pair resulted in more rapid convergence on the target item; subjects took significantly more time to correctly identify the target object when the display did not contain both members of the contrastive pair because contrastive inferences could not be generated in these scenarios.

Figure 1 depicts a display used in Sedivy and colleagues’ work.  The audio command accompanying such a display would be “pick up the tall glass.”  Subjects assigned to the contrast condition for this particular trial would see the display as it is arranged above, and subjects in the no-contrast condition for this trial would see an additional distractor object in the place of the contrast object (the short glass).  Only when the contrast object was present in the visual context did subjects, operating under the assumption that the speaker is striving to avoid over-informativity, tend to assume that the target object was the glass in the upper left corner upon hearing the prenominal adjective “tall.”  It is worth noting that the competitor object—in this case, the pitcher—is a better exemplar of the scalar adjective used (i.e. it is taller than either of the glasses).  Thus, when the contrastive pair was absent and a contrastive inference could not be generated, subjects were forced to rely solely upon semantic information and therefore tended to initially look to the competitor at higher rates than the target upon hearing “tall” (Sedivy et al., 1999).  These findings serve as evidence that humans are capable of incorporating contrastive inferences into their real time processing of language to guide their expectations about a speaker’s utterances.

The speed at which contrastive inferences are generated is quite rapid in comparison to other language processing capacities at the semantics-pragmatics interface—namely, scalar implicatures (refer to sentence (1) for an example of a scalar implicature).  The literature on scalar implicatures presents a somewhat mixed picture regarding their processing speed, but at least in certain contexts these inferences seem to appear only after substantial time-delays that are measureable in eye-tracking experiments (Huang & Snedeker, 2009).  Contrastive inference generation, on the other hand, shows no major lag between the processing of semantic meaning and pragmatic processing (Sedivy et al., 1999).  Thus, contrastive inferences don’t cleanly align with the aforementioned paradigm for categorizing the distinction between semantics (instantaneous processing) and pragmatics (delayed processing).

Two possible accounts could explain the time course discrepancy between these two categories of inferences.  Firstly, it is possible that, even though they have traditionally been categorized as an aspect of pragmatic processing, contrastive inferences should instead be conceived of as an instantaneous semantic process.  This would only be true if contrastive inferences are automatically and rapidly applied to language processing and are not cancelable or susceptible to delay by external factors.  Secondly, it could be the case that Sedivy et al. (1999)’s experiment was simply not sensitive enough to detect delays in processing that may be associated with the contrastive inferences generated by subjects; this study was not specifically designed to compare the timing of pragmatic effects on language processing to the timing of semantic processing, and therefore was not maximally sensitive to the detection of such delays (Huang & Snedeker, 2009).  Cognitive load, which is discussed in the next section, is the experimental tool we ultimately employed to test these two possible accounts of contrastive inference processing.

  1. Working Memory and Cognitive Load

Working memory serves as a useful framework for understanding cognitive processing over a short time interval.  This component of memory temporarily stores and maintains visual and auditory information and allows this perceptual content to interface with long-term memory and information from other cognitive domains (Baddeley, 2003).  The feature of working memory that is relevant to our experimentation is the phonological loop, which can hold a small amount of auditory information in memory for a short period of time.  In order to retain a phonological memory trace for longer than a few seconds, it is necessary for this information to be retrieved and re-articulated using subvocal speech.  The number of independent items that can be stored within the phonological loop is constrained by the real-time nature of this process.  Re-articulation of a phonological item using subvocal speech occurs over a period of time that corresponds to the time that would be required to speak that item aloud.  Therefore, as a person attempts to store more phonological information in the phonological loop, he or she will eventually reach a point at which the first item in the sequence will have faded from working memory before it can be refreshed via the re-articulatory pathway (Baddeley, 2003).

Working memory is relevant to researchers interested in the semantics-pragmatics interface because the processing of information within working memory has been found to compete with elements of pragmatic processing.  While semantic processing is considered instantaneous and requires little cognitive effort, the thought processes driving pragmatic inference-making are believed to impose a substantial drain on cognitive resources.  Thus, working memory offers psycholinguistic researchers a powerful tool to determine which components of meaning are semantic (not dependent on the state of working memory capacity) and pragmatic (dependent on the state of working memory capacity).  Experiments employing working memory manipulations require subjects to simultaneously perform two cognitive tasks:  one task involving the storage of information within working memory in order to induce a cognitive load, and a second task to assess semantic and pragmatic language processing (De Neys & Schaeken, 2007), (Marty, Chemla, & Spector, 2013).



  1. Goals of Research and Potential Outcomes

Our research is the first to apply the cognitive load experimental paradigm to a contrastive inference task in order to yield insight into how best to characterize the processing of these inferences.  We will induce a high degree of cognitive load in some of our subjects and determine whether their processing of contrastive inferences is impaired in relation to subjects with a lower degree of cognitive load.  Broadly speaking, there are two potential outcomes to this research.  The first possibility is that subjects with high cognitive loads will display meaningful deficits in the speed and extent to which they generate contrastive inferences compared to subjects with low cognitive loads.  This would suggest that there is a pragmatic aspect of contrastive interference processing that is disrupted by cognitive load and that contrastive inference generation mirrors the processing scheme of scalar implicature:  an initial semantic-only period followed by enrichment by pragmatic inferences.  The second potential outcome is that cognitive load will not meaningfully impair the processing of contrastive inferences.  This finding would indicate that contrastive inferences are a predominantly semantic process that occurs automatically and with little cognitive effort.




  1. Participants

60 native English speaking adults with normal or corrected-to-normal vision participated in this study.  45 subjects were recruited using the Harvard Psychology Study Pool and received either course credit or payment of $10 as compensation.  15 subjects participated in the study on a volunteer basis.

  1. Materials and Design

Our experiment used a Tobii T-60 eye-tracker, which is capable of measuring real-time eye movements with a great degree of precision by reflecting light off the corneas of participants and detecting the angle at which that light is reflected (while participants view a screen with images of objects).  Eye-tracking experiments like ours operate under the visual-world paradigm, which is the well-attested assumption that the eye movements of participants are reflective of their thoughts and cognitive processing at any given point in time (Tannenhaus et al., 1995).  Eye movements occur extremely frequently and closely correspond to the processing of speech, and consequently they offer a nuanced level of insight into the comprehension of language over time (Huang & Snedeker, 2009).  We designed forty visual displays that were presented on the screen of the Tobii with two auditory commands accompanying each display.  The visual displays closely resembled the stimuli used in Sedivy et al. (1999).  Each display contained four objects arranged on a shelf.  There were sixteen critical displays, each of which had two variants:  a version with an intact contrastive pair and a version without an intact contrastive pair.  Figures 2 and 3 exemplify the two variants of one of the displays used in our study; Figure 2 contains an intact contrastive pair (a large flashlight and a small flashlight), while Figure 3 does not.

The two auditory commands associated with each display were in the form of “click on the ______” and directed the participant to use the computer mouse to select one of the four objects.  For instance, the critical command accompanying the displays in both Figure 2 and Figure 3 was “click on the large flashlight.”  A male native speaker of American English recorded these commands, and he was instructed to minimize stress on the prenominal adjectives (like “large”) while maintaining a natural-sounding prosody.  Notably, since the auditory commands were consistent between the two versions of the display, the semantic information directly encoded in this command was also consistent between the two versions.  This means that any differences in the speed at which participants converged their gazes onto the target objects were attributable to contextual information gleaned from visual input—namely, whether or not an intact contrastive pair was present.

Like Sedivy and colleagues’ 1999 study, each of our critical displays contained a competitor object, which is an additional object that could be described by the prenominal adjective in the command.  In the case of Figures 2 and 3, the competitor object was the paper towel roll, to which the adjective “large” could plausibly also apply.  Upon hearing the adjective “large” in the command “click on the large flashlight” but before hearing the noun “flashlight,” two of the objects remain as potentially viable options—the flashlight and the paper towel roll.  Only by applying a contrastive inference could subjects predict that the speaker is referring to the large flashlight and not the large paper towel roll, as the adjective “large” would be extraneous information in the case of the paper towel roll but not in the case of the flashlight.  This inference is possible only in the case of Figure 2 where an intact contrastive pair is present.  Following the critical auditory command directing participants to click on one of the contrastive pair objects, a second auditory command instructed participants to select a distractor object.  In the case of Figures 2 and 3, the second command was “click on the yo-yo.”  The second auditory command included after the critical command was not relevant for data collection purposes, but rather served to prevent subjects from ascertaining the purpose of the experiment.

In addition to the sixteen critical displays, the study included twenty-four filler displays that were also not included in the data analysis.  Each filler display also had two associated auditory commands; like the second command of the critical display, the filler display commands served to prevent participants from detecting patterns in the displays that could bias their eye movements.  Eight of the filler displays contained contrastive pairs and sixteen of the filler displays consisted of four unrelated objects.  We created two counterbalanced stimuli lists, each of which contained only one variant of a given critical display.  Participants were randomly assigned to one of these two lists.  The first three displays in each list were always filler displays, which enabled subjects to become familiar with the procedure of the study, and the remaining thirty-seven displays appeared in random order.  For a full listing of the objects used to create the critical and filler displays, refer to Appendices A and B.

All participants completed a cognitive load task in conjunction with the contrastive inference task described above.  This task consisted of memorizing one or more letters prior to viewing each of the displays and then typing the letters in reverse order after completing the contrastive inference task.  We randomly assigned subjects to either the low load condition or the high load condition, and an equal number of participants in each condition completed each of the two stimuli lists.  Low load subjects were only required to memorize one letter while completing the contrastive inference task, which produces an extremely minimal strain on working memory capacity.  High load subjects saw strings of letters ranging in size from three to six to be memorized in conjunction each visual display.  The number of letters assigned to each subject was calibrated to their performance on a working memory assessment completed by all participants at the start of the experiment; better performance on this assessment resulted in the assignment of a larger string of letters to memorize before each display.  After completing the contrastive inference task, directions on the screen of the Tobii prompted high load participants to type the string of letters they saw in the reverse order that they appeared.  Requiring subjects to recall the letters in the reverse order that they were presented necessitates that they both encode these numbers in the phonological loop and perform cognitive manipulations to reverse the order of the letters, which further strains available working memory capacity (Baddeley, 2003).  We created the strings of letters presented before each display by randomly selecting letters from the following pool without replacement:  {B, F, H, I, J, L, M, O, Q, R, U, X, Y}.  The letters in this pool are all phonetically distinct from each other, which is relevant because memorizing a similar-sounding set of letters is substantially more difficult than memorizing a dissimilar-sounding set of letters (Conrad & Hull, 1964).  Thus, all strings of letters of the same length chosen from this pool will impose a comparable degree of strain on working memory.

  1. Procedure

The study lasted approximately thirty minutes.  Participants sat in front of the Tobii T-60 eye-tracker and keyboard.  The auditory stimuli played through external speakers.  All participants completed the working memory assessment prior to starting the study; the outcome of this assessment dictated the difficulty of the cognitive load task for subjects in the high load condition.  The researcher used the Tobii Studio software to calibrate the eye-tracker to the particular participant.  Next, subjects saw written instructions on the screen of the Tobii explaining the tasks they would complete, and the researcher gave participants the opportunity to ask questions for further clarification.  The forty visual displays appeared one at a time with a string of letters preceding each display and a page prompting the recall of the letters in reverse order following each display.  In a given trial, subjects first saw the letters flash on the screen for one second each.  They then saw the visual display and the first of two commands directing them to click on one of the objects in the display.  One second after the subject’s first click, the second command began, and again, the subject was prompted to click on another of the objects on the screen.  After the second click, the subjects saw a screen directing them to type the letters that they saw prior to the display in the reverse order that they were presented.  After typing each string of letters, participants received feedback on whether or not they recalled the letter sequence correctly.  The Tobii recorded eye movements at a frequency of sixty Hz from the onset of each command until the subject clicked on the screen.  Figure 4 provides a visual schematic of the presentation of stimuli in this experiment.




We analyzed the proportion of looks to the target object relative to the proportion of looks to the competitor object.  This method of data analysis is in line with prior work on contrastive inferences (Grodner and Sedivy, 2011).  Figures 5 and 6 depict the proportion of looks to the target object minus the proportion of looks to the competitor object over time relative to adjective offset in milliseconds for the low load and high load subjects, respectively.  All times are adjusted by 200 milliseconds in accordance with the time needed to initiate a saccade.  The bars spanning each point in the graph represent 95% confidence intervals.

The vertical lines on the graphs of Figures 5 and 6 at 0 milliseconds and 500 milliseconds correspond to the onset and offset of the critical period.  This critical period reflects where manipulations of discourse contrast have been found in previous contrastive inference studies (Grodner & Sedivy, 2011).  For subsequent analysis of the critical region, we binarized the data and performed a mixed effects logistic regression.  Binarizing the data over the critical region produces a measure of whether the subject spent more time looking at the target object or the competitor object over this time period for a given trial.  We selected this method of data analysis because it is indicative of the constraints of the ocular motor system; realistically, over a 500 millisecond time window, a subject is able to fixate on only a single object in the majority of cases.  The decision to use the mixed effects logistic regression model is reflective of the assumption that the spatial distribution of the eye-tracking data is non-normal—one would expect that a subject would tend to move his or her gaze from one object directly to another as opposed to spending the bulk of the time fixating on points between objects in the display.  Trials in which the subject was looking at neither the contrast object nor the competitor object during the critical region were excluded from analysis.

Analysis of the interaction between the load and contrast variables across the entire 500 millisecond critical region did not yield significant results (z = 1.322, p = .186).  Although this analysis fails to demonstrate a reliable interaction between load and contrast, visual inspection of the data suggests that the possibility of such an interaction cannot be ruled out for the entire critical region.  Consequently, subsequent examination of the data involved dividing the critical region into five 100 millisecond windows to achieve a more fine-grained analysis.  The results of this analysis for the interaction between load and contrast are summarized in Table 1; in line with the prior analyses, the data were binarized for each 100 millisecond time window and a mixed effects linear regression model was used.  Notably, the 200-300 millisecond region displays a significant interaction between the contrast and load variables (z = 2.084, p = .037).




Our research demonstrates that a high cognitive load does disrupt the processing of contrastive inferences.  In the time window of 200-300 milliseconds following the offset of the prenominal adjective, participants in the high load condition demonstrated a substantially lower degree of convergence of their eye movements onto the target objects than did low load participants.  This suggests that high load subjects were less able to utilize contrastive inferences to make predictions about the intentions of the speaker due to the added strain placed on their working memory capacities.  The fact that contrastive inferences are susceptible to impairment by cognitive load indicates that this aspect of language processing is at least partially pragmatic in nature, as semantic processing is immune to cognitive load manipulations.  Our research further demonstrates that contrastive inferences correspond to the conventional scheme of language processing at the semantics-pragmatics interface.  In their immediate processing of the auditory commands, participants seemingly relied solely upon the semantic content of the utterances.  The generation of pragmatic inferences came only after the semantic content of the language had been processed.  High degrees of cognitive load compete with pragmatic processing capabilities for a limited pool of cognitive resources, resulting in less rapid generation of contrastive inferences.

The difference in pragmatic inference-making between the low load and high load groups was only at a significant level for a relatively small window of time.  Why might we expect to only see a significant interaction between the contrast and load effects during the period of 200–300 milliseconds after the offset of the adjective?  It seems that prior to this time window, the effect of contrast, while present, is reasonably small.  The critical period starts at the offset of the adjective, and displaying contrastive inference-making at this stage would require essentially instant processing of the referential context information.  Even subjects in the low load condition were not able to reliably generate contrastive inferences in such a short amount of time.  After this 200-300 milliseconds time window, adequate time had lapsed such that many subjects impeded by the high cognitive load were readily able to generate contrastive inferences and converged their gazes onto the target objects.  In some cases, the onset of the noun had already occurred by the end of the critical period, which provided subjects with information about the identity of the target object beyond just what they are able to glean from pragmatic cues.  Thus, the middle window of 200-300 milliseconds is the “sweet spot” where one would be most likely to observe a strong difference in performance between the low load and high load groups.

This research also provides a novel contribution to the psycholinguistics community with regard to the methodology employed to study contrastive inferences.  All prior work on contrastive inferences relied on eye-tracking schemes in which subjects saw and manipulated physical objects (Sedivy et al., 1999), (Sedivy, 2003), (Grodner & Sedivy, 2011), (Huang & Snedeker, 2013).  No other work to date had indicated that contrastive inference-making could be triggered by virtual representations of objects (as was the case with our Tobii eye-tracking study); in fact, work by a previous researcher suggested that virtual images were incapable of providing individuals with the contextual information necessary to trigger contrastive inferences during real-time language processing (Khan, 2008).  Generating contrastive inferences relies heavily on being able to make rapid and accurate comparisons of the sizes of the different objects in the display.  It makes sense that these comparisons may be more difficult to make with images of items as opposed to real items, but our findings demonstrate that the Tobii eye-tracker coupled with virtual representations of objects constitute a viable modality for the study of contrastive inferences.


Baddeley, A. (2003). Working memory: looking back and looking forward. Nature reviews neuroscience. 4, 829-839.

Carston, R. (1999). The semantics/pragmatics distinction: A view from relevance theory. In Turner, K., The semantics/pragmatics interface from different points of view. 85-126.

Conrad, R., & Hull, A. J. (1964). Information, acoustic confusion and memory span. British Journal of Psychology. 55, 429-432.

De Neys, W., & Schaeken, W. (2007). When people are more logical under cognitive load: Dual task impact on scalar implicature. Experimental psychology. 54, 128-133.

Grodner, D., & Sedivy, J. (2011). The effect of speaker-specific information on pragmatic inferences. The processing and acquisition of reference. 239-272.

Horn, L. (1972). On the semantic properties of logical operators in English. University of California at Los Angeles PhD dissertation.

Huang, Y. & Snedeker, J. (2009). Online interpretation of scalar quantifiers: Insight into the semantics–pragmatics interface. Cognitive Psychology. 58, 376–415.

Huang, Y. & Snedeker, J. (2013). The use of referential context in children’s on-line interpretation of scalar adjectives. Developmental Psychology. 49, 1090-1102.

Khan, M. (2008). Activating prenominal modifiers in contrastive contexts. Unpublished manuscript, Harvard University Department of Psychology.

Marty, P., Chemla, E., & Spector, B. (2013). Interpreting numerals and scalar items under memory load. Lingua. 133, 152-163.

Sedivy, J. (2003). Pragmatic versus form-based accounts of referential contrast: Evidence for effects of informativity expectations. Journal of psycholinguistic research. 32, 3-23.

Sedivy, J., Tannenhaus, M., Chambers, C., & Carlson, G. (1999). Achieving incremental semantic interpretation through contextual representation. Cognition. 71, 109-147.

Tanenhaus, M., Spivey-Knowlton, M., Eberhard, K., & Sedivy, J. (1995). Integration of visual and linguistic information in spoken language comprehension. Science. 268, 1632.



Figures and Tables

Figure 1:  The type of display used in Sedivy et al. (1999); image taken from Grodner and Sedivy (2011).


Figure 2:  A visual display with a contrastive pair present.  The objects in the display are the target (large flashlight), the contrast item (small flashlight), the competitor (paper towel roll), and a distractor (yo-yo).  The critical command accompanying this display was “click on the large flashlight.”


Figure 3:  A visual display without a contrastive pair.  The objects in the display are the target (large flashlight), the competitor (paper towel roll), and two distractors (yo-yo and orange).  The critical command accompanying this display was “click on the large flashlight.”


Figure 4:  Sequence of stimuli presentation in our experiment for a high load condition participant with three letters.


Figure 5:  The proportion of looks to the target object minus the proportion of looks to the competitor object over time relative to adjective offset for the low load condition.  The blue line indicates critical displays with an intact contrastive pair, and the pink line indicates critical displays without an intact contrastive pair.  The vertical lines at 0 and 500 milliseconds denote the critical region.


Figure 6:  The proportion of looks to the target object minus the proportion of looks to the competitor object over time relative to adjective offset for the high load condition.  The blue line indicates critical displays with an intact contrastive pair, and the pink line indicates critical displays without an intact contrastive pair.  The vertical lines at 0 and 500 milliseconds denote the critical region.




Region z-score p-value
0-100 ms .821 .412
100-200 ms 1.046 .296
200-300 ms 2.084 .037
300-400 ms .209 .835
400-500 ms .372 .710

Table 1:  The results of the mixed effects logistic regression model for the interaction between the load and contrast variables across the five 100 millisecond critical time regions.



Appendix A:  Critical Displays

Target Contrast Competitor Distractor 1 Distractor 2
Long pencil Short pencil Long toothbrush Sunglasses Computer mouse
Tall jar Short jar Tall basket Flip-flop Beads
Tall cup Short cup Tall soap dispenser Highlighter Hot glue gun
Small envelope Large envelope Small cell phone Stick Globe
Small whisk Large whisk Small rock Slipper Soda
Short candle Tall candle Short shot glass Notepad Stapler
Large crayon Small crayon Large chalk Shell Drink umbrella
Large hairbrush Small hairbrush Large bowl Playing card Medal
Small deodorant Large deodorant Small candy Camera Power strip
Small post-it Large post-it Small thumb tack Ladle Twine
Large flashlight Small flashlight Large oven mitt Yo-yo Orange
Large binder clip Small binder clip Large lightbulb Lego Battery
Long spoon Short spoon Long hammer Clock Thimble
Small funnel Large funnel Small Chapstick Neck tie Water gun
Thin marker Fat marker Thin paintbrush Tissue box Tennis ball
Small leaf Large leaf Small pocket knife Hair bow Watch

Note:  There were two versions of each critical display.  One version contained the contrast object, and a second version replaced the contrast object with the second distractor object.




Appendix B:  Filler Displays

Object 1 Object 2 Object 3 Object 4
Glass mug Ceramic mug Nail Paper bag
Cloth glove Latex glove Gum Tupperware
Plastic Slinky Metal Slinky Safety goggles Hand sanitizer
Wide tape Narrow tape Key Ziploc bag
Black feather White feather Picture frame Ice cube tray
Red towel Blue towel Match Dollar bill
Short glass Tall glass Teabag Nail clippers
Gray sock White sock Glasses Tweezers
Doll Bracelet Styrofoam ball Baby bottle
Fork Ice cream scoop Mechanical pencil Toy truck
Screwdriver Piggy bank Clothes pin Car air freshener
Paper airplane Tape measure Headphones Wrench
Calculator Hair dryer Swimming goggles Q-tip
Penny Scissors Plastic fork Razor
Straw Nail polish Carabiner Loofa
Wine glass Eraser American flag Pom-pom
Bucket Pizza cutter Duster Paper plate
Pen Toilet paper Tongs Styrofoam cup
Tiara Quarter Stocking Flask
Bottle cap Baseball Lipstick Thermometer
Pencil sharpener Lighter Ping pong ball Shoe
Coffee filter Chopsticks Frisbee Party blower
CD Cork Rubber duck Rope
Mirror Sandal Floss Screw

Note:  The filler displays were not relevant for data analysis.