Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Ann E. Nordmeyer; Michael C. Frank

doi:10.1525/collabra.67931

Nordmeyer, Ann E.; Frank, Michael C.

2023-01-26 00:00:00

Nordmeyer, A. E., & Frank, M. C. (2023). Pragmatic Felicity Facilitates the Production and Comprehension of Negation. Collabra: Psychology, 9(1). https://doi.org/10.1525/collabra.67931 Cognitive Psychology Pragmatic Felicity Facilitates the Production and Comprehension of Negation 1 2 Ann E. Nordmeyer , Michael C. Frank 1 2 Department of Psychology, Southern New Hampshire University, Hooksett, NH, US, Department of Psychology, Stanford University, Stanford, CA, US Keywords: language, psycholinguistics, language comprehension, language production, pragmatics https://doi.org/10.1525/collabra.67931 Collabra: Psychology Vol. 9, Issue 1, 2023 Negation is a fundamental element of language and logical systems, but processing negative sentences can be challenging. Early investigations suggested that this difficulty was due to the representational challenge of adding an additional logical element to a proposition. In more recent work, however, supportive contexts mitigate the processing costs of negation, suggesting that pragmatics can modulate this difficulty. We test the pragmatic hypothesis that listeners’ processing of negation is influenced by expectations about speakers’ production of negation by directly comparing speakers and listeners in two pairs of experiments. In both experiments, speakers produce negative sentences more often when they are both relevant and informative. And in both experiments, listeners in turn are fastest to respond to sentences that they expect speakers to produce. We argue that general pragmatic principles that apply to all sentences can help explain the challenges of processing negation. One explanation is that not all negations are equally Introduction felicitous. On Gricean and neo-Gricean accounts of lan- Language allows us to describe not only the world as we guage use in context, listeners expect speakers to produce see it, but also the world as it is not. Nevertheless, for hu- truthful, informative, and relevant utterances (Grice, 1975; man language users, processing negation is often slow and Horn, 1984; Levinson, 2000; Sperber & Wilson, 1986). effortful. Deciding the truth value of a sentence like “star Many formal theories of pragmatics focus specifically on isn’t above plus” takes much longer than making the same the role of either relevance or informativeness (e.g., Frank decision about a positive sentence (Carpenter & Just, 1975; & Goodman, 2012; Sperber & Wilson, 1986), or conflate the Clark & Chase, 1972; Just & Carpenter, 1971, 1976). Fur- two concepts. In this paper, we attempt to examine the sep- thermore, participants often appear to process the positive arate effect of these pragmatic factors. components of a sentence prior to negating them, suggest- For our purposes in this work, we define an informative ing again that negation is challenging (Ferguson & Sanford, utterance as one that conveys more information about the 2008; Fischler et al., 1983; Hasson & Glucksberg, 2006; referent (i.e., makes the referent easier to identify in con- Kaup et al., 2006; Kaup & Zwaan, 2003; Lüdtke et al., 2008). text; see for a formalization of informativeness, and and Why do adults struggle to process negation despite produc- for a discussion of informativeness as it applies to negative ing negative sentences frequently from a very young age sentences). For example, if we are looking for my car in (Bloom & Capatides, 1993; Pea, 1980)? a parking lot with many types of cars, the utterance “my a Corresponding author: Ann E. Nordmeyer School of Arts and Sciences Southern New Hampshire University 2500 N. River Rd. Hooksett, NH 03106 Phone: (603) 668-2211 x 2058 Email: a.nordmeyer@snhu.edu 1 We define informativeness following the neo-Gricean tradition that is codified in Rational Speech Act models, namely that utterances are informative about an intended meaning (Frank & Goodman, 2012; Goodman & Frank, 2016). For purposes of experimentation and mea- surement, these intended meanings are often operationalized as particular referents within grounded reference games, but nothing in the formal machinery requires reference rather than propositional meaning. One open challenge for these accounts is to generalize this formal pragmatic machinery to be applicable in non-referential cases. Here we rely on these accounts for general theoretical guidance rather than specific numerical predictions, however. Pragmatic Felicity Facilitates the Production and Comprehension of Negation car isn’t a minivan” doesn’t convey as much information The expectations that a listener has about what a as, say, “my car is a convertible”, because “isn’t a minivan” speaker will say are separate from a listeners’ expectations refers to a larger set of cars than “is a convertible”. If we that an event will happen. That is, highly expected out- were in a parking lot almost entirely full of minivans, how- comes are generally uninformative to talk about; for exam- ever, the statement “my car isn’t a minivan” becomes much ple, in the parking lot full of different cars, I expect that more informative, because it helps to uniquely identify the your vehicle is some kind of car but it would be uninfor- car in question in that particular context. mative, and therefore unexpected, to hear the utterance “I We define a relevant feature here as one that addresses drive a car” (even though this might be a relevant response the “question under discussion” (QUD), or topic of dis- to the QUD what type of vehicle do you drive). In contrast, course (Roberts, 2012; Van Rooy, 2003). The QUD can be very unexpected outcomes can be highly informative to talk given implicitly by the context, and negative sentences are about if they occur. For example, a car having an ejector most relevant when the context sets up a polar QUD where seat is very unexpected, so typically you would not expect at least one of the possible answers is a negative response to hear the utterance “my car has an ejector seat”. But if (Xiang et al., 2020). For example, in the parking lot with you are James Bond and you have a sudden need for an ejec- many types of cars, the QUD might be something like what tor seat (where we can assume there might be a spoken or type of vehicle do you drive? where the possible relevant ut- unspoken QUD of does this car have an ejector seat?"), then terances would be a list of different types of vehicle (e.g. the utterance “my car has an ejector seat” is both highly in- sedan, minivan, convertible, etc.), and the utterance “my formative and highly relevant, precisely because of the un- car is not a minivan” would not be relevant. If we are in usual context. the parking lot full of mostly minivans, a reasonable QUD Listeners can find highly unusual information easier to would be is your car a minivan? and the set of relevant ut- process if it is supported by the context; e.g. if a sentence terances would be yes or no. In this latter context both my describes an unusual protagonist who does strange things, car is a minivan and my car is not a minivan are both relevant then listeners are faster to process sentences that describe utterances, but they differ in their informativeness (in fact, unusual events (Rohde et al., 2021), or in a story about the negative utterance is more informative in this specific a dancing peanut singing a song about almond he loves, context, because it more uniquely identifies the referent). describing the peanut character as “in love” invokes less Thus, we take relevance and informativeness to be separate of an N400 response—a marker of semantic processing pragmatic factors, and although they are often aligned (that costs—than describing the peanut as “salted” (Nieuwland is, a cooperative speaker produces utterances that are both & Van Berkum, 2006). In the domain of negation, assuming relevant and informative; Grice, 1975) both can change de- speakers produce true utterances (another of Grice’s max- pending on the context in which an utterance is produced. ims - the maxim of quality), highly expected utterances are A key element of neo-Gricean theories of communica- almost always ones that describe unexpected events be- tion is that listeners expect speakers to produce utterances cause describing a feature that a referent has is almost al- that are informative and relevant. These pragmatic expec- ways more informative and relevant, except when there is tations are present regardless of whether a sentence is af- some violation of the expected state of affairs (e.g. the con- firmative or negative. Context can influence what listeners vertible in the parking lot of mostly minivans). expect to hear, and these expectations influence listening Consistent with this suggestion, presenting negative in- time in turn. These expectations can occur at the word level formation in a supportive context that prepares the listener (Hale, 2001; Levy, 2008), or be influenced by the broader for a violation of expectations can mitigate some of the nonlinguistic context - for example, listeners are faster to processing costs of negation (Glenberg et al., 1999; Wason, process sentences about expected information based on 1965). When a negated feature is explicitly mentioned or both the discourse context and world knowledge (Hald et inferred in preceding sentences (Lüdtke & Kaup, 2006; al., 2007; Lemke et al., 2021). When an utterance appears Orenes et al., 2014), or when negation is presented within to violate expectations, listeners make inferences about the a supportive dialogue (Dale & Duran, 2011), negative sen- speaker’s intended meaning that go beyond the literal tences tend to be processed faster relative to negative sen- meaning of the utterance (Kravtchenko & Demberg, 2022). tences presented without context. And in an ERP exper- Neo-Gricean theories have been used to explain the prag- iment, negations that are expected based on real-world matic inferences that can be drawn from a negative utter- knowledge (e.g., “with proper equipment, scuba-diving ance (Mœschler, 1992, see Tian et al., 2010, 2016 for ex- isn’t very dangerous”) elicited smaller N400 responses than perimental support); for example, if I told you that my car unlicensed negations (e.g., “bulletproof vests aren’t very isn’t a minivan, you might assume that all of the other cars dangerous”; Nieuwland & Kuperberg, 2008). in the parking lot are minivans (because otherwise I would This previous work supports the neo-Gricean idea that a have produced a more informative utterance), or that the listener expects a speaker to produce relevant and informa- QUD is Is your car a minivan? (which would make the utter- tive utterances. These principles are violated when negative ance relevant). When the context rules out this kind of in- sentences are presented without a supportive context, con- terpretation, however, the utterance is simply infelicitous. tributing to negation’s processing cost. Our current exper- Is this kind of pragmatic infelicity generally responsible for iment directly tests two hypotheses. First, speakers tend to the processing cost of negation? produce sentences that are both relevant and informative given the context, and should be less likely to produce a Collabra: Psychology 2 Pragmatic Felicity Facilitates the Production and Comprehension of Negation sentence (negative or affirmative) if the context makes that In the context, at least one character in the set possesses sentence irrelevant or uninformative. Second, expectations target items, so mentioning the presence or absence of the about what speakers would likely say—and their match or target item becomes relevant in this context. We therefore mismatch with what the speaker in fact does say—play an predicted a sharp decrease in the surprisal of producing a important role in the processing costs of negation. To for- true negative utterance and the reaction time to respond to malize this second hypothesis, we are guided by recent a true negative utterance between the and contexts. probabilistic models of language comprehension that de- The informativeness of an utterances changes depending fine a listener’s pragmatic expectations as the probability on how many other characters could be described by that that a speaker would utter a statement in order to convey utterance (e.g., saying a character “has apples” is informa- a particular meaning (Frank & Goodman, 2012; Goodman tive when few other characters have apples, whereas “has & Frank, 2016), and use surprisal, an information-theoretic no apples” is informative when most other characters have measure of expectation-based processing costs (Levy, apples). Mentioning the absence of target items is more in- 2008), to predict processing times. formative in the context compared to the ; however, a negative utterance in the context is still not very infor- The current studies mative in that it describes three of the four characters. We expected that the probability of producing a negative ut- In our experiments, participants viewed sets of four terance would increase as the number of context characters characters who varied in terms of the presence or absence with target items increased. For positive sentences we pre- of a target feature (e.g., boys with or without apples, where dicted the opposite effect of context: In the context, say- apples are the “target item”), and either completed (Exp. 1a ing “has apples” is not very informative, because everyone and 2a) or verified (Exp. 1b and 2b) sentences about one of has apples; in the context, however, “has apples” is very the characters. The critical difference between Experiments informative because the target character is the only person 1 and 2 is that in Experiments 1a and 1b, the characters who has apples. We predicted a corresponding shift in reac- were identical save for the presence/absence of target fea- tion time for listeners, with listeners responding fastest to tures whereas in Experiments 2a and 2b the characters also the sentences that were informative in context. varied in terms of appearance (i.e. hair and shirt color). Varying the presence or absence of the target feature Experiment 1a: Speakers, identical characters created five possible context conditions: , , , , , where the numerator represents the number of characters In Experiment 1a, participants viewed trials where they in the set who e.g., have apples. In Experiments 1a and saw four characters that were all identical except for the 2a (speakers), participants produced written descriptions of presence or absence of identical items (Figure 1). Partici- one of the characters; in Experiment 1a the instructions pants were asked to complete a sentence describing one of asked participants to produce descriptions such that an the characters in such a way that another person would be imaginary listener could identify the referent, whereas Ex- able to identify the character if all of the characters were periment 2a merely asked participants to describe the char- presented in a different order (i.e., not identifying loca- acter without referring to their position in the array. In Ex- tion of the referent). The goal of this study was to measure periments 1b and 2b (listeners) participants evaluated the the probability of speakers producing negations of the form truth value of sentences about the same pictures (see Fig- “[NAME] has no [TARGET ITEM]” while varying both the ure 1). We predicted that speakers would be more likely relevance and informativeness of these utterances based on to mention the presence or absence of the feature (e.g., the visual context. apples) when that feature was relevant and informative. In this first study, we were concerned about whether Furthermore, we predicted that listeners’ processing costs participants would produce negation at all if they had any would be proportional to surprisal. Surprisal is an informa- other way of describing characters, so we kept all other tion-theoretic measure of the amount of information car- character features identical (e.g. hair color and shirt color), ried by an event (in this case, the amount of information effectively making a polar QUD (e.g. does the boy have ap- conveyed by a sentence); in prior work on sentence compre- ples?) more likely, because the presence/absence of objects hension it has been used successfully as a linking hypoth- was the only defining feature of the characters on each trial. esis between production probabilities and reaction times (e.g., Levy, 2008). Method Mentioning the presence or absence of a target feature (e.g. apples) is only relevant when at least one character in Participants the context has apples, setting up a polar QUD (e.g. Does Participants were recruited for Experiment 1a and 1b in boy X have apples?; ). Here, the , , and contexts should tandem, and were randomly assigned to one of the two ex- give rise to QUDs where a true negative sentence such as periments at the start of data collection. We recruited a “X has no apples” is a relevant and felicitous response. In planned sample of 500 participants to participate in these the context, none of the characters have apples, making online experiments through the Amazon Mechanical Turk Does boy X have apples? an unlikely QUD. Thus we expected (mTurk) website. We restricted participation to individuals that very few (if any) speakers would mention the absence in the US and paid 50 cents for this 10 minute study. of target items in the context, and that listeners would be slowest to respond to negative sentences in this context. Collabra: Psychology 3 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 1. An example of a true negative trial with a 0/4 context (left) and a 3/4 context (right). The sentence “Bob has no apples” in the 0/4 context is both uninformative (because the sentence is true of all of the characters) and irrelevant (because apples are not present in the context and therefore the QUD is unlikely to be about apples), whereas the same sentence in the 3/4 context is both informative and relevant. 13 participants were rejected for indicating that they peared with or without target items an equal number of were under 18 after completing the experiment. Another times in each context type. 18 participants were excluded for indicating that their na- Procedure tive language was not English. Of the remaining 470 par- ticipants (across Exp. 1a and Exp. 1b), n = 283 participants Experiment 1a can be viewed at http://anordmey.gith completed Experiment 1a (speakers); of these, 161 were ub.io/negatron/experiments/experiment1/speakers/nega male,120 were female, 2 declined to state gender, and ages tron.html. Presentation and randomization of stimuli and ranged from 18-66+. the collection of reaction time was controlled using javascript and jquery version 1.11.1. Experiment code can Stimuli be found in the supplemental files at https://github.com/a Thirty-two trial items were created in which characters nordmey/negatron. were shown holding either two of the same common, recog- Participants were first presented with a brief overview nizable objects (“target items”; e.g., two apples), or holding screen explaining that they would play a language game. nothing (see Figure 1). Once participants accepted the task, they were randomly Each participant saw trials in which different propor- assigned to Experiment 1a (speakers) or Experiment 1b (lis- tions of characters were holding target items (context con- teners) and saw more detailed instructions which explained dition). These contexts showed , , , , or of the the task and informed them that they could stop at any characters holding objects. The order of characters was time. Participants were told, “First you will see four people. shuffled on each trial, with the referent of the sentence ap- Pay attention to all four people until a red box appears pearing in a random position. Participants saw each image paired with an incomplete around one of the people. When the red box appears, you sentence (e.g. “[NAME] has ______________.”). In half of the will see an incomplete sentence below the pictures. The sentence that you see is about the picture with the box trials, the highlighted character was holding target items around it. Your job is to finish the sentence using only a (“item” trials), and in half of the trials, the highlighted character was holding nothing (“nothing” trials). The ex- few words. You should complete the sentence in a way that periment was fully crossed such that target characters ap- would help someone else identify the character in the red box if they saw the pictures in a different order.” After these 2 At the time that this study was first piloted (in 2014), the pay rate of $3/hour was consistent with other pay for mTurk HITs at the time. Shortly after this study was run (in 2015), our lab policy changed to require studies to pay the equivalent of the US federal minimum wage, which was a rate recommended by many mTurk workers at the time, and has been increased since. Both authors of this study are committed to fair pay for crowd workers. Collabra: Psychology 4 Pragmatic Felicity Facilitates the Production and Comprehension of Negation instructions, participants were shown an example trial which showed them an example of a positive trial, and ex- plained again “Remember, your job is to complete the sen- tence so that another person might be able to identify the character if they saw these pictures in a different order.” Participants saw an array of four pictures on each trial: The target pictures and three context pictures presented in a random horizontal arrangement. Participants looked at these pictures for four seconds, at which point a red box appeared around one of the pictures. One second later, an incomplete sentence appeared, with a textbox for partici- pants to complete the sentence. Data Processing Affirmative responses labeling the target feature were Figure 2. Probability of producing negative sentences coded as “positive” (e.g., “apples,” “two apples,” “red ap- on “nothing” trials (i.e., true negatives) and positive ples,” etc.). Responses negating the target feature (e.g., “no sentences on “item” trials (i.e., true positives) across apples”) were coded as “negative.” All other responses (e.g. different contexts. descriptions of the characters’ clothing or hair color, as Negative sentences are shown in black, and positive sentences in grey. The context is well as other types of positive or negative utterances) were notated by a fraction representing the number of characters in the context who held tar- get items. Error bars show 95% confidence intervals computed by non-parametric boot- coded as “other.” Codes were hand-checked to ensure that strapping. label synonyms or spelling errors were coded correctly. The raw data for this experiment and all experiments reported utterances about characters without target items. There is in this manuscript can be found at https://github.com/ a large jump in the production of true negative utterances anordmey/negatron, which includes the full responses of all between the context, where only 2% of responses were participants in the speaker experiments. We calculated the proportion of positive sentences de- negations of the target item, and the context, where 42% of responses were negative. scribing characters who possessed target items, and the proportion of negative sentences describing characters with To evaluate the reliability of these patterns, we fit two nothing, creating probability distributions for true positive separate binomial mixed-effects models: 1) a model fit only and true negative utterances in each context. We then used to negative utterances describing the absence of a target item (i.e., true negatives), and 2) a model fit only to positive this distribution to calculate the surprisal of hearing a true positive or true negative sentence for each context. Sur- utterances describing the presence of positive items (i.e., prisal (or “self-information” ) for a sentence is defined as true positives). To test the effect of informativeness, we coded context as a continuous variable, e.g. the proportion of characters in the context with target items. To test the Results & Discussion effect of relevance, we created a dummy code to separately test for the effects of the context compared to all of the Figure 2 shows the proportion of participants who pro- other contexts (i.e., the contexts where discussing the pres- duced affirmative sentences to describe the presence of tar- ence or absence of apples is relevant). . get items (true positives), and negation to describe the ab- The first model explored the effect of context on the sence of target items (true negatives). A table of the precise probability of producing a true negative utterance on probabilities displayed in this figure can be found in the “nothing” trials We found a significant positive linear ef- supplemental materials. Participants produced more true fect of context, with the probability of producing a negative positive utterances about the presence of target items than sentence increasing as the proportion of characters with true negative utterances about the absence of target items. target items increases ( , ), indicating a sig- As the number of characters in the context with target nificant effect of informativeness on the production of neg- items increased, however, participants became slightly less ative sentences. We also found that participants were sig- likely to produce true positive utterances about referents nificantly less likely to produce negative utterances in the with target items, and more likely to produce true negative compared to the other context conditions, indicating an 3 All mixed-effects models used the maximal convergent random effects structure (Barr et al., 2013) and were fit using the lme4 package version 1.1-27 in R version 4.1.1. Significance for binomial mixed-effects models was derived in glmer() using the Wald Z test and for lin- ear mixed models was calculated using the standard normal approximation to the t distribution (Barr et al., 2013). 4 The model specification was as follows: negation ~ relevant context + continuous context + (1 | subject) + (1 | item). Collabra: Psychology 5 Pragmatic Felicity Facilitates the Production and Comprehension of Negation The difference between the production of negation in the compared to the demonstrates the impact of “rel- evance”. This pattern is also seen clearly in the surprisal of utterances produced, in Figure 3. The steep gap in surprisal between the and the contexts is driven by the fact that almost no participants produced negation in the context. In this context, because there are no e.g. “apples” in the context, the QUD is unlikely to pertain to apples, and there- fore apples are unlikely to be mentioned by the speaker. In contrast, the context is more likely to give rise to a polar QUD where mentioning the presence or absence of e.g. “ap- ples” is a relevant response. Exploratory Analyses Figure 3. Surprisal for true positive and true negative In our initial analyses we coded utterances such as “Bob sentences across different contexts. has zero apples” and “Bob has nothing” as “other” rather Negative sentences are shown in black, and positive sentences in grey. The context is than “negative” because we wanted to examine speaker ut- notated by a fraction representing the number of characters in the context who held tar- terances that were as close as possible to the utterances get items. Error bars show 95% confidence intervals computed by non-parametric boot- strapping. in the listener condition (Experiment 1b). Furthermore, our discussion of relevance makes specific predictions about the negation of the target item (e.g., a polar QUD licenses effect of relevance ( , ). Comparing this [has apples, has no apples]) and it is less clear what QUDs li- model to an identical model without the relevance factor cense the response has nothing. In an exploratory analysis, indicates that including relevance as a factor in the model however, we found that including other instances of nega- significantly improves model fit ( ). tion in our analysis (e.g., “not apples”, “without apples”, The second model tested the effect of context on the “zero apples”, and “has nothing”) does not change the pat- probability of producing a true positive utterance on “item” tern of results described above for negative sentences (i.e., trials. We found a significant negative linear effect of con- significant effects of relevance, , , and in- text ( , ), indicating that the probability formativeness, , ). of producing a positive utterance decreases as the propor- tion of characters with target items increases. The findings Experiment 1b: Listeners, identical characters of both models together suggest that the probability of pro- ducing both true positive and true negative utterances is in- In Experiment 1b, participants viewed trials that were fluenced by the informativeness of that utterance in con- identical to the trials in Experiment 1a (Figure 1) except text. that instead of being asked to complete sentences, partic- These findings suggest that the production of both nega- ipants were shown a sentence of the form “[NAME] [has/ tive and positive sentences was influenced by the surround- has no] [TARGET ITEM]” and were asked to complete a sen- ing context, although the effect on negative sentences was tence verification task, answering as quickly and accurately more pronounced. This supports our hypothesis that speak- as possible whether the sentence was true of the identified ers will produce informative utterances (i.e. produce sen- referent. tences that are maximally effective at identifying the refer- ent), because true positive utterances about the presence of Method target items are most informative when none of the other characters have target items (e.g. the context; note that Participants participants cannot truthfully describe the presence of tar- Participants were recruited for Experiment 1b at the get items in the context), whereas true negative utter- same time as Experiment 1a. As described in Experiment ances about the absence of target items are most informa- 1a, after excluding participants who reported being under tive when the other characters do have target items (e.g. the age 18 or having a native language other than English, n context; note that participants cannot truthfully describe = 188 remained for analysis in Experiment 1b; of these, 92 the absence of target items in the context). We can also were male and 95 were female, 1 declined to report gender, see the impact of informativeness in Figure 3, which shows and ages ranged from 18-66+. the surprisal of true positive and true negative sentences, with surprisal decreasing for true negative sentences and increasing for true positive utterances as the number of characters with target items increases. 5 The model specification was as follows: positive ~ continuous context + (1 | subject) + (1 | item). Collabra: Psychology 6 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Stimuli Participants saw sets of four characters identical to the images shown in Exp. 1a, except that instead of filling in a sentence to describe a character, participants completed a sentence verification task (see Figure 1). On each trial a sentence of the form “[NAME] [has/has no] [TARGET ITEM]” appeared. Half of the sentences were positive and half were negative (sentence type), and they were paired with pictures such that half were true and half were false (truth value), resulting in four possible trial types (true pos- itive, true negative, false positive, and false negative). Be- Figure 4. Reaction times for each trial type across cause true positive and false negative sentences cannot oc- different conditions. cur in a context (i.e. the referent must have the target Responses to true sentences are shown on the left, and false sentences are shown on the item in these trials), and true negative and false positive right. Negative sentences are shown in black, and positive sentences in grey. The context sentences cannot occur in a context, each trial type oc- is notated by a fraction representing the number of characters in the context who held curred in four possible contexts. The experiment was fully target items. Error bars show 95% confidence intervals computed by non-parametric bootstrap. crossed, with participants receiving eight true positive, eight false positive, eight true negative and eight false neg- ative sentences distributed equally across context types in formed mean, a criterion established in our previous exper- a randomized order over the course of the study. iments (Nordmeyer & Frank, 2014). Procedure Results & Discussion Experiment 1b can be viewed at http://anordmey.github Overall accuracy on this task was very high, with partic- .io/negatron/experiments/experiment1/listeners/negatro ipants responding correctly on 98% of true positive trials, n.html. Presentation and randomization of stimuli and the 92% of true negative trials, 95% of false positive trials, and collection of reaction time was controlled using javascript 96% of false negative trials. Participants’ reaction times can and jquery version 1.11.1. Experiment code can be found be seen in Figure 4. Participants were fastest to respond in the supplemental files at https://github.com/anordmey/ to true positive sentences, and slowest to respond to true negatron. negative sentences in all context conditions. Replicating On each trial, participants saw an array of four pictures previous findings (Clark & Chase, 1972), we also see an for four seconds, at which point a red box appeared around interaction between polarity and truth value, such that par- one of the pictures. One second later, a sentence about that ticipants are slowest to respond to true negatives and false picture appeared. Participants were told to read the sen- positives. Focusing on responses to true sentences (which tence and respond as quickly and accurately as possible can be directly compared to the speaker responses in Ex- with a judgment of whether it was true or false when ap- periment 1a, where participants were asked to produce true plied to the highlighted picture (by pressing either ‘P’ or sentences), listeners’ reaction times to true negative sen- ‘Q’). We used javascript to record reaction times for each tences mirrored the surprisal of true negative sentences trial, measured as the time from when the sentence was in Experiment 1a, with participants responding slowest to presented to the moment when the response was made. true negatives in the context. Participants first saw eight positive sentence practice tri- We fit a linear mixed-effects model to examine the inter- als with feedback about incorrect responses before begin- action between sentence type (positive or negative), truth ning the test trials. The practice trials showed four people value (true or false), and context as predictors of reaction holding objects in various colors, and the target sentence time. All model coefficients are shown in Table 1. In addi- described the color of the referent item in a way that was tion to main effects of sentence type ( , ) accurate or inaccurate. On the test trials, participants saw and truth value ( , ), there was an interac- the same arrays of pictures shown in Experiment 1a. tion between sentence type and truth value such that true positive sentences elicited the fastest responses and true Data Processing negative sentences elicited the slowest responses ( , ). The model showed a significant negative linear We excluded two participants for having an overall accu- effect of context, with reaction times decreasing as the racy below 80%, which left a total of n = 186 participants proportion of characters with target items increased for analysis. At the trial level, we excluded trials with RTs ( , ) and a significant three-way interac- greater than 3 standard deviations from the log-trans- 6 The model specification was as follows: RT ~ sentence × truth × context + (sentence | subject) + (sentence | item). Collabra: Psychology 7 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Table 1. Model output from a mixed-effects model a significant positive relationship between surprisal and re- predicting listeners’ reaction times in response to action time for true negative sentences, , , sentences in different contexts. supporting our prediction that the effects of context on re- action time reflect differences in how speakers would de- Std. scribe the same stimuli. This relationship between surprisal Coefficient err. and reaction time for true negative sentences holds even Intercept 1483 42 35.27 with the outlying context removed from analysis Sentence (Negative) -205 37 -5.51 ( , ). Truth (True) -372 37 -10.00 Experiment 2a: Speakers, varied characters Context -238 43 -5.50 Sentence Truth 692 53 12.99 In Experiment 1 we demonstrated that the probability of Sentence Context 310 61 5.09 speakers (participants in Exp. 1a) producing negations was influenced by both the relevance (i.e. whether the context Truth Context 366 61 6.04 promotes a polar QUD) as well as the informativeness (i.e. Sentence Truth -839 87 -9.68 Context how well a negation identified the referent in context) of these utterances in context. In addition, we demonstrated that the processing time for listeners (participants in Exp. 1b) to respond to negative sentences was highly correlated tion between sentence type, truth value, and context to the surprisal of speakers’ productions of these utter- ( , ). This three-way interaction reflects ances. the fact that both true negatives and false positives in the In Experiment 1 we increased the probability of negation are strikingly slower than any other sentence type in any by only varying the presence and absence of target items context. Both true negative and false positive sentences in across characters. Would participants continue to use nega- the context are using the word e.g. “apples” to describe tion in a more natural context where other features existed a scene in which there are no apples present. Although our to disambiguate the characters, and would the presence focus here is on responses to true sentences, the slow re- of these alternative features influence the effect of rele- action times for both true negative and false positive sen- vance and informativeness? In Experiment 2 we explored tences suggests that listeners expect speakers to describe this question by varying both the shirt and hair color of the relevant features of the context even when the sentence is characters within each trial, in addition to varying the pres- false. ence and absence of target items between trials. In addi- To explore the separate effects of relevance (i.e., the ef- tion, we changed the instructions in Experiment 2a, remov- fect of the context compared to the others) and infor- ing the instruction “complete the sentence in a way that mativeness (i.e. the linear effect of context) on responses would help someone else identify the character” and in- to true utterances, we fit two separate models to reaction stead merely instructing participants to produce sentences times in response to true positive and true negative utter- without referring to the character’s position. Other than ances. We found a significant effect of relevance on reac- these changes, the task in Experiment 2a was identical to tion times for negative utterances, with the context pro- the task in Experiment 1a. ducing significantly slower reaction times compared to the relevant contexts ( , ). We did not find a sig- Method nificant linear effect of context above and beyond the ef- fect of the context ( , ). We did, however, Participants find a significant positive linear effect of context on the the reaction time to respond to positive sentences ( , Participants were recruited for Experiment 2a and 2b in ), indicating a significant effect of informativeness tandem, and were randomly assigned to one of the two ex- on RTs to positive sentences, but not negative sentences. periments at the start of data collection. We recruited a planned sample of 500 participants to participate in these Comparing Speakers and Listeners across online experiments through the Amazon Mechanical Turk Experiment 1 (mTurk) website. We restricted participation to individuals in the US and paid 50 cents for this 10 minute study. 5 par- To test our hypothesis that processing times are a func- ticipants were rejected for indicating that they were under tion of listeners’ expectations about what a speaker will say, 18 after completing the experiment, and an additional 16 we regressed the mean reaction time in response to true participants were excluded for indicating that their native positive and negative utterances in each condition against language was not English. Of the remaining 479 partici- the surprisal for the same utterances (Figure 5). There was 7 Model specification for true negative model: rt ~ relevance context + numeric context + (relevant context + numeric context | subject) + (relevant context + numeric context | item); model specification for true positive model: rt ~ nu- meric context + (numeric context | subject) + (numeric context | item). Collabra: Psychology 8 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 5. Reaction times in the listener condition plotted by surprisal in the speaker condition. Each point represents a measurement for sentence type and context. Negative sentences are shown in black, and positive sentences in grey. Error bars on the horizontal and vertical axes represent 95% confidence intervals on their respective measures. The gray band shows the linear regression and 95% confidence region for all conditions. The inset graph zooms in on the linear regression and 95% confidence region for data excluding the outlying 0/4 condition. pants (across Exp 2a and Exp 2b), n = 233 completed Exp. that you see is about the picture with the box around it. 2a (speakers); of these, 111 were male and 122 were female, Your job is to finish the sentence using only a few words. and ages ranged from 18-66+. Please avoid using descriptions of the person’s position rel- ative to other characters.” After these instructions, partici- Stimuli pants were shown the same example trial as in Experiment 1a, and told again "Please avoid using descriptions of the Stimuli were identical to the stimuli in Exp. 1a except person’s position relative to other characters. So, for exam- that within each trial characters’ shirt and hair colors also ple, do not complete the sentence by saying “second to the varied, providing other referential possibilities for speakers right”. (see Figure 6). Data Processing Procedure We followed the same coding procedure in Exp. 2a as we Experiment 2a can be viewed at http://anordmey.githu did in Exp. 1a. We calculated the proportion of positive sen- b.io/negatron/experimnts/experiment2/speakers/nega tences describing characters who possessed target items, tron.html. Presentation and randomization of stimuli and and the proportion of negative sentences describing char- the collection of reaction time was controlled using acters with nothing, creating probability distributions for javascript and jquery version 1.11.1. Experiment code can true positive and true negative utterances in each context be found in the supplemental files at https://github.com/ and using these distributions to calculate the surprisal of anordmey/negatron. hearing a true positive or true negative sentence for each Although the task was identical to the “complete the context. sentence” task in Experiment 1a, we altered the instruc- tions slightly to emphasize that participants should not re- Results & Discussion fer to the position of the character in the array, instead of emphasizing the ability for a listener to identify the ref- Figure 7 shows the proportion of participants who pro- erent. This change was made to see if participants would duced affirmative sentences to describe the presence of tar- spontaneously produce relevant and informative utterances get items (true positives), and negation to describe the ab- even without being told to imagine a potential listener. Par- sence of target items (true negatives). A table of the precise ticipants were told, “First you will see four people. Pay at- probabilities displayed in this table can be found in the tention to all four people until a red box appears around supplemental materials. As in Experiment 1a, the probabil- one of the people. When the red box appears, you will see ity of producing true negative utterances increases and the an incomplete sentence below the pictures. The sentence Collabra: Psychology 9 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 6. An example of a true negative trial with a 0/4 context (left) and a 3/4 context (right). The sentence “Bob has no apples” in the 0/4 context is both uninformative (because the sentence is true of all of the characters) and irrelevant (because apples are not present in the context and therefore the QUD is unlikely to be about apples), whereas the same sentence in the 3/4 context is both informative and relevant. a model fit only to positive utterances. To test the effect of informativeness, we coded context as continuous, e.g. the proportion of characters in the context with target items. To test the effect of relevance, we created a dummy code to separately test for the effects of the context compared to all of the other contexts (i.e., the contexts where discussing the presence or absence of apples is relevant). For the first model described above, we tested the effect of context on the probability of producing a true negative utterance on “nothing” trials. We found a significant effect of the relevant contexts, indicating that participants were less likely to produce negative sentences in the context compared to the other contexts ( , ). This finding suggests that relevance has a significant effect on the production of true negative sentences. We also found a Figure 7. Probability of producing negative sentences significant positive linear effect of context, with the prob- on “nothing” trials (i.e., true negatives) and positive ability of producing a negative sentence increasing as the sentences on “item” trials (i.e., true positives) across proportion of characters with target items increases different contexts. ( , ), indicating a significant effect of infor- Negative sentences are shown in black, and positive sentences in grey. The context is mativeness on the production of negative sentences. Com- notated by a fraction representing the number of characters in the context who held tar- paring this model to an identical model without the rele- get items. Error bars show 95% confidence intervals computed by non-parametric boot- vance factor indicates that including relevance as a factor strapping. in the model significantly improves model fit ( ). probability of producing true positive utterances decreases The second model tested the effect of context on the as the number of characters with target items increases. probability of producing a true positive utterance on “item” As in Exp. 1a, we fit two separate binomial mixed-effects trials. We found a significant negative linear effect of con- models: 1) a model fit only to negative utterances, and 2) text ( , ), indicating that the probability of 8 The model specification was as follows: negation ~ relevant context + continuous context + (1 | subject) + (1 | item). 9 The model specification was as follows: positive ~ continuous context + (1 | subject) + (1 | item). Collabra: Psychology 10 Pragmatic Felicity Facilitates the Production and Comprehension of Negation apples” was a very rare production on trials where there weren’t any apples in the context; participants in this con- dition tended to produce affirmative sentences describing other features of the referent, such as their clothing or hair color. Participants only consistently produced negative ut- terances in the context, where at least one other charac- ter has target items, thus making does boy X have apples? a likely QUD. Although the pattern of negation production looks different between Experiment 1 and Experiment 2 (as described in the paragraph above), when we look at speaker surprisal, the two experiments show the same pattern. In both Experiment 1a and Experiment 2a, the surprisal of producing a negative utterance is significantly higher in the than in any other context because the probability of pro- ducing a negative utterance in this context (i.e., mention- Figure 8. Surprisal for true positive and true negative ing the absence of an item that is seen nowhere in the con- sentences across different contexts. text) is close to zero in both experiments. Negative sentences are shown in black, and positive sentences in grey. The context is notated by a fraction representing the number of characters in the context who held tar- Exploratory Analyses get items. Error bars show 95% confidence intervals computed by non-parametric boot- strapping. In Experiment 2a, we used the same coding scheme as in Experiment 1a, coding utterances such as “Bob has noth- producing a positive utterance decreases as the proportion ing” as “other” rather than “negative”. Once again we ex- of characters with target items increases. The findings of amined how these coding decisions influenced our results both models together suggest that the probability of pro- in Experiment 2a. Consistent with our predictions, if we ducing both true positive and true negative utterances is in- include other negations of the target item in our analysis fluenced by the informativeness of that utterance in con- (e.g., “not apples”, “without apples”, “zero apples”, etc.), text. the results of the speaker condition are the same as those A key difference between Experiment 1a and Experiment reported above (i.e., significant effects of relevance and in- 2a is the relative informativeness of describing the presence formativeness). If we also include instances of “nothing”, or absence of the target item, as opposed to describing e.g. however, the effect of informativeness remains highly sig- shirt or hair color. In Experiment 1a, where all of the char- nificant ( ), but the effect of relevance is acters were identical, descriptions of clothing color were al- no longer significant ( ). That is, “has no ways relatively uninformative (because they were true of all apples” is not a very relevant utterance in a context where characters in a given trial). That is, in Experiment 1a, neg- no one else has apples, but “has nothing” is, perhaps, a rel- ative utterances on “nothing” trials were therefore at least evant thing to say in the context of an experiment where as informative as describing shirt color in the , , and people sometimes have objects and sometimes do not. contexts. Contrast this with Experiment 2a, where de- In the context, producing the utterance has no apples scribing the color of a characters shirt is always an infor- is not relevant because no one has apples and therefore the mative utterance: the informativeness of a true negative ut- QUD is unlikely to be related to the presence or absence terance is only comparable to these competing utterances of apples. What is the QUD in this context? An analysis of in the context. The effect of the informativeness of these all speaker responses from Experiment 2b, shown in Figure competing utterances in the current study is reflected in 9, can help illuminate the likely QUD in different contexts. the raw probabilities for the speaker data (Figure 7), where In the context when the referent has nothing, speakers there is a jump in the probability of a true negative ut- overwhelmingly described color (89% of responses) - usu- terance between the and the context. In Experiment ally hair or shirt color, which varied across all characters 1a, this sharp increase occurred between the and con- and was therefore highly informative. Participants become texts – again, the point where producing a true positive ut- less likely to describe hair or shirt color (and more likely to terance becomes as informative as producing an utterance produce negation) as the absence of target items becomes about e.g., shirt color. Overall, negation was produced less more informative. The opposite pattern can be seen on tri- frequently in Experiment 2a (17% true negative utterances als where the referent is holding target items; here describ- across all context conditions) compared to Experiment 1a ing color becomes more common as the number of target (38% true negative utterances across all contexts), likely items increases (and describing the presence of target items due in part to difference in the relative informativeness of becomes less informative). That is, as describing the pres- negative utterances in Experiment 2. ence or absence of target items becomes less informative, As in Experiment 1a, and consistent with our hypothesis describing shirt or hair color becomes increasingly likely. that speakers will produce negative sentences that are rele- This pattern of responses suggests that some proportion of vant (i.e. producing an utterance that is a possible response speakers, most commonly in the condition, interpreted to a likely QUD), negative sentences were almost never pro- the QUD as something like What color shirt is the person duced on “nothing” trials in the context. That is, “no Collabra: Psychology 11 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 9. Categories of different types of speaker responses across the different context conditions. Purple bars (“Color” responses) represent the proportion of participants who relied on descriptions of color (i.e. hair or shirt color) to describe the referent. Dark blue bars (the “has [target item]” responses) on the “referent has item” trials represent affirmative descriptions of the target item (i.e. the “true positive” responses in Figure 7 and teal bars (the “has no [target item]” responses) on the “referent has nothing” trials represent negative descriptions of the absence of target items (i.e. the “true negative” responses in Figure 7). wearing or even What does this person look like, which would instead of referring to the color of items, practice trials sim- license more general descriptions of the referent. ply showed photographs of common household items, and sentences took the form e.g. “This [is/is not] a spoon”. This Experiment 2b: Listeners, varied characters allowed us to give participants practice with the sentence verification task without priming them to think about the In Experiment 2b, participants viewed trials that were color of items (which was now a relevant feature of the identical to the trials in Experiment 2a (i.e. characters with characters in Experiment 2). On the test trials, participants varied shirt and hair colors), with the same task and sen- saw the same arrays of pictures shown in Experiment 2a. tence prompts as those in Experiment 1b (i.e. a sentence verification task). Data Processing Method We excluded two additional participants for having an overall accuracy below 80%, which left a total of n = 244 Participants participants for analysis. At the trial level, we excluded tri- als with RTs greater than 3 standard deviations from the Participants were recruited for Exp. 2b at the same time log-transformed mean, a criterion established in our previ- as Exp. 2a. N = 246 participants were randomly assigned to ous experiments (Nordmeyer & Frank, 2014). complete Exp. 2b; of these, 120 were male and 124 were female, 2 declined to report gender, and ages ranged from Results & Discussion 18-66+. Overall accuracy on this task was again very high, with Stimuli participants responding correctly on 98% of true positive trials, 93% of true negative trials, 96% of false positive tri- Participants saw the same set of images as participants als, and 97% of false negative trials. Reaction times for Ex- in Exp. 2a, with the same sentences and instructions as in periment 2b can be seen in Figure 10. As in Experiment Exp. 1b (i.e. a sentence verification task; see Figure 6). 1b, we again see main effects of sentence type, with partic- ipants responding faster to positive sentences than nega- Procedure tive sentences, and an interaction between truth value and polarity. We also see again that the context yields much Experiment 2b can be viewed at http://anordmey.git slower reaction times than any other context, suggesting hub.io/negatron/experiments/experiment2/listeners/nega again a striking effect of relevance on both true negative tron.html. Presentation and randomization of stimuli and and false positive sentences. the collection of reaction time was controlled using We fit a linear mixed-effects model to examine the inter- javascript and jquery version 1.11.1. Experiment code can action between sentence type (positive or negative), truth be found in the supplemental files at https://github.com/ value (true or false), and context as predictors of reaction anordmey/negatron. time. All model coefficients are shown in Table 2. In addi- The instructions and procedure were identical to Experi- tion to main effects of sentence type ( , ) ment 1b, except that the practice trials were altered so that and truth value ( , ), there was an inter- Collabra: Psychology 12 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 10. Reaction times for each trial type across different conditions. Responses to true sentences are shown on the left, and false sentences are shown on the right. Negative sentences are shown in black, and positive sentences in grey. The context is notated by a fraction representing the number of characters in the context who held target items. Error bars show 95% confidence intervals computed by non-parametric bootstrap. Table 2. Model output from a mixed-effects model to true utterances, we fit two separate models to reaction predicting listeners’ reaction times in response to times in response to true positive and true negative utter- sentences in different contexts. ances. We found a significant effect of relevance on reac- tion times for negative utterances, with the context pro- Std. ducing significantly slower reaction times compared to the Coefficient err. relevant contexts ( , ). We did not find a sig- Intercept 1598 41 39.16 nificant linear effect of context above and beyond the ef- Sentence (Negative) -205 38 -5.36 fect of the context ( , ). We did, however, find a significant positive linear effect of context on the Truth (True) -384 36 -10.57 the reaction time to respond to positive sentences ( , Context -341 42 -8.14 ), indicating a significant effect of informativeness Sentence Truth 663 54 12.16 on RTs to positive sentences, but not negative sentences. Sentence Context 377 59 6.40 Truth Context 454 59 7.71 Comparing Speakers and Listeners across Sentence Truth Experiment 2 -901 83 -10.79 Context As in Experiment 1, we regressed the mean reaction time in response to true positive and negative utterances in each condition against the surprisal for the same utterances in action such that true positive sentences elicited the fastest order to test our hypothesis that processing times are a responses and true negative sentences elicited the slowest function of listeners’ expectations about what a speaker responses ( , ). The model showed a signif- will say (Figure 11). There was a significant positive rela- icant negative linear effect of context, with reaction times tionship between surprisal and reaction time for true neg- decreasing as the proportion of characters with target items ative sentences, , , supporting our predic- increased ( , ). A significant three-way in- tion that the effects of context on reaction time reflect teraction between sentence type, truth value, and context differences in how speakers would describe the same stim- indicates that this pattern was driven primarily by re- uli. This relationship between surprisal and reaction time sponses to true negative sentences, with context having for true negative sentences holds even with the outlying the most pronounced effect on true negative utterances context removed from analysis ( , ). ( , ). To explore the separate effects of relevance (i.e., the ef- fect of the context compared to the others) and infor- mativeness (i.e. the linear effect of context) on responses 10 The model specification was as follows: RT ~ sentence × truth × context + (sentence | subject) + (sentence | item). 11 Model specification for true negative model: rt ~ relevance context + numeric context + (relevant context + numeric context | subject) + (relevant context + numeric context | item); model specification for true positive model: rt ~ nu- meric context + (numeric context | subject) + (numeric context | item). Collabra: Psychology 13 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Figure 11. Reaction times in the listener condition plotted by surprisal in the speaker condition. Each point represents a measurement for sentence type and context. Negative sentences are shown in black, and positive sentences in grey. Error bars on the horizontal and vertical axes represent 95% confidence intervals on their respective measures. The gray band shows the linear regression and 95% confidence region for all conditions. The inset graph zooms in on the linear regression and 95% confidence region for data excluding the outlying 0/4 condition. likely QUD, and were more likely to mention the presence General Discussion or absence of features when mentioning those features What makes negation hard to process? Our work sug- would help uniquely identify the referent. The fact that in- gests that the same general pragmatic mechanisms that formativeness played a role in the production of both pos- govern positive sentences are responsible for much of the itive and negative utterances supports our argument that difficulty associated with negative sentences. Negative sen- these pragmatic pressures are general, rather than specific tences presented without context are uninformative and to negation. not relevant; thus, they are unlikely to be produced by Past work suggests that the context in which a negative speakers. In turn, listeners respond to unlikely utterances sentence occurs can influence the QUD inferred by partic- with increased processing times. In contexts where nega- ipants (Tian et al., 2010; Xiang et al., 2020). In our exper- tion is relevant and more informative, processing costs are iment, speakers in both experiments were most likely to lower. Overall, this evidence supports a Gricean interpreta- produce true negatives in the , , and contexts. This is tion of negation processing, with pragmatic principles play- consistent with previous work suggesting that these types ing a role in the processing of both positive and negative of contexts give rise to a polar QUD (Xiang et al., 2020). The sentences. structure of the experiment may have also played a role in While previous work has shown that contextual factors QUD inference; e.g., our prompt for speakers asked them facilitate the processing of negation (Dale & Duran, 2011; to fill in the sentence “[NAME] has ______________.” which Nieuwland & Kuperberg, 2008; Orenes et al., 2014; Wason, may have biased speakers towards a polar QUD (e.g. Does 1965), our findings here go further. First, by using actual [NAME] have [ITEM]?). The within-subjects structure may language productions as the predictor of processing dif- also have contributed to this, as participants may have in- ficulty, our work implicates specifically pragmatic factors. ferred the polar QUD due to the presence/absence of items Because the field of pragmatics is concerned with language varying across different trials. The question of how a dis- use, demonstrating a relationship between actual speaker course participant uses context to infer the QUD is a topic productions and listener processing time is critical to the for future research. argument that pragmatic factors are responsible for the Listeners’ reaction times were highly correlated with the processing cost of negation. Second, rather than treating surprisal of a speaker producing the same utterances in pragmatics as a black box, we show that two different com- context in both Experiment 1b and Experiment 2b. For true ponents—informativeness and relevance—each contribute negative sentences, however, this effect was driven primar- to the relative (un-)likelihood of hearing a negation. Speak- ily by the effect of relevance (i.e., the difference between ers in both experiments were unlikely to mention the ab- the and all other contexts). Why didn’t informativeness sence of a feature unless it was a relevant response to a play a role in listeners’ response times? One possibility is Collabra: Psychology 14 Pragmatic Felicity Facilitates the Production and Comprehension of Negation that only relevance plays a role in forming listeners’ expec- In the domain of negation, a number of theories have tations about what a speaker will say. Another possibility is been proposed to explain how negation is represented and that the effect of informativeness is small compared to the processed at the mechanistic level. Two-step theories of effect of relevance, and that the reaction time measure is negation processing have argued that negation is processed too noisy to identify this effect. Although our results can- by first representing the negated state of affairs, and then not disentangle these two possibilities, the fact that we did rejecting this in favor of the actual state of affairs; these find a significant effect of informativeness on listeners’ re- representations have been argued by some to be proposi- sponses to true positive sentences suggests that informa- tional in nature (Clark & Chase, 1972) or simulations of the tiveness does play some role in listeners’ expectations. true and negated state of affairs (Dudschig & Kaup, 2020; One possible explanation for the difference in reaction Kaup et al., 2006, 2007; Kaup & Zwaan, 2003). These theo- time between the context and the other contexts is that ries can explain why, in many tests of negation processing – including ours – there is an interaction between the polar- in the contexts where apples are present, the visual pres- ity and truth value of a sentence. Other theories argue that ence of apples primes the semantic processing of sentences with the word “apples” (see e.g.,). In our experiments, we under the right experimental and pragmatic conditions, the cannot disentangle the possible lower-level effect of this true state of affairs is represented and accessible to com- prehenders immediately, suggesting that it is possible for type of lexical activation from the facilitation that comes negation to be processed fluently in a single step, especially from listener expectations about what a speaker might say. What it means for a word or sentence to be facilitated by in supportive contexts (as is also reflected in the contexts “predictability” likely occurs at multiple levels of represen- effects seen in our data, though we note that even in our most supportive contexts, negation was processed slower; tation. That is, the lower-level semantic priming that might ). Although this body of literature is highly complementary occur upon seeing a picture of an apple might facilitate the higher-level expectations that a listener makes about a to the data we have presented here, our studies here are speaker utterance (see for a discussion of prediction at mul- not designed to test between competing theories about the mechanism by which negation is represented. tiple levels of representation). Although we focus here on negation, our findings have Much like other psycholinguistic experiments, our tasks take place in a restricted world where the set of possible implications for sentence processing more generally. De- referents (and by extension, the set of possible utterances bates about the effects of pragmatics on linguistic process- ing exist in other domains, such as the processing of scalar that a speaker is likely to produce) is limited. In real world implicatures (the pragmatic inference that e.g., “some” im- contexts, a speaker could produce many different utter- ances to describe many different referents at any given mo- plicates “some but not all”; Grodner et al., 2010; Huang ment. Our goal in these experiments is to explore whether & Snedeker, 2009, 2011). provide an informative compar- ison between scalar implicature and negation, presenting Neo-Gricean pragmatic principles can at minimum explain mouse-tracking trajectories for each. Their negation data how people communicate in these restricted worlds. Our theory would predict that when people produce negative show the same pattern of processing difficulties we observe, and critically, their data on the processing of underinfor- utterances in the “real world” they are usually producing mative “some” utterances look almost identical. We hy- them in contexts where those utterances are highly infor- pothesize that, in both cases, participants’ processing dif- mative and relevant; thus, we would predict that people have less difficulty processing negative sentences in natural ficulty is a function of the violation of their pragmatic expectations about what speakers will say. conversation than they do in constrained psycholinguistic Our findings here suggest that the processing difficulties experiments. of negative sentences arise at least in part from the relative Our analysis is intended as a computational level analy- sis (Marr, 1982). We are not committed to a specific account pragmatic felicity of negation in context. Neo-Gricean prin- ciples of informativeness and relevance impact the proba- of how negative sentences or pragmatic inferences are rep- bility of producing negative sentences in different contexts, resented. Our results don’t tell us whether listeners actu- ally simulate a speaker or a specific QUD when they form and listeners expect speakers to abide by these principles. expectations about what a speaker will say, or are doing These findings demonstrate that when logical words are used in a communicative context, we often have no diffi- something that approximates such a simulation. Our goal culty understanding them. is to show at least that the weaker of these two possibilities is true; that is, listeners are forming expectations about what a speaker will say, even though we don’t know pre- cisely what form these representations take. Several mech- Author Contributions anistic theories of pragmatics have been proposed, such as an in-the-moment alignment of linguistic representations Both authors developed the study concept and con- between speakers and listeners (Pickering & Garrod, 2004), tributed to the study design. Data collection was conducted or cached expectations about how speakers use language by AEN. AEN performed the data analysis and interpre- leading to preferred interpretations by listeners (Levinson, tation under the supervision of MCF. Both authors con- 2000). Either of these possibilities, or any number of others, tributed to the development of the manuscript and ap- could be consistent with our data. proved the final version of the manuscript for submission. Collabra: Psychology 15 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Funding Information Data Accessibility Statement This work supported by the NSF GRFP to AEN and ONR All the stimuli, experiment code, raw data, and analysis N00014-13-1-0287. scripts described here can be found in the supplemental materials on this paper’s project page at Competing Interests https://github.com/anordmey/negatron. The authors declare no competing interests. Submitted: August 12, 2021 PST, Accepted: December 20, 2022 PST This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International License (CCBY-4.0). View this license’s legal deed at http://creativecommons.org/licenses/by/4.0 and legal code at http://creativecom- mons.org/licenses/by/4.0/legalcode for more information. Collabra: Psychology 16 Pragmatic Felicity Facilitates the Production and Comprehension of Negation References Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Hald, L. A., Steenbeek-Planting, E. G., & Hagoort, P. Random effects structure for confirmatory hypothesis (2007). The interaction of discourse context and testing: Keep it maximal. Journal of Memory and world knowledge in online sentence comprehension. Language, 68(3), 255–278. https://doi.org/10.1016/j.j Evidence from the N400. Brain Research, 1146, ml.2012.11.001 210–218. https://doi.org/10.1016/j.brainres.2007.02.0 Bloom, L., & Capatides, J. B. (1993). Language development from two to three. Cambridge University Hale, J. (2001). A probabilistic earley parser as a Press. psycholinguistic model. Second Meeting of the North American Chapter of the Association for Computational Carpenter, P. A., & Just, M. A. (1975). Sentence Linguistics on Language Technologies 2001 - NAACL comprehension: A psycholinguistic processing model ’01. Second meeting of the North American Chapter of verification. Psychological Review, 82(1), 45–73. htt of the Association for Computational Linguistics. htt ps://doi.org/10.1037/h0076248 ps://doi.org/10.3115/1073336.1073357 Clark, H. H., & Chase, W. G. (1972). On the process of Hasson, U., & Glucksberg, S. (2006). Does comparing sentences against pictures. Cognitive understanding negation entail affirmation? Journal of Psychology, 3(3), 472–517. https://doi.org/10.1016/00 Pragmatics, 38(7), 1015–1032. https://doi.org/10.101 10-0285(72)90019-9 6/j.pragma.2005.12.005 Dale, R., & Duran, N. D. (2011). The cognitive dynamics Horn, L. R. (1984). Toward a new taxonomy for of negated sentence verification. Cognitive Science, pragmatic inference: Q-based and R-based 35(5), 983–996. https://doi.org/10.1111/j.1551-6709.2 implicature. In D. Schiffrin (Ed.), Meaning, form, and 010.01164.x use in context: Linguistic applications (pp. 11–42). Dudschig, C., & Kaup, B. (2020). Negation as conflict: Georgetown University Press. Conflict adaptation following negating vertical Huang, Y. T., & Snedeker, J. (2009). Online spatial words. Brain and Language, 210, 104842. http interpretation of scalar quantifiers: Insight into the s://doi.org/10.1016/j.bandl.2020.104842 semantics–pragmatics interface. Cognitive Psychology, Ferguson, H. J., & Sanford, A. J. (2008). Anomalies in 58(3), 376–415. https://doi.org/10.1016/j.cogpsych.20 real and counterfactual worlds: An eye-movement 08.09.001 investigation. Journal of Memory and Language, 58(3), Huang, Y. T., & Snedeker, J. (2011). Logic and 609–626. https://doi.org/10.1016/j.jml.2007.06.007 conversationrevisited: Evidence for a division between Fischler, I., Bloom, P. A., Childers, D. G., Roucos, S. E., semantic and pragmatic content in real-time & Perry, N. W. (1983). Brain potentials related to language comprehension. Language and Cognitive stages of sentence verification. Psychophysiology, Processes, 26(8), 1161–1172. https://doi.org/10.1080/ 20(4), 400–409. https://doi.org/10.1111/j.1469-8986.1 01690965.2010.508641 983.tb00920.x Just, M. A., & Carpenter, P. A. (1971). Comprehension of Frank, M. C., & Goodman, N. D. (2012). Predicting negation with quantification. Journal of Verbal Pragmatic Reasoning in Language Games. Science, Learning and Verbal Behavior, 10(3), 244–253. http 336(6084), 998–998. https://doi.org/10.1126/scienc s://doi.org/10.1016/s0022-5371(71)80051-8 e.1218633 Just, M. A., & Carpenter, P. A. (1976). Eye fixations and Glenberg, A. M., Robertson, D. A., Jansen, J. L., & cognitive processes. Cognitive Psychology, 8(4), Johnson-Glenberg, M. C. (1999). Not propositions. 441–480. https://doi.org/10.1016/0010-0285(76)9001 Cognitive Systems Research, 1(1), 19–33. https://doi.or 5-3 g/10.1016/s1389-0417(99)00004-2 Kaup, B., Lüdtke, J., & Zwaan, R. A. (2006). Processing Goodman, N. D., & Frank, M. C. (2016). Pragmatic negated sentences with contradictory predicates: Is a language interpretation as probabilistic inference. door that is not open mentally closed? Journal of Trends in Cognitive Sciences, 20(11), 818–829. http Pragmatics, 38(7), 1033–1050. https://doi.org/10.101 s://doi.org/10.1016/j.tics.2016.08.005 6/j.pragma.2005.09.012 Grice, H. P. (1975). Logic and conversation. Speech Acts, Kaup, B., Yaxley, R. H., Madden, C. J., Zwaan, R. A., & 41–58. https://doi.org/10.1163/9789004368811_003 Lüdtke, J. (2007). Experiential simulations of negated Grodner, D. J., Klein, N. M., Carbary, K. M., & text information. Quarterly Journal of Experimental Tanenhaus, M. K. (2010). “Some,” and possibly all, Psychology, 60(7), 976–990. https://doi.org/10.1080/1 scalar inferences are not delayed: Evidence for immediate pragmatic enrichment. Cognition, 116(1), Kaup, B., & Zwaan, R. A. (2003). Effects of negation and 42–55. https://doi.org/10.1016/j.cognition.2010.03.01 situational presence on the accessibility of text information. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29(3), 439–446. http s://doi.org/10.1037/0278-7393.29.3.439 Collabra: Psychology 17 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Kravtchenko, E., & Demberg, V. (2022). Informationally Orenes, I., Beltrán, D., & Santamaría, C. (2014). How redundant utterances elicit pragmatic inferences. negation is understood: Evidence from the visual Cognition, 225, 105159. https://doi.org/10.1016/j.cogn world paradigm. Journal of Memory and Language, 74, ition.2022.105159 36–45. https://doi.org/10.1016/j.jml.2014.04.001 Lemke, R., Schäfer, L., & Reich, I. (2021). Modeling the Pea, R. D. (1980). The development of negation in early predictive potential of extralinguistic context with child language. In D. R. Olson (Ed.), The Social script knowledge: The case of fragments. PLOS ONE, Foundations of Language and Thought (pp. 156–186). 16(2), e0246255. https://doi.org/10.1371/journal.pon Norton New York. e.0246255 Pickering, M. J., & Garrod, S. (2004). Toward a Levinson, S. C. (2000). Presumptive meanings: The theory mechanistic psychology of dialogue. Behavioral and of generalized conversational implicature. The MIT Brain Sciences, 27(02), 169–190. https://doi.org/10.10 Press. https://doi.org/10.7551/mitpress/5526.001.000 17/s0140525x04000056 Roberts, C. (2012). Information structure in discourse: Levy, R. (2008). Expectation-based syntactic Towards an integrated formal theory of pragmatics. comprehension. Cognition, 106(3), 1126–1177. http Semantics and Pragmatics, 5, 6–1. https://doi.org/10.3 s://doi.org/10.1016/j.cognition.2007.05.006 765/sp.5.6 Lüdtke, J., Friedrich, C. K., De Filippis, M., & Kaup, B. Rohde, H., Futrell, R., & Lucas, C. G. (2021). What’s (2008). Event-related Potential Correlates of new? A comprehension bias in favor of informativity. Negation in a Sentence–Picture Verification Cognition, 209, 104491. https://doi.org/10.1016/j.cogn Paradigm. Journal of Cognitive Neuroscience, 20(8), ition.2020.104491 1355–1370. https://doi.org/10.1162/jocn.2008.20093 Sperber, D., & Wilson, D. (1986). Relevance: Lüdtke, J., & Kaup, B. (2006). Context effects when Communication and cognition. Blackwell Publishing. reading negative and affirmative sentences. Tian, Y., Breheny, R., & Ferguson, H. J. (2010). Why we Proceedings of the 28th Annual Conference of the simulate negated information: A dynamic pragmatic Cognitive Science Society, 1735–1740. account. Quarterly Journal of Experimental Psychology, Marr, D. (1982). Vision: A computational investigation 63(12), 2305–2312. https://doi.org/10.1080/1747021 into the human representation and processing of visual 8.2010.525712 information. Henry Holt and co. Inc. Tian, Y., Ferguson, H., & Breheny, R. (2016). Processing Mœschler, J. (1992). The pragmatic aspects of linguistic negation without context – why and when we negation: Speech act, argumentation and pragmatic represent the positive argument. Language, Cognition inference. Argumentation, 6(1), 51–76. https://doi.or and Neuroscience, 31(5), 683–698. https://doi.org/10.1 g/10.1007/bf00154259 080/23273798.2016.1140214 Nieuwland, M. S., & Kuperberg, G. R. (2008). When the Van Rooy, R. (2003). Questioning to resolve decision truth is not too hard to handle. Psychological Science, problems. Linguistics and Philosophy, 26(6), 727–763. 19(12), 1213–1218. https://doi.org/10.1111/j.1467-92 https://doi.org/10.1023/b:ling.0000004548.98658.8f 80.2008.02226.x Wason, P. C. (1965). The contexts of plausible denial. Nieuwland, M. S., & Van Berkum, J. J. A. (2006). When Journal of Verbal Learning and Verbal Behavior, 4(1), peanuts fall in love: N400 evidence for the power of 7–11. https://doi.org/10.1016/s0022-5371(65)80060-3 discourse. Journal of Cognitive Neuroscience, 18(7), Xiang, M., Kramer, A., & Nordmeyer, A. E. (2020). An 1098–1111. https://doi.org/10.1162/jocn.2006.18.7.10 informativity-based account of negation complexity. Journal of Experimental Psychology: Learning, Memory, Nordmeyer, A. E., & Frank, M. C. (2014). A pragmatic and Cognition, 46(10), 1857–1867. https://doi.org/10.1 account of the processing of negative sentences. 037/xlm0000851 Collabra: Psychology 18 Pragmatic Felicity Facilitates the Production and Comprehension of Negation Supplementary Materials Peer Review History Download: https://collabra.scholasticahq.com/article/67931-pragmatic-felicity-facilitates-the-production-and- comprehension-of-negation/attachment/134865.docx?auth_token=lGIjssA_JmtTUGKMzNDB Collabra: Psychology

http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png

Collabra Psychology University of California Press

http://www.deepdyve.com/lp/university-of-california-press/pragmatic-felicity-facilitates-the-production-and-comprehension-of-ArGLpLaPuc

Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Loading next page...

Publisher: University of California Press
eISSN: 2474-7394
DOI: 10.1525/collabra.67931
Publisher site: See Article on Publisher Site

Abstract

Journal

Collabra Psychology – University of California Press

Published: Jan 26, 2023

Keywords: language; psycholinguistics; language comprehension; language production; pragmatics

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Pragmatic Felicity Facilitates the Production and Comprehension of Negation

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies