The influence of affect on the production of referring expressions

ABSTRACT

Many studies have provided evidence for the influence of affect on cognitive processing. However, experimental investigations of the relationship between affect and speech production are scarce. This study explores whether a speaker’s affective state influences the production of referring expressions. In two experiments, affective states were elicited using film excerpts, after which speakers referred to target stimuli in a way that differentiated them from distractors. The affective states were opposites, either in terms of valence (happiness versus sadness) or approach-avoidance motivation (anger versus disgust). Affective conditions were then compared with respect to the frequency with which participants referred to a target’s affect, whether this affect was congruent with the speaker's affective state, the number of modifiers per expression, the ambiguity of referring expressions, and overspecification. Results revealed no differences between different affective states concerning these factors, suggesting that a speaker’s affective state does not influence the production of referring expressions.

KEYWORDS:

• Emotion speech production
• emotional congruence
• emotional reference
• underspecification
• overspecification

Introduction

Suppose there is a man you wish to talk about, and suppose that unlike the other men in the room, this man is wearing glasses. The easiest way to refer to him would probably be The man with the glasses. This is a textbook example of a definite referring expression, with the man as the definite referent, and with the glasses as a modifier to distinguish this particular man from the other men in the room. Suppose, however, that except for glasses, a description of the man’s facial expression (e.g. The laughing man) could also be used as a modifier to distinguish him from the other men. Which modifier is chosen then? Does the speaker choose one of the two available modifiers, or are both included? Where does this decision fit in the larger framework of speech production?

An attempt to answer these questions was made by Levelt (1989) in his book Speaking: From Intention to Articulation. In it, he managed to organise a plethora of academic research about speech production, and described in detail how he thought it functions. This theoretical description was later developed into a computational model, WEAVER++ (Levelt et al., 1999; Roelofs, 1997), a cascading model, in which speech production is represented as a series of stages, each of which the speaker needs to pass through, before moving to the next one. The earliest stage in WEAVER ++ is conceptual preparation in terms of lexical concepts, which is the moment a speaker chooses ideas he/she wishes to express, for which there are words in the speaker’s language. Evidently, this is the moment of content selection for a referring expression, but questions regarding what content is selected and what factors play a part in this process, have not been addressed explicitly.

The production of referring expressions

Levelt (1989) stated that according to Grice’s maxims of quantity (Grice, 1975), (1) any referring expression should contain sufficient information to avoid ambiguity, but (2) not more information than necessary. In accordance with the first maxim of quantity, several studies revealed that underspecification (giving too little information for a listener to identify a target) is a rare phenomenon, that only occurs in approximately 0.5–6% of referring expressions, depending on the experimental context (Deutsch & Pechmann, 1982; Koolen et al., 2011, 2013; Pechmann, 1989). However, there seems to be less evidence for the second maxim of quantity, because overspecification (giving more information than necessary) occurs in approximately 21–60% of referring expressions (Deutsch & Pechmann, 1982; Engelhardt et al., 2006; Koolen et al., 2011; Pechmann, 1984, as cited in Levelt, 1989; Pechmann, 1989), and in those cases, it does not even always contribute to a listener’s identification of a target (Belke & Meyer, 2002; Pechmann, 1989). Given that overspecification clearly challenges Grice’s second maxim of quantity, a better understanding of factors that cause overspecification would improve our understanding of factors involved in content selection.

First, several studies suggest that overspecification sometimes occurs due to speakers selecting (technically) redundant modifiers, perhaps because they are aware they will genuinely help listeners identify a target (Arts et al., 2011; Levelt, 1989; Mangold & Pobel, 1988; Paraboni et al., 2007; Sonnenschein, 1982; Sonnenschein & Whitehurst, 1982). Evidence provided by Pechmann (1989) however, suggests that overspecification occurs because of the incremental nature of speech production: Speakers already start speaking before they have fully analysed what modifiers best distinguish a target from distractors, and consequently, they mention modifiers that eventually become redundant as more distinguishing modifiers are identified and added to the expression later. Koolen et al. (2011) found evidence this is especially likely in visually complex scenes. They argue that the complexity makes the scene difficult to analyse, prompting participants to add redundant modifiers in order to be communicatively successful.

What modifiers catch the speaker’s eye before others is another matter of debate. Some studies suggest priming effects are responsible, as participants were likely to use redundant modifiers, if they had been primed earlier during a conversation (Goudbeek & Krahmer, 2012; Herrmann & Deutsch, as cited in Levelt, 1989). On the other hand, it is also suggested that certain target features, such as colour, are highly salient to begin with, and will be included even if they are not useful (Belke & Meyer, 2002; Nadig & Sedivy, 2002; Pechmann, 1989).

In short, multiple factors are influential when selecting content for referring expressions, underspecification is unlikely, while overspecification occurs often. Hence, the laughing man with the glasses may indeed be referred to as The laughing man, but an overspecified expression such as The laughing man with the black glasses, would also be very likely.

A hiatus in the literature

It is interesting to note that one aspect of human psychology that could also influence content selection for speech production, has largely been ignored, namely, affect. Indeed, the studies on WEAVER++ and overspecification that were discussed earlier did not mention this topic at all. While a small number of studies have investigated the influence of affect on language production, few of them focussed on referring expressions specifically. To our knowledge, a 2013 study by Kempe, Rookes, and Swarbrigg, is the only study to have done so, and they found that if speakers were happy, the ambiguity of their referring expressions increased. Even though the study by Kempe et al. is the only one to directly investigate the influence of affect on referring expression production, there is much indirect evidence that is also suggestive of such a relationship. This work will be discussed in the next section.

Affect and cognition

Many studies have found that different affective states can influence an individual’s information processing style in a variety of ways (e.g. Bless et al., 1996; Bodenhausen, Kramer, et al., 1994; Converse et al., 2008; Forgas & East, 2008; Gable & Harmon-Jones, 2010a; Moons & Mackie, 2007; Niedenthal et al., 1997). Based on this, it is not difficult to imagine that this differential influence on information processing, also extends to speech production, given that information processing and speech production are intimately linked (as evidenced by Levelt (1989)). We will first clarify certain key concepts about affect, before examining the research concerning the influence affect can have on information processing.

One of the most obvious differences between affective states, is their valence, i.e. whether the affective state is positive or negative. However, over the years, research has revealed that such a view is limited, as there are at least two other dimensions on which affect can differ, namely approach-avoidance motivation (Calcott & Berkman, 2014; Elliot et al., 2013; Gable & Harmon-Jones, 2010c) and motivational intensity (Gable & Harmon-Jones, 2010a, 2010b). To give a short working definition of these terms: The approach-avoidance dimension refers to an affective state’s tendency to elicit a reaction to either approach or avoid an object (note that a precise definition is still the subject of debate (Elliot et al., 2013)), whereas motivational intensity entails the degree to which an affective state motivates an individual to take action. Note that while motivational intensity is an important dimension of affect, this study focuses on the other two equally important dimensions of valence and approach-avoidance motivation. The following subsection will therefore discuss how several affective states that are opposites in terms of these dimensions can influence information processing.

The influence of affect on information processing

Valence: Happiness vs sadness. Happiness is an affective state that is thought to be relatively low in both approach motivation and motivational intensity (Harmon-Jones et al., 2013), with an increased reliance on heuristic cues (Bless et al., 1990, 1996; Bodenhausen, Kramer, et al., 1994), and reduced perspective taking during communication (Converse et al., 2008). Bless et al. (1996, 1990) and Bodenhausen, Kramer, et al. (1994) argue that the increased reliance on heuristic cues is not due to a cognitive inability to implement a more analytical way of information processing, but due to a lack of motivation to do so. They found that if happy participants knew they would be held accountable for their judgments and could be asked to defend them, or were explicitly asked to analyse arguments thoroughly, they tended to process information more analytically.

Similar to happiness, sadness is generally thought to be an affective state that is low in approach motivation and motivational intensity (Harmon-Jones et al., 2013), but, in contrast to happiness, promotes analytical thinking and perspective taking (Bless et al., 1990, 1996; Clore & Huntsinger, 2007; Converse et al., 2008; Forgas & East, 2008).

Approach-avoidance motivation: Anger vs. disgust. Anger is an approach-oriented affective state with high motivational intensity (Carver & Harmon-Jones, 2009; Gable & Harmon-Jones, 2010b) that is usually associated with high arousal, a reliance on heuristic cues, and a lack of analytical thinking (Bodenhausen, Sheppard, et al., 1994; DeSteno et al., 2004). However, a study by Moons and Mackie (2007), called into question whether a reliance on heuristic cues and a lack of analytical processing are truly characteristic of anger, as the angry participants in their study displayed a surprisingly analytical information processing style. Moreover, although they agreed that angry individuals used heuristic cues to judge information, their experiments showed that such cues were only used, if participants considered them relevant. While these results seem to be at odds with those of Bodenhausen, Sheppard, et al. and DeSteno et al., Moons and Mackie noted they may be due to the fact that their study elicited a more moderate form of anger. Studies by Baron et al. (1992), and Sanbonmatsu and Kardes (1988), support this explanation, as they found a negative correlation between affective arousal and analytical information processing. Hence, intense anger may be associated with reduced analytical tendencies, but moderate anger appears to have the opposite effect.

Disgust is characterised as an avoidance-oriented affective state with high motivational intensity. While little is known about the influence of disgust on information processing, a study by Calcott and Berkman (2014) demonstrated that individuals with avoidance-oriented affect (e.g. disgust) differ from those with approach-oriented affect (e.g. [moderate] anger) in how easily they can switch between global and local elements of stimuli. This suggests that although disgust and anger both lead to high motivational intensity and are negative in terms of valence, they need not necessarily elicit the same analytical tendencies. Which of the two affective states leads to the greatest analytical tendencies, remains an empirical question, but we speculate that disgust is the lesser one in this regard. The reasoning behind this is that if someone is moderately angry, the cause (e.g. deception, a lie, etc.) may be more likely to necessitate analytical processing of information to construct an argument, while in the case of disgust, the cause of discomfort is more likely to just be avoided instead of addressed with an argument. As a result, disgust may be less likely to elicit analytical tendencies than moderate anger.

Affective congruence. While the previous subsections have focused mainly on the influence of specific affective states on information processing, the current section presents a phenomenon that applies, presumably, to all of the discussed affective states. Several studies have shown that an individual’s affective state can interact with similarly affectively charged information, and influence how this information is recalled or perceived; a phenomenon this study will refer to as affective congruence (Bower, 1981; Bower et al., 1981; Forgas & Bower, 1987; Niedenthal et al., 1997; Niedenthal & Setterlund, 1994). For example, when hearing a story about characters describing happy and sad incidents, happy participants remembered more details about happy incidents, while sad participants demonstrated similar results for sad incidents (Bower et al., 1981). Interestingly, Bower, and Forgas and Bower found that if the affective states of participants and stimuli differed, but were the same in terms of valence (e.g. both were negative), effects of affective congruence could be observed. In contrast, Niedenthal et al., and Niedenthal and Setterlund found that affective congruence only occurred if the affective states of participants and stimuli were exactly the same. These opposing results may be due to differences in the experimental context.

Cognition and language production

While not aimed specifically at investigating referring expressions, other research discovered that proprioceptive cues (e.g. bodily feelings) altered a person’s information processing style, which in turn was found to influence language production (Beukeboom & de Jong, 2008; Friedman & Förster, 2000). More specifically, Friedman and Förster (2000) stated that flexing or extending one’s arm gives rise to a conditioned neural response, which subsequently influences the information processing style. For example, when individuals flexed their arm, they displayed an improved ability to solve verbal analogy tasks, suggesting that they engaged in a more creative, global style of information processing that enabled them to detect common features and relations more effectively. However, during a task that required careful processing of information to determine the veracity of statements, individuals who extended their arm, outperformed those who flexed their arm, suggesting extending one’s arm promoted non-creative, analytical processing of information.

Beukeboom and de Jong (2008) subsequently investigated whether the change in processing style, caused by flexing or extending one’s arm, also influenced language production. They discovered that individuals who extended their arm and engaged in analytical processing, were more likely to display lower levels of abstraction in descriptions of their own behaviour in certain social situations (e.g. “I talk and dance” (p. 8), revealing a more detailed focus on specific actions). Conversely, they also found that individuals who flexed their arm, and engaged in a more global style of information processing, produced descriptions with a higher level of abstraction (e.g. “I am outgoing” (p. 8), which is unspecific regarding their exact behaviour).

The studies in this section and the ones above, demonstrate that affect can influence cognition, and that cognition, in turn, can influence language production. With this in mind, it is reasonable to suggest affect can therefore influence language production. The next section will discuss studies that explored this relationship.

Affect and language production

While there is ample indirect evidence that affect is linked to language production, only a few studies have investigated this relationship directly. For example, research by Forgas (1999a, 1999b) revealed that speakers in a negative affective state would produce more polite requests, especially if there was a risk of offending someone, whereas positive affective states would lead to more impolite requests. Additionally, Beukeboom and Semin (2005) found that the affective state of speakers influenced how abstract their description of an event would be. For example, speakers who were experiencing negative affect were observed to use words in their descriptions that signified lower levels of abstraction, such as verbs (e.g. “A punches B”, which concretely describes the action), whereas speakers who were experiencing positive affect, used words that signified greater levels of abstraction, such as adjectives (e.g. “A is aggressive”, which focus more on what “A punches B” implies). Forgas (2007) also revealed that when asked to persuade others of a certain point of view, speakers produced more persuasive messages if they were experiencing a low-intensity negative affective state,than if they were experiencing a positive affective state.

Research question and hypotheses

So, a substantial body of literature provides evidence that affect can influence language production. However, not many studies have experimentally investigated to what extent the content selection stage of speech production is influenced by this relationship, making it unclear to what extent models of speech production should take affect into account, especially regarding the production of referring expressions. Therefore, the current study aims to offer more insight into this matter. This leads to the following research question: Does the affective state of a speaker influence the content of referring expressions? Based on the previously discussed literature, we address this question by analysing how several affective states, which contrast either in terms of valence (operationalised as happiness versus sadness) or approach-avoidance motivation (operationalised as disgust versus anger), differ with respect to five dependent variables, which will be discussed in the following subsection.

Hypotheses

Multiple studies have demonstrated that information is recalled or noticed more easily if the affective state that is associated with that information, is either entirely congruent with the affective state the individual is experiencing at the time (Bower et al., 1981; Niedenthal et al., 1997; Niedenthal & Setterlund, 1994), or at least congruent in terms of valence (Bower, 1981; Forgas & Bower, 1987). This so-called affective congruence may also extend to speech production, or more specifically, the production of referring expressions. Therefore, we predict that unlike speakers who are neutral in terms of affect, speakers who experience affect, will show a preference for target properties that are at least congruent with their own affective state in terms of valence, and will refer to these properties more frequently than speakers in an incongruent (or neutral) affective state. For instance, happy speakers could be more likely to refer to a target stimulus’s happiness, while sad or moderately angry speakers could be more likely to refer to a target’s sadness. Contingent on this hypothesis, the increased focus of affective speakers on (congruent) affect in target stimuli, should increase the frequency with which they refer to a target’s affective state, compared to speakers in the neutral condition.

The use of modifiers in referring expressions is also thought to be modulated by affect. Given that sad and moderately angry speakers have a greater tendency to be analytical (Bless et al., 1990; Forgas, 2007; Forgas & East, 2008; Moons & Mackie, 2007), we expect that they will spend more time analysing the contribution of modifiers before including them in a referring expression. So, compared to happy, disgusted, and neutral speakers, sad and moderately angry speakers will use fewer modifiers per referring expression, regardless of whether the expression is under- or overspecified. By contrast, happy and disgusted speakers are expected to use more modifiers than sad, moderately angry, or neutral speakers, as their less analytical information processing style may make them more prone to adding modifiers based on quick heuristics (Bless et al., 1990, 1996; Bodenhausen, Kramer, et al., 1994).

With respect to the degree of underspecification, we hypothesise that happy and disgusted speakers are more likely to underspecify their referring expressions compared to sad, angry, and neutral speakers. This prediction is based on Kempe et al. (2013), who found that happy individuals were more likely to produce underspecified referring expressions than neutral individuals. By contrast, the strong analytical tendencies of sad and moderately angry speakers are thought to lead to less underspecification compared to happy, disgusted, and neutral speakers, as they are predicted to be more inclined to analyse how much information a listener needs to distinguish a target from distractors.

Finally, regarding overspecification, the analytical tendencies of sad and moderately angry speakers are thought to enable them to better determine the optimal properties for a (minimally) distinguishing referring expression, resulting in referring expressions with less overspecification, compared to happy, disgusted, and neutral speakers. Conversely, the reduced analytical tendencies of happy and disgusted individuals will make it harder to find the most efficient modifiers, leading to an increase in overspecification compared to the sad, moderately angry, and neutral speakers.

The current study

To test these hypotheses, two experiments are conducted, during which participants are required to refer to human faces and distinguish them from distractor faces. The first experiment focuses on differences in valence by comparing happy and sad speakers, while the second experiment focuses on approach-avoidance motivation by comparing moderately angry and disgusted speakers (leaving the valence dimension untouched). In both experiments, neutral speakers are used as a control group.

Experiment 1

Method

Participants

The data of 90 students (78 undergraduate students and 12 graduate students, 62 females) from Tilburg University’s Department of Communication and Cognition were collected. Each condition (positive, negative, neutral) contained exactly 30 participants, 15 per stimulus order. Undergraduate students participated to obtain ‘participant pool credit’ (a mandatory part of their bachelor’s programme), whereas graduate students volunteered without further compensation. An additional 27 participants had been tested, but they were not included in the final data analysis due to severe technical difficulties with equipment (e.g. software that failed to run the experiment, or headphones that did not record audio properly).

Materials

Consent form

To keep participants unaware of the true purpose of the experiment, the consent form stated that the purpose of the experiment was to investigate the influence of affect and speech production on memory. After informing participants about the experiment’s procedure and duration (approximately 25 min), the consent form also stated that participants would see film footage that could make them feel uncomfortable, and that they would be offered a funny video at the end of the experiment, if they needed something to lift their spirits. It was also explained data would be stored anonymously, would only be accessible to the authors of this paper, and would only be used for scientific research and publications. Lastly, it was explained that participation was voluntary, and that participants could quit the experiment at any moment without negative consequences, should they feel unable to continue.

Affect elicitation

Since film excerpts have been shown to be an excellent medium to elicit affect (i.e. Westermann et al., 1996), we selected several seven-minute film excerpts based on their expected potential for eliciting a particular affective state (see Appendix A for a more detailed description of these excerpts). To make sure our speakers were in the intended affective state throughout the experiment, the affect elicitation was repeated halfway through the experiment: Each participant was shown two different film excerpts, one at the beginning of the experiment, and one in the middle.

A 7-point Likert-scale with four items was used after each excerpt to measure how effective it had been at eliciting the intended affective state (1 = sadness, 7 = happiness, 4 = neutrality). To prevent response strategies, the anchors of each Likert item were reversed consecutively. The internal consistency of the reported affect scores was assessed by taking the scale items of both Likert scales (eight scale items in total) and calculating their Cronbach’s alpha. This resulted in α = .910 for the happiness condition, α = .931 for the sadness condition, and α = .897 for the neutral condition, indicating good to excellent reliability.

Stimuli

The stimuli consisted of 30 experimental and 10 filler trials. Experimental trials consisted of four human faces (two male, two female), with one target face and three distractors, while filler trials consisted of four geometric shapes, again with one target and three distractors (see Figure 1). The human faces were obtained from a subset of the Karolinska Directed Emotional Faces dataset (Goeleven et al., 2008). The facial expressions of these stimuli varied from neutral, to happy and sad, and it was ensured that the affective states of the target stimuli appeared in equal numbers throughout the experiment (i.e. sad targets appeared as frequently as happy and neutral targets). Additionally, non-affective salient attributes, such as hats, glasses, bowties, and light or dark hair colours, were added to both targets and distractors. Combining these non-affective attributes with affective states in stimuli enabled participants to refer to the target by mentioning the affective state, the non-affective attribute, or a combination of both (see Figure 1). Each trial would be visible for 5 s, forcing participants to mention features that were most salient to them, and preventing them from mindlessly piling up modifiers in hopes of distinguishing the target from the distractors. To prevent affective states from being more salient than non-affective attributes as a result of their frequency, no particular affective state occurred more than twice per trial (e.g. there were only two happy faces per trial).

Figure 1. Above, an example of an experimental trial, and below, an example of a filler trial. Participants needed to unambiguously refer to the stimulus with the arrow (e.g. The laughing man, The laughing man with dark hair, etc.).

Regarding non-affective attributes, it was ensured that when a particular non-affective attribute was distinguishing, a distractor of the opposite gender would also display this attribute (e.g. both a male and female would have black hair). This was done to prevent the attribute from drawing a disproportionate amount of attention to itself by virtue of its uniqueness (e.g. when the target is the only one with black hair).

After the experiment was conducted, four of the experimental trials were found to require participants to mention two (instead of one) modifiers to properly distinguish the target from distractors. This was not in accordance with the preset criterion that it should be possible to accurately refer to a target with only one modifier. Therefore, these trials were omitted from the statistical analysis, resulting in 26 usable experimental trials per participant.

Variables and design

Independent variables

The experiment used a 3 × 2 between-subject design, with affective condition as the independent variable and presentation order of stimuli (henceforth known as stimulus order) as a control variable. Affective condition consisted of three levels (happiness, sadness and neutrality), each presented in two counterbalanced stimulus orders. Participants were randomly assigned to the conditions.

Dependent variables

The dependent variables we assessed were: (1) affective congruence, (2) affective word usage, (3) modifier usage, (4) underspecification, and (5) overspecification. Regarding affective congruence, whenever a participant described the affective state of a target as either positive or negative, it was coded as 1 (positive affect) or 0 (negative affect). Affective word usage refers to the mentioning of a target’s affective state in a referring expression (this includes neutral affect, e.g. The man with a neutral face), and it was coded as 1 (affect mentioned) or 0 (affect not mentioned). Modifier usage was defined as the number of modifiers in a referring expression. A modifier was defined as a word that modifies either the head noun of a referring expression or another modifier. For example, in the expression The woman with the black glasses, glasses was considered a modifier of woman, but black was also considered a modifier, as it added a level of specificity to glasses. Underspecification was defined as a situation in which the referring expression was not sufficiently specified to distinguish a target from distractors, and it was coded as 0 (at least minimally specified, i.e. the expression contained the bare minimum of information needed to unambiguously distinguish the target from distractors) or 1 (underspecified). Lastly, overspecification took referring expressions into account that were at least minimally specified. If a referring expression was minimally specified, it was coded as 1, whereas the coding for overspecification was equal to the number of superfluous modifiers in the expression. In preparation of the statistical analysis, the mean score of each dependent variable was calculated per participant.

Number of referring expressions

Ninety participants had responded to 26 valid experimental trials, which yielded a total of 2340 recordings. However, not all these recordings contained referring expressions that could be analysed. First, some recordings did not contain any audio, either due to a technical malfunction or because the participant did not say anything. Secondly, some recordings were incomplete, because the participant did not finish the referring expression within the 5 s time limit. After a clean-up, 2314 recordings could still be used for the analyses of affective congruence and affective word usage, because participants had already mentioned the target’s affective state, or it was clear they were going to mention it (e.g. The woman who looks … [happy]). Additionally, 2292 recordings could be used for the analyses of modifier usage, underspecification, and overspecification, because the modifiers used in the referring expression could be reasonably inferred. Such predictions were only made if all expected modifiers, except for one, had been mentioned, or if other participants in the same affective condition frequently mentioned the same set of modifiers for that particular stimulus, and one of those modifiers was still missing from the referring expression. However, if the other participants used varying sets of modifiers, there was too much uncertainty to predict what modifiers would be used in the incomplete referring expression, and hence, it was omitted from the statistical analysis.

Statistical analysis

For our main analyses, we used mixed effect models. Data were analysed using the R Statistical Package (version 3.6.1, R Core Team, 2019) in conjunction with the lme4 package (version 1.1.-10, Bates et al., 2015) for mixed effects analysis, and the lmerTest package (version 3.1-1, Kuznetsova et al., 2017) to obtain p values. Implementing mixed models enabled us to take into account the effects of multiple stimuli across affective conditions, and multiple measurements per participant within an affective condition. In addition, we assessed the evidence for the null hypothesis using Bayesian analysis of variance, as implemented in JASP (version 0.11.1, JASP Team, 2019).

For affective congruence, we ran a weighted Generalised Linear Mixed Model by using the frequencies with which participants had referred to negative or positive affect in their referring expressions as weights in the analysis. Consequently, the scores of participants who had referred more frequently to the positive or negative affective state of target stimuli received proportionally more weight. This was done to ensure that participants who only rarely mentioned positive or negative affect, would not disproportionately influence the analysis. The general specification of the full models that were used for the analyses is given below: $\begin{array}{rl}(\mathrm{g}& \left)\mathrm{l}\mathrm{m}\mathrm{e}\mathrm{r}\right(\mathrm{d}\mathrm{e}\mathrm{p}\mathrm{e}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{v}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{b}\mathrm{l}\mathrm{e}\sim \mathrm{h}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}\\ & \times stimulusorder\end{array}+\mathrm{s}\mathrm{a}\mathrm{d}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}\times \mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{e}\mathrm{r}+\left(1+\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}|\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\right)+\left(1+\mathrm{s}\mathrm{a}\mathrm{d}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}|\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\right)+\left(1|\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{p}\mathrm{a}\mathrm{n}\mathrm{t}\right))$Note that the variables happiness, sadness, and stimulus order are dummy variables, and that the type of dependent variable dictated whether a Generalised Linear Mixed Model (glmer) or a Linear Mixed Model (lmer) was used. Appendix C reports on tests that assess whether models with varying slopes provided a better fit than models with only random intercepts, and provides parameter estimates, standard errors and p values for all fixed effects.

Additionally, while the mixed models provide a valid test for the presence of an effect (i.e. rejecting the null hypothesis), they cannot -easily- quantify evidence in favour of the null hypothesis. So, we supplemented each mixed model with a 3 × 2 two-way Bayesian ANOVA, with affective condition (happiness, sadness, neutrality) and stimulus order (1 and 2) as independent variables, using the statistical software package JASP (version 0.11.1, JASP Team, 2019). Because of the exploratory nature of this study, choosing the right priors was difficult, hence, aside from the default prior used by JASP (0.5), we conducted each analysis with multiple priors to assess the robustness of the results (priors were entered in JASP under the option r scale fixed effects).

Because JASP does not offer the possibility to add weights in a Bayesian ANOVA, we conducted two Bayesian analyses for affective congruence. In the first one, we only selected participants who referred to the affective state of targets at least 15 times across experimental trials, and in the second one, we adjusted this number to 10 to include more participants. These cut-off points prevented participants who only rarely mentioned positive or negative affect, from skewing the results of the analysis, while still allowing us to retain a reasonable number of participants per affective condition. Moreover, using two cut-off points enabled us to assess the robustness of results, and potentially, bolster any conclusions. For reasons of brevity, the results of the full model were reported in the running text and the complete results of the Bayesian analyses were reported in Appendix C.

Procedure

After providing informed consent, indicating that they understood they were allowed to quit the experiment at any moment without negative consequences, should they feel unable to continue, participants disposed of any chewing gum and shut off (or handed in) their cell phone. After this, they entered a soundproof booth, sat down at the computer and received instructions. The experiment started with a brief example where the experimenter told participants they had 5 s to refer to the intended target, without mentioning its location. Participants then viewed a film excerpt (the affect elicitation), for which they required the headset connected to the computer in the booth. To avoid problems regarding the loudness of the audio in the film excerpt, participants could adjust the volume as they saw fit. After the film excerpt, they rated the affect they were experiencing on a digital 7-point Likert scale with four items. Participants then continued with three test trials to get accustomed to the experimental procedure. After this, they would see a number of experimental stimuli, each displaying four faces in a row, and were required to refer unambiguously to the face the red arrow was pointing at. These trials were separated by 3-second intervals.

After repeating the entire procedure twice, participants exited the booth to take a mock memory test, which served to maintain the illusion that the experiment was about the influence of affect and speech production on memory (which was how the experiment had been advertised). For the memory test, participants wrote down the three most salient things they could remember about the film excerpts they had seen. This concluded the experiment for participants in the happiness and neutral conditions. However, participants in the sadness condition were also offered a funny video (the Friends episode the happiness condition had viewed) to lift their spirits, because they had seen film footage that could have made them feel uncomfortable.

Results and discussion

Affect elicitation

A two-way ANOVA with affective condition and stimulus order as factors, revealed a significant main effect of affective condition on reported affect, F(2, 84) = 99.96, MSE = 0.738, p < .001. Post-hoc analyses (Bonferroni corrected) confirmed that the happiness (M = 5.92, SD = 0.79), sadness (M = 2.79, SD = 0.85) and neutral conditions (M = 4.48, SD = 0.91) differed significantly from one another (p < .001) in the expected direction (i.e. sadness had the lowest scores, happiness the highest, and the neutral condition sat in between). In conclusion, it appears the affect elicitation was successful.

Affective congruence and affective word usage

A weighted generalised linear mixed model (GLMM) with random slopes and intercepts was fitted to test whether participants in the happiness and sadness conditions displayed affective congruence in their referring expressions (i.e. whether they referred more frequently to affective states that were congruent with their own emotional state in terms of valence). The analysis suggests there was no main effect of affective condition on the degree of affective congruence. Compared to the neutral condition, the ratio of positive to negative affective words in referring expressions did not differ significantly for the happiness condition (B = 3.02, SE B = 2.17, p = .165) and sadness condition (B = −1.35, SE B = 2.94, p = .645). This suggests that participants in the happiness and sadness conditions were not more likely to refer to positive or negative affect in target stimuli respectively. Therefore, no support was found for the role of affective congruence in referring expression generation.

To test whether participants who were experiencing affect were more likely than neutral participants to refer to a target’s affective state, a random intercept GLMM was fitted. The results show that participants in the happiness (B = 0.50, SE B = 0.83, p = .547) and sadness conditions (B = −0.10, SE B = 0.82, p = .899) did not differ from participants in the neutral condition in how frequently they referred to a target’s affective state. In conclusion, there was no support for the hypothesis that participants who were experiencing affect, were more likely to mention the affective state of a target, compared to neutral participants.

Modifier usage, underspecification and overspecification

An analysis of modifiers (e.g. The happy man with the pretty, black glasses) was carried out to investigate whether happy participants used more modifiers in their referring expressions, and whether sad participants used fewer, compared to the neutral condition. A random intercept linear mixed model (LMM) suggested that participants in the happiness (B = −0.45, SE B = 0.23, p = .055) and sadness conditions (B = 0.02, SE B = 0.23, p = .928), used similar numbers of modifiers in their referring expressions as participants in the neutral condition. Hence, there was no support for our hypothesis regarding modifier usage.

Next, it was analysed whether participants in the happiness and sadness conditions underspecified more and less respectively, compared to the neutral condition. A general analysis found that only 5% of all referring expressions were underspecified, and the subsequent random intercept GLMM revealed there was no significant effect of affective condition on underspecification. Compared to the neutral condition, participants in the happiness (B = −0.12, SE B = 0.47, p = .806) and sadness conditions (B = −0.06, SE B = 0.47, p = .901), did not differ in their degree of underspecification. Therefore, our hypothesis regarding underspecification was not supported.

Lastly, we analysed whether participants in the happiness and sadness conditions overspecified more and less respectively, compared to the neutral condition. In total, 79% of the referring expressions were overspecified, and the subsequent random intercept LMM suggested that participants in the happiness condition (B = −0.19, SE B = 0.17, p = .267) and sadness condition (B = 0.002, SE B = 0.16, p = .988) did not display significantly differing degrees of overspecification, compared to the neutral condition. Consequently, there was no support for our hypothesis regarding overspecification.

For two of the analyses above, we also observed significant effects for the order in which stimuli were presented. Participants in the second stimulus order condition were more likely to display affective congruence in their referring expressions than participants who viewed the first stimulus order, and in the happiness condition they used fewer modifiers. However, the statistical significance of the affective congruence result was not high, and in both cases, the effects were relatively small and difficult to interpret.

Bayesian analyses

Although the analyses above indicate the results are likely under the null hypothesis, they do not actually provide evidence in favour of it. Therefore, multiple Bayesian two-way ANOVAs were conducted to shed light on this matter. Depending on the prior, these analyses yielded Bayes Factors between <0.01–0.18 and 0.05–0.67 for affective congruence with cut-off points of 15 and 10, respectively, Bayes Factors between 0.03-0.93 for affective word usage, 0.05–1.01 for modifier usage, 0.01–1.26 for underspecification, and Bayes Factors of < 0.01–0.12 for overspecification (see Appendix C for more detailed results). Bayes Factors below 1 indicate evidence in favour of the null hypothesis, hence regardless of the prior, our results indicate either weak to strong support for the null hypothesis, but almost never support for the alternative hypothesis (van Doorn et al., 2019). In fact, even when Bayes Factors provided evidence for the alternative hypothesis, the evidence was weak and disappeared when other priors were used.

In summary, the mixed models suggest that affective states that are opposites in terms of valence, do not influence the production of referring expressions, and these results were bolstered by those of the subsequent Bayesian ANOVAs. In Experiment 2, we assess the role of the approach-avoidance motivation dimension in the production of referring expressions.

Experiment 2

Method

Participants

The data of 103 students (101 undergraduate students and 2 graduate students, 68 females) from Tilburg University’s Department of Communication and Cognition (n = 74, 52 females) and Department of Social Psychology (n = 29, 16 females) were analysed. There were 34 participants in the anger condition, 37 in the disgust condition, and 32 in the neutral condition. Undergraduate students participated to obtain ‘participant pool credit’ (a mandatory part of their bachelor’s programme), whereas graduate students volunteered without further compensation. Five additional participants had been tested, but they were not included in the final data analysis due to technical difficulties with equipment (e.g. software that failed to run the experiment, or headphones that did not record audio properly).

Materials

Consent form

The consent form was identical to the one in experiment 1.

Affect elicitation

To elicit anger and disgust, we used multiple validated film excerpts from an online dataset (Mahau, 2016) accompanying a paper by Schaefer et al. (2010). Schaefer et al. had not only investigated to what extent film excerpts elicited a particular affective state, but also to what extent they elicited multiple affective states. The excerpts of the current experiment were selected on the basis of their ability to elicit anger or disgust, without strongly eliciting other affective states, while neutral affect was maintained using the same film excerpts from experiment 1 (see Appendix B for more detailed information regarding the film excerpts).

Affect was measured the same way as in Experiment 1, with two slight differences. First, this experiment used two digital slider scales (with four items each), as their continuous nature allowed participants to report their affective state in a more nuanced way than would be possible using Likert scales (Treiblmaier & Filzmoser, 2011). Each time a slider item appeared, the question “What are you feeling?” was written above it. Scores close to 1 indicated disgust, scores close to 7 indicated anger, and scores close to 4 represented neutrality. Secondly, unlike the Likert scales in the first experiment, the slider scales in the current experiment explicitly stated that the number ‘4’ represented a neutral state of mind (e.g. “What are you feeling?”, 4 = “nothing”) to ensure participants would interpret the scale correctly. As was the case in experiment 1, slider scale item anchors were reversed consecutively. The internal consistency of the reported affect scores was assessed by taking the scale items of both slider scales (eight scale items in total) and calculating their Cronbach’s alpha. This resulted in α = .694 for the anger condition, α = .807 for the disgust condition, and α = .926 for the neutral condition, indicating adequate to excellent reliability.

Stimuli

36 experimental and 12 filler stimuli were used. These were similar to the ones used in experiment 1, but four slight changes were made to simplify them and reduce cognitive load for participants. (1) All non-affective attributes were present in equal numbers throughout the experiment (whereas glasses occurred more frequently than other attributes in experiment 1). (2) To reduce stimulus complexity, only one non-affective attribute would vary per trial (e.g. hair colour could vary, but none of the other non-affective attributes would). (3) When a non-affective feature was distinguishing, both the target and a distractor of the opposite gender would be the same regarding this feature (e.g. both would have black hair), whereas the other two distractors would differ from the target regarding this feature (e.g. they would have blonde hair). (4) Only two affective states were displayed per experimental trial. For instance, if the target and a distractor of the opposite gender were laughing, the other distractors would look sad.

Variables and design

The (in)dependent variables and design of the experiment were the same as in experiment 1.

Number of referring expressions

As each of the 103 participants completed 36 experimental trials, 3708 audio recordings were obtained. However, for the same reasons as in experiment 1, not all recordings contained referring expressions that could be used for analysis. If a recording contained an incomplete referring expression, the rationale for including or excluding it from the analysis was the same as in experiment 1. This yielded 3693 referring expressions that were usable for the analyses of affective congruence and affective word usage, and 3599 expressions that could be used for the analyses of modifier usage, underspecification, and overspecification.

Statistical analysis

Except for the fact that the affective states under inspection differed, the statistical analysis was the same as in experiment 1. Complete results for both the mixed models and Bayesian analyses can be found in Appendix D. The general specification of the full mixed-models is given below: $\left(\mathrm{g}\right)\mathrm{l}\mathrm{m}\mathrm{e}\mathrm{r}(\mathrm{d}\mathrm{e}\mathrm{p}\mathrm{e}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{v}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{b}\mathrm{l}\mathrm{e}\sim \mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{r}\times \mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{e}\mathrm{r}+\mathrm{d}\mathrm{i}\mathrm{s}\mathrm{g}\mathrm{u}\mathrm{s}\mathrm{t}\times \mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{e}\mathrm{r}+\left(1+\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{r}|\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\right)+\left(1+\mathrm{d}\mathrm{i}\mathrm{s}\mathrm{g}\mathrm{u}\mathrm{s}\mathrm{t}|\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{u}\mathrm{s}\right)+\left(1|\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{p}\mathrm{a}\mathrm{n}\mathrm{t}\right))$

Procedure

The procedure was almost identical to the procedure of experiment 1. However, in the current experiment, the duration of film excerpts in the anger and disgust conditions was relatively short, hence, multiple excerpts were shown consecutively during each phase of the affect elicitation to elicit an affective state of sufficient intensity. To create a less abrupt transition between film excerpts, and to encourage participants to stay focused on the experiment, digital slider scale items appeared between excerpts, asking whether the participant was feeling positive or negative. These items did not collect any data. At the end of each phase of the affect elicitation, participants indicated whether they were feeling angry, disgusted, or neutral on a digital slider scale with four items. Additionally, participants in both the disgust and anger condition were offered to watch the Friends episode from experiment 1 to cheer them up.

Results and discussion

Affect elicitation

A two-way ANOVA revealed a significant main effect of affective condition on reported affect, F(2, 97) = 25.42, MSE = 0.655, p < .001. Post-hoc analyses (Bonferroni corrected) revealed there was no significant difference between the anger (M = 3.03, SD = 1.03) and disgust conditions (M = 2.61, SD = 0.68), p = .099, but the analyses also indicated both of them differed significantly from the neutral affect condition (M = 3.98, SD = 0.63), p < .001. Given that scores close to 1, 4, and 7 indicated disgust, neutrality, and anger, respectively, the pairwise comparisons suggested that the affect elicitation was successful at eliciting disgust and maintaining neutrality, but not at eliciting anger. In fact, the mean score of the anger condition (3.03), in conjunction with the pairwise comparisons, suggested that participants in this condition were, on average, merely less disgusted than participants in the disgust condition. This was not entirely unexpected, as the online dataset (Mahau, 2016) that accompanied the study by Schaefer et al. (2010), already indicated that some of the film excerpts they had associated with anger, could also elicit some degree of disgust.

Affective congruence and affective word usage

We fitted a weighted generalised linear mixed model (GLMM) with random slopes and intercepts to test whether participants in the anger and disgust conditions displayed affective congruence in their referring expressions. The analysis suggested there was no main effect of affective condition on the degree of affective congruence. Compared to the neutral condition, the ratio of positive to negative affective words in referring expressions did not differ significantly for the anger condition (B = −1.00, SE B = 1.05, p = .339) and disgust condition (B = −1.13, SE B = 0.80, p = .868). This suggests that participants in these conditions were not more likely to refer to positive or negative affect in target stimuli respectively. Consequently, there was no evidence to support the hypothesis that anger and disgust lead to affective congruence in referring expression generation.

To test whether participants in the anger and disgust conditions were more likely to refer to a target’s affective state than participants in the neutral condition, a random intercept GLMM was conducted. It indicated that participants in the anger (B = −1.54, SE B = 1.54, p = .316) and disgust conditions (B = −0.55, SE B = 1.50, p = .715) did not refer to a target’s affective state more or less frequently than participants in the neutral condition. Consequently, there was no support for the hypothesis that participants who were experiencing affect, were more likely to refer to a target’s affective state than neutral participants.

Modifier usage, underspecification and overspecification

To test whether, compared to the neutral condition, participants in the disgust condition used more, and participants in the anger condition use fewer modifiers, a random intercept linear mixed model (LMM) was fitted. The results suggested that participants in the anger (B = −0.12, SE B = 0.49, p = .805) and disgust conditions (B = −0.18, SE B = 0.47, p = .698) used similar numbers of modifiers in their referring expressions as participants in the neutral condition. Therefore, our hypothesis regarding modifier usage was not supported.

With respect to underspecification, we analysed whether participants in the disgust condition underspecified more, and those in the anger condition underspecified less, compared to the neutral condition. Only 2.6% of all referring expressions were underspecified, and the subsequent random intercept GLMM revealed there was no significant effect of affective condition on underspecification. Compared to the neutral condition, participants in the anger (B = 2.52, SE B = 1.66, p = .128) and disgust conditions (B = 1.60, SE B = 1.58, p = .311), did not differ significantly in terms of underspecification. Consequently, no support was found for our hypothesis regarding underspecification.

As for overspecification, we analysed whether participants in the disgust and anger conditions overspecified more and less respectively, compared to the neutral condition. In total, 74% of all referring expressions were overspecified, and the subsequent random intercept LMM suggested that participants in the anger condition (B = −0.13, SE B = 0.49, p = .783) and disgust condition (B = −0.18, SE B = 0.48, p = .710) did not differ significantly from the neutral condition in their tendency to overspecify. Therefore, our hypothesis regarding overspecification was not supported.

For two of the analyses above, the order in which stimuli were presented also seemed to exert a significant influence. In both cases, there was a relatively large discrepancy between stimulus order 1 and 2 in the anger condition, regarding affective congruence and underspecification. However, these results were difficult to interpret and their statistical significance was not high.

Bayesian analyses

As was the case in the first experiment, multiple Bayesian two-way ANOVAs were conducted to analyse whether the data offer support for the null-hypothesis or alternative hypothesis. Using several priors, these analyses yielded Bayes Factors between 0.03–0.92 and 0.02–0.98 for affective congruence with cut-off points of 15 and 10, respectively, Bayes Factors between <0.01–0.07 for affective word usage, <0.01–0.04 for modifier usage, <0.01–0.11 for underspecification, and Bayes Factors of <0.01–0.04 for overspecification (see Appendix D for more detailed results). Bayes Factors below 1 indicate evidence in favour of the null hypothesis, hence regardless of the prior, these results either indicate weak to strong support for the null hypothesis, but never support for the alternative hypothesis (van Doorn et al., 2019).

In sum, because the affect elicitation procedure did not successfully elicit anger, it is still unknown how an approach-oriented affective state would have influenced the production of referring expressions. However, both the mixed-models and the Bayesian analyses indicated that compared to the neutral control group, avoidance-oriented affect did not significantly influence the production of referring expressions.

General discussion

Many studies have demonstrated that affect can have a notable influence on an individual’s information processing style (Bless et al., 1990, 1996; Bodenhausen, Kramer, et al., 1994; Bodenhausen, Sheppard, et al., 1994; Calcott & Berkman, 2014; Carver & Harmon-Jones, 2009; Converse et al., 2008; DeSteno et al., 2004), and some have even directly demonstrated the influence of affect on language production (Beukeboom & Semin, 2005; Forgas, 1999a, 1999b, 2007). However, the number of studies that directly investigated this link is limited, and even fewer studies have ever investigated whether the influence of affect extends specifically to the production of referring expressions. Therefore, the aim of the current study was to provide more insight into this matter, by conducting two experiments that sought to answer the following research question: Does the affective state of a speaker influence the content of referring expressions?

Although this study corroborated results of previous studies regarding the degree of underspecification and overspecification in referring expressions, the data did not support any of our hypotheses. Neither affective states that differ in terms of valence (happiness and sadness), nor affective states that differ in terms of approach-avoidance motivation (anger and disgust), yielded significant differences compared to neutral participants, regarding affective congruence, the frequency with which a target’s affective state was mentioned, the total number of modifiers used per referring expression, underspecification, and the degree of overspecification. Consequently, no evidence was found to suggest that the affective state of a speaker influences the content of his/her referring expressions. Of course, the most obvious explanation for the observed null results is that the hypothesised effects simply do not exist. However, this is not necessarily the case, as the next section will discuss other plausible causes that stem from methodological limitations.

Limitations and strengths

The first reason for the lack of an experimental effect, pertains to the effectiveness of the affect elicitation procedure. It is possible that the scores of self-reported affect were the result of demand characteristics. However, this seems unlikely, given that at the end of the experiments, participants frequently and spontaneously said the film excerpts were funny, boring, sad, etc. Moreover, participants were never explicitly told to enter an affective state, which has indeed been shown to elicit demand characteristics (Westermann et al., 1996). A more likely explanation for the observed results, is that the elicitation procedure was simply not capable of eliciting an affective state that was strong enough to cause the hypothesised results, or perhaps only in the first trials after the affect elicitation, after which the affect may have diminished. While we presented our participants with affect inducing film excerpts twice during the experiment (which is more than most studies on affect do), affective states induced by film excerpts might not last long enough to have an optimal effect on all, or even most of our stimuli. Perhaps these issues are inherent to the type of affect we elicited in our experiments. Indeed, two types of affective response can be distinguished: incidental, and integral. In our experiments, we elicited an incidental affective response, which means the response was caused by a stimulus (e.g. a film excerpt) that was unrelated to the target under inspection (Cohen et al., 2007). By contrast, integral affective responses are directly caused by and aimed at a target object, such as a decision, a person or a company. Given that the cause of the affective response was no longer visible during the course of our experiments, this may have caused the affective intensity to wane as the experiment progressed, leading to the observed null results.

Furthermore, due to practical constraints, the current study only attempted to investigate affective states that were opposites in terms of valence or approach-avoidance motivation. However, these may not be the relevant dimensions. Of course, the current study has not yielded enough evidence to suggest research into the approach-avoidance dimension is a dead end, but it is worth noting that comparing affective states that differ in terms of motivational intensity, may provide another fruitful line of research. After all, previous research has shown that affective states that were both negative in valence and low in approach motivation, yet differed in terms of motivational intensity, produced different results in terms of attentional focus (Gable & Harmon-Jones, 2010a). More specifically, low motivational intensity was associated with attentional broadening, while high motivational intensity was associated with attentional narrowing. Additionally, sadness, which is low in motivational intensity (Gable & Harmon-Jones, 2010a), is associated with increased analytical tendencies (Bless et al., 1990, 1996), whereas (intense) anger, which is high in motivational intensity, is associated with heuristic processing strategies (Bodenhausen, Sheppard, et al., 1994). Further evidence for the importance of motivational intensity is provided by Arts et al. (2008), who demonstrated that speakers who are highly motivated to communicate successfully with a recipient, are more likely to overspecify their referring expressions.

Besides factors related to affect, the experimental design and procedure may also have limited the influence of affect on referring expression generation. For example, the time limit of 5 s for producing a referring expression, could have made the experimental task too cognitively demanding, which may have caused participants to keep adding modifiers in hopes of producing an unambiguous referring expression. This is plausible, given the high percentage of overspecified referring expressions in our experiments (79% and 74%, respectively), which is substantially higher than even the highest percentage (60%) found in earlier studies (Pechmann,1984; as cited in Levelt, 1989). More specifically, approximately 34.4% and 24.7% of all referring expressions in experiments 1 and 2, respectively, contained 2 superfluous modifiers, while 8.7% and 8.3% of the referring expressions contained 3 or more (max. = 5) superfluous modifiers. These percentages suggest time pressure may have been an issue during both experiments. This is further supported by incidental conversations with participants after the experiments, because they said time pressure was a challenge at the beginning of the experiment. Furthermore, it could be the case that by explicitly instructing participants to refer to a target in an unambiguous way, the natural tendencies of certain affective states may have been overridden, causing happy, and disgusted individuals to become substantially more analytical, even though they were hypothesised to be less so. This notion is based on Bodenhausen, Kramer, et al., (1994) who found that happy participants were capable of and likely to engage in analytical processing, if they were sufficiently motivated to do so. Likewise, participants of the current study may have felt motivated enough to use analytical processing strategies, thus obscuring the hypothesised effects. On the other hand, Kempe et al. (2013) found that even while being motivated to communicate effectively, happy speakers were still more likely to produce ambiguous referring expressions than neutral speakers. These contradictory results make it difficult to determine whether motivating participants to produce unambiguous referring expressions led to null results, but the possibility cannot be ruled out.

Lastly, the composition of our stimuli may also have influenced results. For example, simple attributes such as glasses, may be less cognitively demanding to refer to, and therefore more likely to be mentioned than more cognitively challenging features, such as facial expressions. After all, a facial expression with its affective meaning, presumably requires more cognitive resources to process than simple features with no affective meaning.

Besides these limitations, our study also had a number of strengths, including a careful attempt to induce affective states in speakers using validated video excerpts, substantial sample sizes (certainly for a language production study), a large number of diverse, original stimuli, and elaborate statistical analyses. That being said, more research is needed to determine whether the null effects we reported are due to the set-up of the experiments, or due to weakness of any influence of affective state on language production in general, and referring expressions in particular. In view of these empirical uncertainties, it is difficult to formulate strong implications for theory at the moment, but we nevertheless want to make some suggestions.

Implications

Kempe et al. (2013) found that happy participants were more prone than neutral participants to be ambiguous when producing referring expressions, but the current study did not replicate this finding. In fact, it did not observe any significant differences between affective states, regarding several dependent variables related to speech production. As of yet, the amount of research that shows affective states influence speech production is sparse. However, it would be premature to conclude that speech production models, such as described in WEAVER++ (Levelt, 1989; Levelt et al., 1999) or A Blueprint of the Speaker (Levelt, 1999), do not need to take affect into account as an influential factor in referring expression production.

In Levelt’s model, speech production is a process that can be divided into three main stages: content selection, message formulation, and articulation. During the stage of content selection, speakers decide what to say, i.e. what attributes they wish to incorporate into their message. This results in a so-called preverbal message. During the subsequent stage, message formulation, this preverbal message is used for lexical retrieval and grammatical encoding to form the words of the actual utterance. Lastly, during the stage of articulation, these words are encoded phonologically and articulated. There is certainly evidence for the influence of affect on the early and final stages of language production in this model. For example, Stirman and Pennebaker (2001) demonstrated that suicidal poets were more likely to use first person singular pronouns, words related to death, and fewer references to other people in their poems, suggesting affect influences the stage of content selection. Additionally, work by Scherer (2003) revealed that sad people are more likely to speak in a softer voice, demonstrating that affect can influence articulation.

Regarding our study, we assumed that affect would be influential during the stage of content selection, either by making certain attributes of targets more salient than others, or by enabling speakers to assess the relevance of attributes for distinguishing the target more effectively, causing them to be incorporated into the preverbal message. However, we did not observe any evidence for this: Speakers who were experiencing affect, did not refer to a target’s affective state more frequently than neutral speakers, they did not display affective congruence, use different numbers of modifiers, or overspecify and underspecify at different rates than neutral speakers. Therefore, affect does not appear to influence content selection, for this particular set of dependent variables. Interestingly, Stirman and Pennebaker (2001) also did not find much evidence for affective congruence among suicidal and non-suicidal poets, aside from a slight preference of suicidal poets for words related to death.

Moreover, perhaps the influence of affect could also be found at a later stage, such as message formulation. Instead of influencing ‘what’ to say, affect may also influence ‘how’ speakers say it. For example, it could be the case that affective speakers are indeed more inclined to focus on the affective state of a target, but that this does not manifest itself in mentioning it more frequently. Instead, perhaps they simply mention it earlier in a referring expression than neutral speakers: The smiling girl vs The girl who is smiling. In this case, an affective speaker may be more inclined to use the target’s affective state as a pre-modifier, whereas a neutral speaker may be more likely to use it as a post-modifier. Moreover, affective congruence could also manifest itself in a different way, with happy speakers being more likely to use adjectives related to happiness (e.g. happy, smiling, laughing) as pre-modifiers, and using negative adjectives (e.g. unfriendly, sad, angry looking) as post-modifiers. Furthermore, if speakers are experiencing an affective state that elicits a more analytical processing style, they may prefer absolute target properties, such as colour, over relative ones, such as size, as they do not require any comparisons between the target and distractors, and would therefore have a greater chance of leading to communicative success. Hence The green, big box may be preferred over The big, green box.

Future directions

Given the amount of research that has shown affect can influence information processing, and given that little research has focused specifically on the influence of affect on the production of referring expressions, we consider this line of research worth pursuing. That being said, it may be interesting to apply different affect elicitation procedures and experimental paradigms. For example, even though Westermann et al. (1996) demonstrated that the affect elicitation procedure used in the current study was the most successful procedure of the ones they had analysed, we were unable to elicit moderate anger and elicited mild disgust instead. Based on these findings, future studies should carefully assess whether their affect elicitation procedure elicits anger, and not a mixture of affective states. This may be difficult when using film excerpts, because the dataset of Schaefer et al. (2010) (see Mahau (2016)) already indicates that the film excerpts they associated with anger, were also capable of eliciting disgust to some extent. However, alternative affect inducing procedures, such as cued recall, might be more specific.

It may also be advisable to reduce the number of stimuli after affect elicitation to make sure the affect is still strong enough to influence the production of referring expressions when speakers produce their descriptions. However, if reducing the number of stimuli is not desirable, one could also elicit affect more frequently during the course of the experiment, thereby countering the dissipation of affect. This can be achieved by either repeating the affect elicitation procedure more frequently, or by focusing on an integral type of affective response, instead of an incidental one. After all, as was noted in the section Limitations and strengths, if an incidental affective response is used, stimuli during the experimental trials are not responsible for the speaker’s affective state, which means affect could be experienced less intensely as the experiment progresses, leading to null results. With an integral affective response, affect may be experienced more intensely, given that the stimuli that cause the affective state will also be the stimuli under consideration during the experimental trials.

Furthermore, instead of focusing solely on the valence and approach-avoidance motivation dimensions of affect, future studies could also take other dimensions into account, such as motivational intensity. After all, as was already discussed in the section Limitations and strengths, previous research has shown that affective states that differ in terms of motivational intensity, yielded differences regarding several dependent variables (Arts et al., 2008; Bless et al., 1990, 1996; Bodenhausen, Sheppard, et al., 1994; Gable & Harmon-Jones, 2010a). Hence, it is plausible that focusing on motivational intensity will provide a fruitful line of research.

To analyse whether explicitly instructing participants to produce unambiguous referring expressions can cause happy and disgusted participants to become more analytical, this instruction could be added as a control variable, by having half of the participants do the experiment after getting this instruction and the other half without getting it. Additionally, to reduce cognitive load, it may also be useful to relax the time limit participants are given to produce a referring expression. Finally, it also seems worthwhile to explore the influence of affect on the formulation stage of speech production by analysing the position of different types of modifiers in referring expressions.

Acknowledgements

We would like to extend our gratitude to Lisa Niederdorfer for providing us with the stimuli and experimental design that formed the foundation of the stimuli and design of the current study. We are also grateful to Jacqueline Dake for providing technical assistance, and the equipment that was used to run the experiments. Lastly, we would like to thank Tünde van Hoek for putting the participant pool of Tilburg University’s School of Social and Behavioral Sciences at our disposal.

Disclosure statement

No potential conflict of interest was reported by the author(s).