Comment: Can We Model What an Emotion Is? Comment on Suri & Gross (2022)

The question “what is emotion?” has long been at the core of theoretical debates. The IAC-E is a useful framework for understanding relationships among responses in emotional situations. However, this approach cannot address the nature of emotion. Researchers determine what counts as emotion in the IAC-E, and this decision impacts the relationships detected and inferences made. The assumptions of researchers about emotion change the output. Further, the model is not theoretically agnostic and is best suited to examine emotion perception/knowledge, as in the simulations presented. According to some theories, experienced emotion is qualitatively different than situations that involve perceiving others’ emotion or semantic knowledge. Addressing the nature of emotion requires empirical examination of the assumptions made in each theory.

Emotions are a challenge to operationalize, similar to any relatively high-level psychological construct, such as “thoughts” or “personality.” Emotion theorists have argued for decades about what an emotion is and how emotions unfold, as Suri and Gross (2022) eloquently point out. These arguments over how to define emotion are in the service of forwarding the scientific study of emotion by creating a coherent account that operationalizes emotion (Lench, Bench, & Flores, 2013). The Interactive Activation and Competition framework for Emotion (IAC-E) is put forward as a potential solution to arguments over the nature of emotion.

This solution is intuitively appealing. The framework is sufficiently flexible to include all theoretical views of emotion, including the two extremes that emotions are biologically innate or socially constructed, and everything in between. The framework is described as agnostic on these issues, instead estimating the degree to which different assumptions about emotion hold in various situations. In this comment, we argue that the framework cannot provide an answer to what an emotion is and is not agnostic. We do not intend to suggest that the framework is not useful, but that modelling ipso facto cannot address theoretical issues about the nature of emotion.

Theory Informs Models

The IAC-E cannot address the question of what an emotion is. This is because theories about emotion guide decisions by researchers about what to include in simulations, and thereby determine the output. This is a classic problem in any modelling approach, often captured in the idea of garbage-in, garbage-out. Inferences from modelling will only be as good as the data and assumptions fed into the model.

The first required step to use the IAC-E is to define an emotional instance of interest, and the framework then estimates associations in the accompanying dataset. At this first step, the researcher's assumptions have a tremendous impact on the output. For example, simulations in Suri and Gross (2022) focus on the ability to identify facial expressions of emotion. Does perceiving emotions on other people's faces represent an emotional instance? Can the results of this study tell us anything about the nature of emotion? Other examples of the challenges inherent in identifying emotional instances include how to conceptualize mixed experiences like nostalgia, or how to resolve the fact that not all participants experience the target emotion.

Researchers’ choices at this first step about what constitutes an emotional instance will determine the relationships detected. One could argue that it does not matter because the model is flexible enough to include any definition. But then it is also true that the framework cannot reveal what an emotion is. Inferences from the model will only be as good as the information put in, and that is based on the emotion theory guiding researchers’ choices.

Further, it is not clear what conclusions we would be able to draw after running simulations on hundreds of datasets that address emotional instances. An average computed across all of the simulations would most likely reveal that there is some consistency and some variability in emotion, rather than the nature of emotion itself. What is perhaps more interesting would be to examine the sources of variance in estimates of consistency. In other words, what factors contribute to greater or lesser consistency?

This is consistent with many theories of emotion that hold that both innate and learned responses contribute to emotional instances. The issue from a modelling perspective is that evidence of consistency in one emotion instance does not necessarily extend to other emotion instances. For example, there is evidence that snake-like movements trigger fear responses that appear biologically based (Öhman & Mineka, 2001). But identifying one instance of fear that appears innate does not lead to the conclusion that all fear instances have that quality. Including moderators is one potential path forward to address this issue. Here again, the emotion theory of the researcher will guide choices about which moderators to include and how to categorize them. The assumptions that researchers make, based on their theory of emotion, will determine what comes out of the simulations.

The Framework is Not Agnostic

Suri and Gross (2022) begin their paper by referencing a well-established conceptualization of emotion as “… a response to an event that has perceived significance for the individual” (Keltner & Gross, 1999). However, the description of the IAC-E never conceptualizes emotions this way. This limitation is reinforced in several simulations, which appear to primarily illustrate emotion perception or knowledge, rather than the actual experience of emotion as defined.

The first simulation involves participants selecting emotion corresponding to a facial expression. This is not consistent with the provided definition of emotion; there is no response and no perceived significant event. In the second simulation, participants recalled a past emotional episode. Although recall can evoke emotion in the present, memory for past events increasingly relies on semantic beliefs and knowledge as time passes from the event. Therefore, the simulation is likely partially capturing an "on-line" experience and partially capturing beliefs about emotion (Robinson & Clore, 2002). Similarly, none of the supplemental simulations capture emotion as defined.

This issue appears to extend beyond the specific studies chosen for simulation. The framework itself assumes hidden knowledge constructs (conjunction units) that are connected to responses (feature units). The intent of these hidden units is to permit the framework to capture variability in responses. This is presumed to occur when emotions are more learned and less innate (in which case, pools would directly connect to each other). However, any source of shared variability, including from theoretical and methodological moderators, could appear in these hidden pools. As a result, the framework could be used to compare variability estimates across study types or emotions, but the reason(s) for that variability will be unclear.

The framework is also described as capturing emotions that are appraisal-led, as well as other possible sequences. However, the framework appears to assume that external events elicit emotions and that appraisals are a pool within the framework. A prototypic example is given of walking into a job interview. This scenario clearly has external inputs, as the person is sensing the room and their interactions around interviewing. These inputs activate different pools, some of which could be appraisals. But what about emotion when contemplating possible future job interviews? The framework does not easily account for situations where emotion derives from ongoing internal processes rather than external input (Lazarus, 1991; Parkinson, 2001).

Emotion is Whatever You Think It Is

Answering questions about the nature of emotion requires experiments that directly test core assumptions of each perspective (Lench, Bench, Darbor, & Moore, 2015). If emotions are constructed through learned interactions, for example, then it should be possible to teach people completely new “emotions” that consist of randomly selected responses. This type of evidence would begin to address what emotion is.

Debate about the nature of emotion has raged for hundreds of years. There is a widespread parable about the blind men who are introduced to an elephant for the first time. The men touch parts of the elephant to identify the creature. Each touches a different part, and their identification corresponds to the part of the elephant they are holding, such as perceiving the elephant as rope-like when holding the tail. The men argue over the nature of the elephant, each believing their own identification based on the part that they can reach. But the whole of the elephant can only be understood by conceptualizing the nature of the entire creature. Similarly, emotion theory is the view that allows us to understand the pieces and to make predictions about how emotions function and their consequences.