Salons & Methods

Jeffrey Girard, University of Kansas

Large language models for sentiment analysis

This 60-minute workshop introduces Large Language Models (LLMs) as an accessible, state-of-the-art tool for estimating sentiment in unstructured text data, including essays, social media posts, and transcripts. We will explore how modern LLMs can outperform traditional methods by utilizing “zero-shot in-context learning,” a technique that yields highly accurate sentiment estimates without requiring labeled training data, advanced programming skills, or specialized hardware. Participants will learn practical workflows for applying this technique using two approaches: cloud-based models for speed and power, and locally hosted models for strict data privacy and security. The session will also demonstrate how to leverage the R programming language to automate these tasks, enabling the efficient processing of large file batches. Furthermore, we will review findings from the instructor’s recent publication in Affective Science, which provides a rigorous validation and fairness audit of LLM-based sentiment analysis across naturalistic speech datasets from social and clinical psychology. The workshop will conclude with a dedicated Q&A period to address specific implementation queries, ensuring attendees leave with the rationale and technical know-how to apply these methods in their own research. 

Vlad Chituc, Yale University

Affective psychophysics: Or, how (not) to measure subjective experience

People can feel more or less happy, hold opinions with more or less conviction, and find some experiences more or less painful. Though all of these examples concern a magnitude, it is far from clear how such magnitudes could be accurately measured, given their inherently subjective nature. In this workshop, I highlight two common and underappreciated problems with the labeled (e.g., Likert) scales that are ubiquitous in affective science, and I provide straightforward solutions adapted from sensory psychophysics. First, these scales are interpersonally relative, in that they can mean different things to different people. This can produce spurious group differences: for example, labeled scales suggest that women experience anger more intensely than do men, but this finding disappears when tested with psychophysical measures designed to avoid such artifacts. Second, these scales are nonlinear: across dozens of studies spanning perceptual, evaluative, affective, and motivational domains, Likert ratings are a logarithmic compression of underlying psychological magnitude, meaning that averaging such ratings systematically underestimates the mean. Both problems have solutions requiring minor changes to measures and analyses, involving at most a dozen lines of code. More broadly, these findings illustrate how insights from classical psychophysics can be productively applied to the measurement challenges that are most central to affective science.

Joey Heffner, Yale University

(Beginner-Friendly) Computational Modeling of Emotion

How do our choices and their outcomes translate into subjective feelings? How does our happiness depend on our choices and what happens to us? Computational accounts of emotion aspire to answer these questions with a rigorous framework informed by formal principles. This methods workshop provides an accessible introduction to modeling emotion within the context of decision-making. We will cover how momentary happiness and affect ratings can be modeled during risky decision-making tasks, while discussing applications to other domains.

This workshop is aimed at graduate students and researchers who are new or interested in the field and want to learn how to use computational approaches to better understand emotions. Prior programming experience is helpful but not required. The workshop focuses on a popular computational model of happiness (Rutledge et al., 2014), using interactive Shiny apps to show how expectations, rewards, and prediction errors combine to influence happiness. By gaining an intuition for the abstract logic used in computational modeling, participants will leave with a clearer understanding of how to implement these tools in their own research. Example data and programming scripts will be provided.