Hurting BERT’s Feelings: Toward a Computational Model of Emotion Regulation in Context Público

Abstract

Emotion regulation is a complex phenomenon, typically operationalized using ontologies of different strategies (e.g., cognitive reappraisal) and assessed using low dimensional assays, with affect ratings serving as the sole measure of regulation success. Such low-dimensional conceptualizations of emotion regulation limit the development computationally explicit accounts of the cognitive and neural processes involved in different regulation strategies. Here we take an alternative approach, using deep language models to test the hypothesis that emotion regulation changes the meaning of events as reflected in a high-dimensional semantic space. We conducted an online study in which participants either used reappraisal, mindfulness, or provided an objective description of stimuli after viewing short affective videos. Fine-tuned Bidirectional Encoder Representations from Transformers (BERT) were able to classify regulation strategy as well as emotional situation based on text descriptions of events. Fine-tuning in regulation specific language changed representations across layers of a BERT model, with a main effect of situation in the last layer and an interaction between strategy condition and situation. These findings suggest that regulation systematically alters the language produced when describing emotional events. The nature of these changes depends both on the type of emotional situation and the emotion regulation strategy employed. Importantly, these changes increase deeper into a language model, with implications for brain mappings of early vs late cortical areas. We further assessed generalizability of our models to independent archival data and found that model classifications better explained variations in affect compared to the experimental labels. This work paves the way for objective modeling of emotion regulation, with applications in a variety of settings in order to facilitate a deeper understanding of how emotion and emotion regulation are operationalized behaviorally and in the brain.

Table of Contents

1.Abstract Cover Page

2.Abstract

3.Thesis Cover Page

4.Introduction

8.Decoding emotional situations from context-sensitive BERT embeddings.

11.Emotion regulation alters the meaning conveyed by BERT embeddings

15.BERT embeddings predict regulation strategy and self-reported affect in independent studies

17.Discussion   

21.Methods

26.References

About this Master's Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Psychology
Subfield / Discipline	Clinical Psychology
Degree	M.A.
Submission	Master's Thesis
Language	English
Research Field	Computer Science Psychology, Cognitive Psychology, Clinical
Palavra-chave	language emotion deep language models emotion regulation
Committee Chair / Thesis Advisor	Philip Kragel, Emory University
Committee Members	provide evidence that artifi- cial neural networks trained on language tasks develop a rich set of both semantic and syntactic knowledge (Reif et al., 2019; Rogers et al., 2020), Emory University Patricia Brennan, Emory University

Última modificação

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Hurting BERT’s Feelings: Toward a Computational Model of Emotion Regulation in Context ()	2024-01-12 14:00:33 -0500	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions