PaperlayerView exact revision on arXiv

Source-linked AI summary

Language-Specific Representation of Emotion-Concept Knowledge Causally Supports Emotion Inference

Ming Li, Yusheng Su, Hsiu-Yuan Huang, Jiali Cheng, Xin Hu, Xinmiao Zhang, Huadong Wang, Yujia Qin, Xiaozhi Wang, Kristen A. Lindquist, Zhiyuan Liu, Dan Zhang

arXiv:2302.09582v5cs.AIcs.CL

TL;DR

Whether language-derived emotion-concept knowledge contributes causally to emotion inference remains an open question. The study examines LLM neuron representations and finds that language-specific knowledge can support emotion inference without sensory-motor representations.

Abstract

from arXiv · show

Humans no doubt use language to communicate about their emotional experiences, but does language in turn help humans understand emotions, or is language just a vehicle of communication? This study used a form of artificial intelligence (AI) known as large language models (LLMs) to assess whether language-based representations of emotion causally contribute to the AI's ability to generate inferences about the emotional meaning of novel situations. Fourteen attributes of human emotion concept representation were found to be represented by the LLM's distinct artificial neuron populations. By manipulating these attribute-related neurons, we in turn demonstrated the role of emotion concept knowledge in generative emotion inference. The attribute-specific performance deterioration was related to the importance of different attributes in human mental space. Our findings provide a proof-in-concept that even a LLM can learn about emotions in the absence of sensory-motor representations and highlight the contribution of language-derived emotion-concept knowledge for emotion inference.

Introduction

The introduction asks whether language-derived emotion-concept knowledge supports emotion inference beyond communication and outlines an LLM-based test using emotion-specific prompts and comparisons with human ratings of emotion attributes. It describes manipulating related artificial neurons to examine their causal role and whether attribute contributions track their importance in human mental space, without relying on human sensory-motor representations.

Introduction

  • Traditional accounts treat language as communicating emotions after the fact, whereas constructivist accounts propose that language shapes discrete emotion concepts and their contextual application (Ekman, 1992; Keltner, Dacher, Cordaro, 2015; Barrett, 2017; Lindquist, 2013).Source: Introduction · p. 3
  • Causal evidence for language-grounded knowledge in emotion inference is lacking; priming changes access to emotion concepts rather than directly manipulating their representations (Lindquist, 2017; Firestone & Scholl, 2014, 2015).Source: Introduction · pp. 3–4
  • Semantic dementia patients cannot infer categorical meaning from facial behaviors, but the anterior temporal lobe’s sensory connections complicate isolating language-derived knowledge (Lindquist et al., 2014; Mahon & Caramazza, 2008; Sabsevitz et al., 2005).Source: Introduction · p. 4
  • The study uses LLMs and selective manipulation of attribute-related artificial neurons to test their causal role in emotion inference about novel situations.LLMs can be examined as systems that acquire knowledge through human language use alone, and their neurons can be selectively manipulated for functional tests.Source: Introduction · p. 4
  • The study selected RoBERTa as a pre-trained base model because it performs emotion-related tasks, is untuned by user feedback, and has open parameters for manipulation experiments.Source: Introduction · p. 5

Prompted emotion inference

  • The LLM performed 27 emotion-inference tasks using emotion-specific prompts trained on the training set, with model parameters and prompts fixed for testing.Each task used the same task dataset and a prompt tailored to the emotion being inferred.Source: Introduction · p. 5

Representational comparison

  • Human ratings of 14 emotion-concept attributes—including core affects, prototypical expressions, and antecedent appraisals—were compared with LLM representations using representational similarity analysis.The comparison used a higher-level functional perspective.Source: Introduction · p. 5

Neuron manipulation

  • Guided by the representational space of human ratings, the researchers located and manipulated LLM neurons relevant to emotion attributes to investigate their causal contribution to inference.Valence is given as an example of a conceptual attribute.Source: Introduction · p. 5

Attribute importance

  • The researchers explored whether the language-based contribution of different emotion attributes was associated with their importance in human mental space.This analysis was intended to inform understanding of emotion-concept constitution and inference.Source: Introduction · p. 5

LLM Infers Emotions Based on Shared Conceptualization

The LLM inferred emotions across 27 tasks, with accuracy varying by task and tracking human-rater agreement. Its artificial neurons represented all 14 emotion-concept attributes, with attribute-specific neurons distributed across layers and less overlapping across attributes.

LLM Infers Emotions Based on Shared Conceptualization

  • Average accuracy across 27 emotion-inference tasks ranged from 68.04% (realization) to 96.43% (gratitude).The LLM was more accurate on tasks with greater human-rater agreement (Pearson's r(25) = .797, p < .001).Source: LLM Infers Emotions Based on Shared… · p. 6

Language-Specific Representations of Emotion-Concept Knowledge

  • Each of 14 emotion-concept attributes was significantly related to a subset of neurons distributed across all LLM layers (FDR-corrected q < .01).The analysis compared neuron-activation representational dissimilarity matrices with human-rated attribute matrices.Source: Language-Specific Representations of… · p. 6
  • The most relevant neurons for different attributes overlapped less, with the top 4,000 attribute-specific neurons illustrated in Figure 2c.Results were also examined across different top-N selections.Source: Language-Specific Representations of… · p. 6

Emotion-Concept Knowledge Causally Contributes to Emotion Inference

Selectively manipulating attribute-specific neurons reduced emotion-inference accuracy relative to both the original model and random same-count neuron manipulations. Attribute-specific contributions were related to human mental-representation weights, although effects varied with the number of manipulated neurons.

Emotion-Concept Knowledge Causally Contributes to Emotion Inference

  • Selective manipulation reduced inference accuracy versus the original model and random manipulation of the same number of neurons, with the largest deterioration at the top 4,000 attribute-specific neurons (FDR-corrected p < .05).The reported significance test used accuracy drops averaged across random seeds and conceptual attributes.Source: Emotion-Concept Knowledge Causally… · p. 6
  • There was no significant heterogeneity in reliance on emotion-concept knowledge across the 27 inference tasks for any tested attribute or neuron count (minimum p = .127).The task-to-task heterogeneity assessment used Hartigan's dip test.Source: Emotion-Concept Knowledge Causally… · pp. 6–7

Importance in Mental Space Predicts Language-Based Contribution of Different Attributes

  • In the LLM, 12 of 14 attributes significantly contributed to inference under some tested neuron counts; self-relativity and disgust-face did not.Significance was based on accuracy drops averaged across random seeds and inference tasks, with FDR correction.Source: Importance in Mental Space Predicts… · p. 7
  • Human mental representations reflected 12 of 14 attributes, excluding arousal and other-relativity (FDR-corrected p < .001).The reported test used bootstrap sampling and a two-tailed signed-rank test.Source: Importance in Mental Space Predicts… · p. 7
  • Language-based contributions to inference were related to human mental-representation weights, with the strongest correlation at 4,000 manipulated neurons (Fisher-based average r = .473, p < 10^-15).The reported statistic was t(26) = -16.714, SD = .030.Source: Importance in Mental Space Predicts… · p. 7
  • The relationship between attribute knowledge contribution and manipulated-neuron count was inverted-U shaped from 1,500 to 6,000 neurons, possibly reflecting floor and ceiling effects.The authors describe floor and ceiling effects as a possible explanation.Source: Importance in Mental Space Predicts… · p. 7

Discussion

The LLM’s language-specific emotion-concept representations causally supported inference and could differentiate emotions from meaningful contexts without sensory-motor representations. The findings also connect the contribution of this knowledge to human mental representations while bounding what language-based inference alone establishes.

Discussion

  • The findings support the possibility that emotion-concept attributes can derive from statistical regularities among linguistic symbols.This provides computational evidence for learning emotion concepts from language experience.Source: Discussion · p. 8
  • LLM emotion-inference performance was consistent with human raters’ agreement, suggesting the language corpus reflected shared understanding that included cognitive and experiential attributes.The reported attributes included core affects and similarity to prototypical facial expressions.Source: Discussion · p. 8
  • Emotion-attribute-related neurons were distributed across LLM layers, and populations for different attributes tended to be distinct.The authors suggest this may reflect low- and high-level linguistic regularities in emotion conceptualization.Source: Discussion · p. 8
  • Language-specific emotion-concept representations were sufficient in principle to differentiate discrete emotions from meaningful contexts without sensory-motor representations.Manipulating attribute-specific neurons examined their functional role in emotion inference.Source: Discussion · p. 8
  • Language-based emotion inference is not exclusive of sensory-motor or other semantic processes integrated with sensory-motor experience.The study also could not establish whether the language-specific representations of self-relativity and disgust-face need other experiential information to function.Source: Discussion · p. 9
  • Language-based knowledge contributions were related to the importance of conceptual attributes in human mental representations of emotion.The authors also report weak heterogeneity in knowledge contributions across 27 emotions.Source: Discussion · p. 9

Limitations of the Study

The LLM setup excludes sensory-motor representations, so the study cannot determine whether emotional feelings are rooted in language. Its findings may also not generalize to LLMs trained using other strategies; the data and code are publicly available, and participants gave informed consent.

Limitations of the Study

  • Because the LLM excludes sensory-motor representations, the study cannot determine whether emotional feelings are rooted in language.The authors also do not intend the study to answer that question.Source: Limitations of the Study · p. 9
  • The authors caution that other LLM training strategies may not provide evidence of the same strength.Source: Limitations of the Study · p. 9

Data and Code Availability

Informed Consent

Methods

The methods recast GoEmotions as 27 binary emotion-inference tasks, train emotion-specific prompts while freezing the LLM, and use prompt-elicited neuron activations to represent emotion-concept knowledge. Inference performance is evaluated against human rater agreement.

Emotion Inference Tasks

  • GoEmotions contains 58,009 English Reddit comments manually labeled with 27 emotions; each emotion defines a separate category-membership inference task, and data were split 80%/10%/10% into training, development, and test, with development unused.Source: Emotion Inference Tasks · p. 11

Training of Emotion-Specific Task Prompts

  • Emotion-specific prompts were optimized on training data while the LLM parameters were frozen, preserving its learned knowledge; prompts need not use emotion words or be human-readable.For each sample, only emotions with which most raters agreed were considered, and each task prompt was trained with 12 random seeds.Source: Training of Emotion-Specific Task… · p. 11

Evaluation of LLM's Inference Performance

  • Inference performance on the test set was related to human rater agreement for each emotion, estimated with Cohen’s Kappa.The 27 prompts across random seeds were evaluated on the test set.Source: Evaluation of LLM's Inference… · p. 11

Model-Based Knowledge Representation

Human Ratings of Emotion-Concept Knowledge

The authors gathered human ratings for 14 attributes across 27 emotions, then compared attribute and neuron representations and selectively manipulated attribute-related neurons during emotion inference. Task reliance was assessed by comparing accuracy drops with random-neuron controls.

Human Ratings of Emotion-Concept Knowledge

  • Human ratings covered 14 attributes of 27 emotions: two core affects, six prototypical expressions, and six antecedent appraisals.Ratings came from three independent experiments with English-speaking participants recruited through Prolific.Source: Human Ratings of Emotion-Concept… · p. 12

Searchlight RSA for LLM's Representation of Emotion-Concept Knowledge

  • They built symmetric 27-by-27 emotion RDMs using Euclidean distances between random-seed-averaged neuron activations to emotion-specific prompts, plus RDMs for 14 conceptual attributes, and used one-tailed sign-rank testing of RDM relatedness.They did not perform bootstrap because of many calculations.Source: Searchlight RSA for LLM's… · p. 12

Artificial Neuron Manipulation Experiment

  • During inference, attribute-specific neuron activations were set to zero and compared with random manipulation of the same number of neurons.The number of manipulated neurons varied across levels.Source: Artificial Neuron Manipulation… · p. 13

Task Reliance on Emotion-Concept Knowledge

  • Task reliance on emotion-concept knowledge was estimated from accuracy drops averaged across random seeds and conceptual attributes, with paired tests and FDR correction.The authors assessed drops separately for each emotion task and manipulation level.Source: Task Reliance on Emotion-Concept… · p. 13

Heterogeneity Test for the Reliance of Different Tasks

The analyses tested how emotion-concept attributes support inference, how that support relates to human mental representations, and whether reliance varies across tasks. Selective neuron manipulation reduced inference accuracy, and attribute-level contributions were compared with human representational weights.

Heterogeneity Test for the Reliance of Different Tasks

Language-Based Knowledge Contribution

  • Language-based contribution was estimated from accuracy drops averaged across random seeds and inference tasks at each manipulated-neuron count.False discovery rate correction was used to assess significance.Source: Language-Based Knowledge Contribution · p. 13

Concept Similarity Judgment Experiment

RSA for Knowledge Weight in Mental Representation of Emotion Concepts

  • RSA estimated how well each of 14 attribute representational dissimilarity matrices fit participants' emotion-concept representations using Kendall's tau.The attribute-wise tests used bootstrap sampling and false discovery rate correction.Source: RSA for Knowledge Weight in Mental… · p. 14

Figures

  • The figures show attribute-related neurons ranked by association and manipulated during test-set inference, with accuracy drops compared against random-neuron controls.Source: Figures · pp. 23–24
  • Selective manipulation of attribute-specific neurons reduced emotion-inference accuracy, while task prompts represented 14 emotion-concept attributes.The analysis used neuron populations associated with those attributes to test their contribution during inference.Source: Figures · p. 22, Figure 3 · p. 24

Mental Representation

  • Language-based contributions tracked attribute weights in human mental representations, with self-relativity and disgust-face as exceptions.Source: Figures · p. 25, Figure 4 · p. 25

Contact Information

The paper reports equal author contributions and describes training emotion-specific prompts for RoBERTa.

Contact Information

Training Emotion-Specific Task Prompts

Collecting LLM-Based Knowledge Representation

The study derived human ratings for 14 emotion-concept attributes and related them to RoBERTa neuron activations. Selectively manipulating attribute-related neurons reduced inference performance, while emotion-level accuracy was positively related to human rater agreement.

Collecting LLM-Based Knowledge Representation

  • RoBERTa's FFN hidden-state activations were treated as artificial neurons, and activations across layers were stacked for each emotion-specific prompt.The study used the resulting representations to identify neurons associated with conceptual attributes.Source: Collecting LLM-Based Knowledge… · p. 27

Core Affects

  • The study collected core-affect ratings from 30 participants for 27 emotions, using nine-point scales for valence and arousal.Participants rated each emotion after seeing its literal definition.Source: Core Affects · p. 27

Prototypical Expressions

  • Thirty participants rated how six prototypical expressions matched 27 emotion concepts using a nine-point scale and averaged male and female faces.The expression images came from the AKDEF stimulus set.Source: Prototypical Expressions · p. 28

Antecedent Appraisals

  • For antecedent appraisals, 299 participants each recalled an event causing one randomly assigned emotion from 27 and rated 38 event items; six appraisal factors were retained from the resulting matrix.Antecedent-appraisal reliability was calculated as ICC(1,k) and was significant at level .001 after FDR correction.Source: Antecedent Appraisals · pp. 28, 32

The LLM's Inference Accuracies and Human Rater Agreements on 27 Emotions

  • Across the 27 emotion tasks, inference accuracy was positively correlated with human rater agreement (Pearson's r(25) = .797, p < .001).Rater agreement was estimated using Cohen's Kappa.Source: Figure S1 · p. 29, Antecedent Appraisals · p. 29

Collecting LLM-Based Knowledge Representation

Antecedent Appraisals

  • Manipulating top-N attribute-specific neurons reduced inference performance relative to both no manipulation and random-neuron controls, indicating attribute-specific causal contributions.The reported contribution is the accuracy drop relative to random manipulation; the figure tests multiple neuron counts and reports FDR-corrected significance.Source: Antecedent Appraisals · pp. 36–38, Figure S9 · p. 37

Heterogeneity Degree in Conceptual Attributes' Causal Contribution to Various Emotion Inference Tasks

This section assesses whether an attribute's causal contribution varies across the 27 emotion-inference tasks by testing the distribution of its accuracy drops for multimodality.

Heterogeneity Degree in Conceptual Attributes' Causal Contribution to Various Emotion Inference Tasks

  • Heterogeneity was assessed by testing whether each conceptual attribute's causal contribution across 27 tasks was multimodal, using Hartigan's dip statistic with unimodality as the alternative hypothesis.Causal contribution is defined here as the accuracy drop relative to the random-manipulation condition.Source: Heterogeneity Degree in Conceptual… · p. 42
Loading 2302.09582v5…