Language-specific representation of emotion-concept knowledge causally supports emotion inference

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Language is central to human emotional experience and communication.
• The role of language in shaping emotion understanding remains debated.
• Artificial intelligence, specifically large language models (LLMs), offers a novel approach to studying these questions.

Purpose of the Study:

• To investigate whether language-based representations of emotion causally influence an AI's ability to infer emotional meaning.
• To determine if LLMs can learn and utilize emotion concepts without sensory-motor grounding.
• To explore the contribution of language-derived emotion knowledge to generative emotion inference.

Main Methods:

• Utilized large language models (LLMs) as a computational tool.
• Identified distinct artificial neuron populations representing 14 attributes of human emotion concepts within the LLM.
• Manipulated these attribute-related neurons to assess their causal role in emotion inference.

Main Results:

• LLMs represent human emotion concept attributes using distinct neural populations.
• Manipulating these neural populations demonstrated a causal link between emotion concept knowledge and generative emotion inference.
• Performance degradation was linked to the importance of specific attributes in human mental spaces.

Conclusions:

• LLMs can learn about emotions independently of sensory-motor experiences.
• Language-derived emotion concept knowledge significantly contributes to an AI's capacity for emotion inference.
• Findings provide a proof-of-concept for the causal role of language in emotion understanding.