Emergence of Hierarchical Emotion Organization in Large Language Models

2024-10-01 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

Large Language Models (LLMs) develop hierarchical emotion organizations that align with human psychological models, a study by Zhao et al. reveals. Inspired by emotion wheels, researchers analyzed probabilistic dependencies in LLM outputs, finding that larger models, such as Llama 3.1 405B, form more complex and nuanced emotion trees compared to smaller models like Llama 3.1 8B. The Llama 405B model achieved an overall emotion classification accuracy of 15.2% for 135 emotion words and 87.1% for six broad categories. The study also uncovered systematic biases in LLM emotion recognition across diverse socioeconomic personas. For instance, Llama 405B showed lower accuracy for underrepresented groups like low-income Black female personas, with misclassifications compounding for intersectional identities. These biases, including misclassifying negative emotions as "shame" for Asian personas or "frustration" for physically disabled personas, mirror patterns observed in human studies, suggesting LLMs internalize aspects of social perception. A strong positive correlation ($r=0.84$, $p<0.001$) was found between the complexity (path length) of an LLM's emotion tree and its recognition accuracy.

Key takeaway

For AI Scientists and AI Ethicists developing conversational agents, you must prioritize evaluating your models' emotional understanding beyond simple classification. Recognize that LLMs inherently form hierarchical emotion structures and replicate human biases, especially for underrepresented groups. Use cognitively-grounded evaluation methods, like the proposed tree-construction algorithm, to uncover these internal representations and systematically test for demographic misclassification patterns. This proactive approach is crucial for mitigating risks of manipulation and ensuring ethical, equitable deployment of emotionally intelligent AI.

Key insights

LLMs spontaneously organize emotions hierarchically, mirroring human psychology, but also replicate human biases in emotion recognition.

Principles

• Emotion hierarchies emerge and grow more complex with LLM scale.
• LLM emotion tree geometry predicts recognition performance.
• LLMs internalize and replicate human social perception biases.

Method

A novel tree-construction algorithm analyzes LLM logit activations to infer hierarchical emotion structures by computing a "matching matrix" ($C=Y^{T}Y$) and conditional probabilities to define parent-child relationships between emotions.

In practice

• Use cognitively-grounded theories for LLM evaluation.
• Analyze LLM logit activations to uncover internal concept structures.
• Evaluate LLM emotion recognition across diverse personas to detect biases.

Topics

• Large Language Models
• Emotion Recognition
• Hierarchical Emotion Organization
• AI Bias
• Cognitive Psychology
• Model Evaluation

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• Mechanistic Decoding of Cognitive Constructs in LLMs — LLMs encode complex emotions like jealousy as structured linear combinations of psychological antecedents.
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.
• Beyond Sentiment Classification: A Generative Framework for Emotion Intensity Evaluation in Text — Generative LLMs fine-tuned on continuous emotion intensity scores outperform classification and generalize to unseen affective dimensions.
• Do Emotions Influence Moral Judgment in Large Language Models? — LLMs' moral judgments are systematically biased by emotional context, unlike human judgments.
• The Mind Behind the Machine: A Deep Look at How Large Language Models Actually Work — Transformers utilize attention for context-aware representations, enabling LLMs to learn language statistics and scale for emergent abilities.
• NICE: A Theory-Grounded Diagnostic Benchmark for Social Intelligence of LLMs — NICE provides a theory-grounded diagnostic benchmark to identify specific social intelligence weaknesses in LLMs, crucial for their application in social contexts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.