Section 3: Hierarchical Representation of Emotions

This interactive demo visualizes the results discussed in Section 3.2: Emotion Trees in LLMs. You can select one of four models from the dropdown menu: GPT-2 (1.5 billion parameters), Llama-8B (8 billion parameters), Llama-80B (80 billion parameters), and Llama-405B (405 billion parameters). As model size increases, more complex hierarchical structures emerge. Each node represents an emotion and is colored according to groups of emotions known to be related. The grouping of emotions by LLMs aligns closely with well-established psychological frameworks, as indicated by the consistent color patterns for emotions with shared parent nodes.

Section 4: Bias in Emotion Recognition

Misclassification Patterns on 135 Emotion Labels

This interactive demo visualizes the results discussed in Section 4. Building on our exploration of emotion representations in LLMs, we investigate how these representations and their resulting emotion perceptions are shaped by demographic attributes such as gender and socioeconomic status. Use the dropdown menu to select from various personas. Hover over an emotion label to reveal which emotions have been incorrectly classified under it.

Misclassification Patterns for Broad 6 Emotion Labels

This interactive demo presents the results discussed in Section 4, including the "User Study: Comparing Emotion Recognition in Humans and LLMs" (Section 4.1). It showcases recognition patterns for six broad emotions across various personas. Additionally, you can explore emotion recognition patterns for different demographics by selecting "Human " from the dropdown menu.

This interactive demo illustrates the results from the "Expanding to Realistic Datasets" section in Section 4.1. It showcases our extended analysis in a realistic setting using the GoEmotions dataset. You can compare the recognition patterns of LLM personas with human-labeled data.