Are Human Conversations Special? A Large Language Model Perspective

TL;DR

Are Human Conversations Special? A Large Language Model Perspective investigates whether human-human conversations require different attention strategies than web, code, and mathematics data. The authors quantify attention distance $\overline{D}_{\alpha}$, attention entropy $\text{Entropy}_{\alpha}$, and a novel Interdependency Factor (IF), using LLaMa-2 13b as a representative decoder-only model, and they visualize hidden-state representations with t-SNE. They find that human-human conversations induce longer-range dependencies in deeper layers, higher attention dispersion, and stronger interdependencies, while authentic conversational data is scarce in web-scale pretraining. They argue for domain-specialized models and larger, higher-quality conversational data to bridge the gap in modeling natural dialogue.

Abstract

This study analyzes changes in the attention mechanisms of large language models (LLMs) when used to understand natural conversations between humans (human-human). We analyze three use cases of LLMs: interactions over web content, code, and mathematical texts. By analyzing attention distance, dispersion, and interdependency across these domains, we highlight the unique challenges posed by conversational data. Notably, conversations require nuanced handling of long-term contextual relationships and exhibit higher complexity through their attention patterns. Our findings reveal that while language models exhibit domain-specific attention behaviors, there is a significant gap in their ability to specialize in human conversations. Through detailed attention entropy analysis and t-SNE visualizations, we demonstrate the need for models trained with a diverse array of high-quality conversational data to enhance understanding and generation of human-like dialogue. This research highlights the importance of domain specialization in language models and suggests pathways for future advancement in modeling human conversational nuances.

Figures (14)

• Figure 1: Heatmap of the Attention Distance Difference matrix ($\Delta \overline{D}_{\alpha}$) calculated with the first domain fixed as web data, and the second domain as human-human conversations, code, and math respectively. Higher values of difference in attention distances in deeper layers (left) indicate that human-human conversations demand deeper modeling of long-term contextual relationships than general web data. The comparison with code (center) indicates higher distances in the first half of the layers, but lower values in deeper layers, indicating more long-term relationships in structural aspects, but localized contextual relationships. However, the attention distances are quite spread across math (right) in comparison to the web.
• Figure 2: Attention Distance Difference by Layer across all heads calculated with the first domain as web data, and the second domain as human-human conversations (left), code (middle), and math (right) respectively.
• Figure 3: Attention Distance Difference by Head across all layers calculated with the first domain as web data, and the second domain as human-human conversations (left), code (middle), and math (right) respectively.
• Figure 4: Heatmap of mean attention entropy for web, human-human conversations, code, and math domains respectively. Higher values indicate more attention diffusion. Human-human conversations show the highest diffusion in attention, especially in the middle and end layers. Attention diffusion in web, code, and math domains is similar, with small differences.
• Figure 5: Mean attention entropy by layer across all heads with first token attention removed for web, human-human conversations, code, and math domains respectively. Higher values indicate more attention diffusion in the layer.