Tracking Evolving Relationship Between Characters in Books in the Era of Large Language Models

AbstractThis work aims to assess the zero-shot social reasoning capabilities of LLMs by proposing various strategies based on the granularity of information used to track the fine-grained evolution in the relationship between characters in a book. Without gold annotations, we thoroughly analyze the agreements between predictions from multiple LLMs and manually examine their consensus at a local and global level via the task of trope prediction. Our findings reveal low-to-moderate agreement among LLMs and humans, reflecting the complexity of the task. Analysis shows that LLMs are sensitive to subtle contextual changes and often rely on surface-level cues. Humans, too, may interpret relationships differently, leading to disagreements in annotations.