When Life Paths Cross: Extracting Human Interactions in Time and Space from Wikipedia

Zhongyang Liu, Ying Zhang, Xiangyi Xiao, Wenting Liu, Yuanting Zha, Haipeng ZhangarXiv 2025 · DOI · arXiv

Introduction

Historical data on real-world human interactions—the "who, whom, when, and where" of our collective past—is surprisingly difficult to quantify at scale. While digital footprints provide insights into modern social networks, historical records are often trapped in unstructured text. Extracting these interactions is a "needle in a haystack" problem: a person’s name, a year, and a city might appear in the same paragraph without having any functional relationship.

Figure 4: Examples of correct (a) and incorrect (b) spatio-temporal interaction extractions. In (b), the model must distinguish between a birth event and a marriage event to avoid false positives.

The FALCON Framework

To address this, researchers developed FALCON (AR-Bert utilizing Feature TrAnsfer and Multi-Task Learning). FALCON is designed to move beyond simple co-occurrence models, which often fail when entities are mentioned multiple times in different contexts.

The core of the framework is AR-Bert (Attention-enhanced R-Bert). Unlike standard models that might struggle with an entity appearing in multiple sentences within one biography, AR-Bert uses an attention mechanism to dynamically weight and aggregate these mentions. This creates a robust representation of an entity's role throughout a narrative.

Figure 6: The architecture of AR-Bert, which utilizes attention layers and mean pooling to aggregate entity representations before classification.

FALCON also employs Multi-Task Learning, training the model to simultaneously detect interactions (the main task) and individual "trajectories" (the auxiliary task of confirming if a single person was at a specific time and place). This approach treats trajectory detection as foundational evidence for interaction detection.

Large-Scale Extraction and Results

Using English Wikipedia as a primary source, FALCON extracted 685,966 interaction records. The model achieved an F1 score of 86.51%, significantly outperforming standard BERT (78.49%) and even specialized large language models like GPT-4o-mini (76.44%).

The researchers also tested FALCON's generalizability on Encyclopedia Britannica. The model's performance dropped by only 3.75%, suggesting it has learned the underlying logic of human interaction narratives rather than just the specific linguistic patterns of Wikipedia.

Quantifying History: A Case Study in Polarization

By converting prose into structured data, the authors conducted "macro-history" analyses. A notable case study involved 14,084 interactions between US politicians from 1960 to 2024. The data revealed a stark trend: adversarial inter-party interactions (such as debates and election competitions) rose from roughly 33% in the 1960s to 67% in the 2020s, while cooperative interactions declined.

Conclusion

FALCON represents a significant step forward in automated historical analysis. While currently limited by the language and cultural biases inherent in English Wikipedia, future work aims to expand the framework to multi-lingual sources and more granular interaction categorization. By mapping the intersections of human lives across time and space, we can gain new, data-driven perspectives on social dynamics and historical shifts.

Summary updated

BibTeX
@misc{liu2025when,
  author = {Zhongyang Liu and Ying Zhang and Xiangyi Xiao and Wenting Liu and Yuanting Zha and Haipeng Zhang},
  title = {{When Life Paths Cross: Extracting Human Interactions in Time and Space from Wikipedia}},
  year = {2025},
  eprint = {2510.00019},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.00019},
}
Figure