When Trump Shakes Biden’s Hand: Tracking Real-World Interactions in Time and Space from Textual Records

ACL ARR 2026 January Submission5293 Authors

Keywords: real-world interaction extraction, real-world interaction analysis, political polarization

Abstract: Interpersonal interactions reveal critical insights across political, cultural, and historical domains. Existing research predominantly analyzes online interactions (e.g., retweets, mentions, follows), but these virtual signals poorly capture real-world encounters (e.g., collaborations, negotiations, debates) and suffer from biases (fake accounts, bots, unrepresentative samples) while spanning only recent decades. In contrast, real-world interactions in historical texts (encyclopedias, news, books) capture actual encounters across centuries with temporal and spatial contexts. However, extracting structured interaction quadruplets (Person1, Person2, Time, Location) is challenging due to complex spatio-temporal dependencies across scattered and indirect mentions. We construct the WikiInteraction dataset (4,507 annotated quadruplets) and propose FALCON, integrating AR-BERT, multi-task learning, and feature transfer to model these dependencies. FALCON achieves 86.51% F1, outperforming all baselines including GPT-o1, while incurring significantly lower computational costs than LLMs, and demonstrates cross-domain applicability on Encyclopedia Britannica and news corpora. Applied at scale, it extracts 658,966 Wikipedia quadruplets over 1,000 years and 77,836 news quadruplets—the largest spatio-temporal interaction dataset to date. A case study on US political polarization (1960-2024, 7,891 figures) reveals long-term geo-temporal trends invisible in online data. Code and datasets are available.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: Computational Social Science and Cultural Analytics, Resources and Evaluation

Contribution Types: Data resources, Data analysis

Languages Studied: English

Submission Number: 5293