Abstract
We introduce a context-weighted entity function analysis model that infers entity purpose and operational roles without relying on predefined categories. A multi-span encoder aggregates discursive and situational cues over long passages, while a function inference layer computes role distributions using contextual importance metrics. Experiments on three corpora—OpenBooks-ZS, IndustryReports-ZS, and MedText-ZS—show relative improvements of 18.2%, 15.9%, and 17.6% over paragraph-level embedding baselines. The model achieves a 31.4% reduction in context fragmentation and maintains 93.2% stability under document shuffling stress tests. Manual evaluation confirms a 22.7% increase in role inference coherence.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Anna Schneider, Felix Wagner (Author)