GA-HEL: A Knowledge-Driven Expert System for Open-Set, Granularity-Aware Hierarchical Event Linking in Social Media

arXiv:2611.16194 · cs.IR · Submitted 2026-11-12 · Read on arXiv

cs.IR

Submitted: 2026-11-12

Updated: 2026-11-12

Comments: 16 pages, 6 figures

Code: https://github.com/SHbiao/GAHEL_release

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

The gist: Social-media event intelligence requires a system to determine whether a post belongs to a registered event space, identify the supported hierarchy path…

Abstract

Social-media event intelligence requires a system to determine whether a post belongs to a registered event space, identify the supported hierarchy path, and stop at the deepest level warranted by the evidence. We formulate this requirement as open-set hierarchical event linking over frozen Storyline--Episode--Atomic Event Cards. GA-HEL implements a three-stage, knowledge-driven expert-system funnel: open-set scope control limits candidate exposure before hierarchical reasoning, hierarchy-aware decoders select a supported path and depth, and deterministic legal-path validation checks each record before database-compatible use. A 200-post Pilot supports task and mechanism development, while a separately annotated 497-post Expansion set evaluates transfer across four held-out Storylines and a hard-negative pool. Method implementations, prompts, Event Cards, candidate order, and six systems' predictions were frozen before Expansion Gold annotation. The resulting evaluation combines end-to-end hierarchical matching, open-set rejection, structural validation, ablation, cross-Storyline analysis, and an independent expert reannotation of a fixed audited subset. GA-HEL remains competitive with strong global full-card reasoning while providing explicit scope control and hierarchy-valid structured records. Error analysis identifies unsupported descent from Episode-level evidence to Atomic Events as the dominant remaining challenge. The task definition, frozen resource, and auditable protocol provide a reproducible foundation for scope-aware event intelligence. Code, Event Cards, labels, frozen predictions, and evaluation scripts are available at https://github.com/SHbiao/GAHEL release.

Sources

Related papers