RhetConnect

Independent research, extending earlier work on rhetorical structure for multimedia documents carried out at CERIST. RhetConnect addresses a structural safety problem in multimedia document adaptation: importance-based selection can silently orphan units when media-type constraints exclude their rhetorical anchor, producing output that is complete in bytes but incoherent in meaning.

Contributions

• A formal definition of rhetorical connectivity for multimedia documents modeled as Extended Rhetorical Graphs.
• A proof that naive weight-threshold selection is structurally unsafe in the general case.
• RhetConnect, a polynomial-time algorithm with proved guarantees for rhetorical connectivity, budget compliance, and media-type compliance.
• Dual empirical validation: 270 synthetic experimental pairs and the RhetConnect Corpus v1.0, a novel 40-document annotated corpus of real multimedia educational documents.

RhetConnect Corpus v1.0

The first publicly available corpus of multimedia educational documents annotated with cross-media rhetorical structure: 40 documents, 559 elementary units, 396 rhetorical relations, spanning 10 academic domains. Sourced from CK-12 Foundation open educational resources plus original constructed documents. Released under CC BY 4.0 (annotations) and MIT (code).

Code and corpus on GitHub →
Browse the corpus interactively →