Element Graph v2.1 — Multimodal Academic Document Retrieval

Typed element graphs derived from three multimodal academic document QA datasets: SPIQA, SciEGQA, and MMDocIR. Each document is decomposed into a typed graph with cross-modal edges (caption_of), reference edges (refer_to), and section-aware structural edges (contains, reading_next, section_next).

The graphs are designed for retrieval research:

• Training supervision via Graph-Relevance Contrastive Loss (GRCL)
• Inference-time graph propagation of retrieval scores
• Same edge-weight schema drives both (single source of truth)

Contents

Per-document section count

Schema

See SCHEMA.md for full schema definition (node types, edge types, attributes, target-attribute modifiers).

Provenance

• SPIQA: graph extracted from original LaTeX (\section{}, \ref{}) and all_figures metadata. Caption-figure pairs are author-provided.
• SciEGQA: graph derived from docling layout parsing + heuristic extraction.
• MMDocIR: graph derived from MMDocIR_layouts.parquet + heuristic extraction.

Reference (original datasets)

• SPIQA: Pramanick et al., NeurIPS 2024 D&B
• SciEGQA: arXiv:2511.15090
• MMDocIR: Dong et al., EMNLP 2025

Images

Image files are NOT included. Download from original dataset sources and look up by (doc_id, eid) — the eid is preserved as the figure filename (SPIQA) or layout_id (MMDocIR).

License

Each subset inherits the license of the original dataset. See per-subset README files.