pengyu3/beyond-images-enriched

Beyond Images — Enriched MMKG Datasets

Textual enrichment of three public multi-modal knowledge graph (MMKG) benchmarks — MKG-W, MKG-Y, and DB15K — produced by the Beyond Images pipeline (ESWC 2026): entity images (both the originals shipped with each benchmark and newly crawled Wikipedia images) are converted into textual descriptions with BLIP-2, giving every entity a language view of its visual content.

Feeding these descriptions to standard MMKG completion models (MMRNS, MyGO, NativE, AdaMF) as an enhanced text modality improves link prediction by up to +7% Hits@1, and by up to +333% Hits@1 on entities whose images are logos or symbols. See the paper for details.

Usage

from datasets import load_dataset

# One row per image caption (default config)
captions = load_dataset("pengyu3/beyond-images-enriched", "captions")

# One row per entity, with all captions merged into one text field
entities = load_dataset("pengyu3/beyond-images-enriched", "entities")

# Provenance for every newly crawled Wikipedia image (URLs + file metadata)
metadata = load_dataset("pengyu3/beyond-images-enriched", "image_metadata")

# Filter to one benchmark / image source
mkgw_new = captions["train"].filter(
    lambda row: row["dataset"] == "MKG-W" and row["source"] == "new"
)

Configs

captions — 918,827 rows

One row per image caption (BLIP-2, prompt: "Describe the scene, objects, colors, and other details in detail.").

Column	Description
dataset	MKG-W, MKG-Y, or DB15K
source	original (benchmark's own images) or new (crawled from Wikipedia)
entity_name	DBpedia entity name
entity_qid	Wikidata QID
dbpedia_url / wikidata_url	entity URIs
image_id	<QID>_<index> image identifier
caption	generated description

entities — 85,684 rows

One row per (dataset, source, entity): all captions concatenated into merged_descriptions, plus num_images. Entities without images have an empty string and num_images = 0.

image_metadata — 316,352 rows

Provenance for every newly crawled image: page_url (Wikipedia file page), image_url, extracted description / date / author / source_field, and the full file-page summary table as metadata_json.

Statistics

	MKG-W	MKG-Y	DB15K
Entities	15,000	15,000	12,842
Captions (original images)	36,462	42,229	497,636
Captions (new images)	107,596	56,675	178,229
Crawled image metadata	81,324	56,680	178,348

Provenance and attribution

• Source benchmarks: MKG-W and MKG-Y from MMRNS; DB15K from mmkb (BSD-3-Clause). Retain their attribution requirements when redistributing derived artifacts.
• New images were collected from Wikipedia / Wikimedia Commons; per-image attribution is in the image_metadata config (author, source, and the file page URL). This dataset redistributes generated text, not the images themselves.
• raw/img_text_summary.zip is the bit-exact original release, canonical on Zenodo: 10.5281/zenodo.14847095 (MD5 e5464dcb76501c10b692afd309139a50).

Links

• Code & pipeline: https://github.com/pengyu-zhang/Beyond-Images
• Live demo (human-auditing tool): https://pengyu-zhang.github.io/Beyond-Images/
• Paper: arXiv:2603.16974 · DOI 10.1007/978-3-032-25156-5_5 (ESWC 2026)

Citation

@inproceedings{zhang2026beyond,
  title     = {Are a Thousand Words Better Than a Single Picture? Beyond Images --
               A Framework for Multi-Modal Knowledge Graph Dataset Enrichment},
  author    = {Zhang, Pengyu and Zaporojets, Klim and Liu, Jie and Huang, Jia-Hong and Groth, Paul},
  booktitle = {The Semantic Web -- ESWC 2026},
  publisher = {Springer},
  year      = {2026},
  doi       = {10.1007/978-3-032-25156-5_5}
}