AbsVis: Benchmarking How Humans and Vision-Language Models "See" Abstract Concepts in Images

• Paper: AbsVis – Benchmarking How Humans and Vision-Language Models "See" Abstract Concepts in Images (EMNLP 2025, Main Conference)
• Authors: Tarun Tater, Diego Frassinelli, Sabine Schulte im Walde
• Code & raw CSVs: github.com/TarunTater/AbsVis_Emnlp2025

Abstract concepts like mercy and peace often lack clear visual grounding. AbsVis is a dataset of 675 images annotated with concept–explanation pairs from humans and two vision-language models (Qwen2-VL-7B-Instruct and LLaVA-NeXT llava-v1.6-mistral-7b-hf). Each attributed concept comes with a free-text explanation of why the image evokes it.

The images depict 225 English nouns spanning the concreteness scale (75 abstract, 75 mid-range, 75 concrete; three images per noun), sourced from YFCC100M.

Key findings from the paper:

• Humans attribute more varied concepts than VLMs.
• Concepts attributed by both humans and VLMs are the most preferred.
• Explanations clarify and strengthen the perceived attributions, for both humans and VLMs.
• VLMs can approximate human preferences, and DPO fine-tuning on these preferences improves alignment with preferred concept–explanation pairs.

Quick start

from datasets import load_dataset

images = load_dataset("Tarun/AbsVis", "images", split="train")             # 675 images + noun + attribution
annotations = load_dataset("Tarun/AbsVis", "annotations", split="train")   # concept–explanation pairs
nouns = load_dataset("Tarun/AbsVis", "nouns", split="train")               # 225 nouns with concreteness ratings

ex = images[0]
ex["image"]          # PIL image
ex["noun"], ex["category"], ex["concreteness"]

# all human explanations for this image
human = annotations.filter(lambda r: r["image_id"] == ex["image_id"] and r["source"] == "human")

Configs and fields

images (default): one row per image

field	description
image	Original YFCC100M image (unmodified file)
image_id	YFCC100M photo hash; joins with annotations.image_id
noun	Target noun the image was selected for
concreteness	Concreteness rating of the noun (1 = most abstract, 5 = most concrete; Brysbaert et al., 2014)
category	abstract, mid-range, or concrete
yfcc_photo_id	Flickr / YFCC100M photo ID
flickr_owner_id, flickr_owner_name	Flickr account of the photographer (for attribution)
flickr_title	Original Flickr title
flickr_page_url	Original Flickr page
license_name, license_url	The image's Creative Commons license

annotations: one row per concept–explanation pair

field	description
image_id	Image the concept was attributed to
noun	Target noun of that image
source	human, qwen (Qwen2-VL-7B-Instruct), or llava (llava-v1.6-mistral-7b-hf)
concept	Attributed (abstract) concept
explanation	Free-text explanation of how the image conveys the concept

Human annotations come from 220 native English speakers recruited via Prolific: five annotators per image, three concept–explanation pairs each, giving up to 15 per image. One-word explanations were removed during quality control, so a few images have fewer than 15 (10 for 1 image, 13 for 7, 14 for 28).

nouns: the 225 target nouns

noun, concreteness, category.

Licensing

AbsVis combines two kinds of content, each under its own license:

• Annotations, nouns and metadata (everything except the image files) are released under CC BY 4.0.
• Images are redistributed unmodified from YFCC100M and keep their original Flickr Creative Commons license, given per image in license_name / license_url. Only these four licenses occur: CC BY 2.0, CC BY-SA 2.0, CC BY-NC 2.0, CC BY-NC-SA 2.0.

Some images are NonCommercial, so commercial use of the full image set is not permitted. If you reuse or display an image, credit its photographer (flickr_owner_name) and link flickr_page_url and license_url. See LICENSE.md.

Citation

If you use AbsVis, please cite:

@inproceedings{tater-etal-2025-absvis,
    title = "{A}bs{V}is {--} Benchmarking How Humans and Vision-Language Models ``See'' Abstract Concepts in Images",
    author = "Tater, Tarun  and
      Frassinelli, Diego  and
      Schulte im Walde, Sabine",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.417/",
    doi = "10.18653/v1/2025.emnlp-main.417",
    pages = "8260--8281",
    ISBN = "979-8-89176-332-6"
}

The images come from YFCC100M; please also cite:

@article{thomee2016yfcc100m,
    title = {{YFCC100M}: The New Data in Multimedia Research},
    author = {Thomee, Bart and Shamma, David A. and Friedland, Gerald and Elizalde, Benjamin and Ni, Karl and Poland, Douglas and Borth, Damian and Li, Li-Jia},
    journal = {Communications of the ACM},
    volume = {59},
    number = {2},
    pages = {64--73},
    year = {2016}
}

Related work

• Tater et al. (2024). Unveiling the Mystery of Visual Attributes of Concrete and Abstract Concepts: Variability, Nearest Neighbors, and Challenging Categories. EMNLP 2024.
• Tater et al. (2024). Evaluating Semantic Relations in Predicting Textual Labels for Images of Abstract and Concrete Concepts. CMCL 2024.
• Tater et al. (2022). Concreteness vs. Abstractness: A Selectional Preference Perspective. AACL-IJCNLP 2022 SRW.

Contact

Tarun Tater: tarun.tater@ims.uni-stuttgart.de