The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support. If this is not possible, please open a discussion for direct help.

Dataset Card for the High-Level Dataset

Dataset Description

The High-Level (HL) dataset aligns object-centric descriptions from COCO with high-level descriptions crowdsourced along 3 axes: scene, action, rationale

The HL dataset contains 14997 images from COCO and a total of 134973 crowdsourced captions (3 captions for each axis) aligned with ~749984 object-centric captions from COCO.

Each axis is collected by asking the following 3 questions:

  1. Where is the picture taken?
  2. What is the subject doing?
  3. Why is the subject doing it?

The high-level descriptions capture the human interpretations of the images. These interpretations contain abstract concepts not directly linked to physical objects. Each high-level description is provided with a confidence score, crowdsourced by an independent worker measuring the extent to which the high-level description is likely given the corresponding image, question, and caption. The higher the score, the more the high-level caption is close to the commonsense (in a Likert scale from 1-5).

Supported Tasks

  • image captioning
  • visual question answering
  • multimodal text-scoring
  • zero-shot evaluation



Dataset Structure

The dataset is provided with images from COCO and two metadata jsonl files containing the annotations

Data Instances

An instance looks like this:

  "file_name": "COCO_train2014_000000138878.jpg",
  "captions": {
    "scene": [
      "in a car",
      "the picture is taken in a car",
      "in an office."
    "action": [
      "posing for a photo",
      "the person is posing for a photo",
      "he's sitting in an armchair."
    "rationale": [
      "to have a picture of himself",
      "he wants to share it with his friends",
      "he's working and took a professional photo."
    "object": [
      "A man sitting in a car while wearing a shirt and tie.",
      "A man in a car wearing a dress shirt and tie.",
      "a man in glasses is wearing a tie",
      "Man sitting in the car seat with button up and tie",
      "A man in glasses and a tie is near a window."
  "confidence": {
    "scene": [
    "action": [
    "rationale": [
  "purity": {
    "scene": [
    "action": [
    "rationale": [
  "diversity": {
    "scene": 25.965358893403383,
    "action": 32.713305568898775,
    "rationale": 2.658757840479801

Data Fields

  • file_name: original COCO filename
  • captions: Dict containing all the captions for the image. Each axis can be accessed with the axis name and it contains a list of captions.
  • confidence: Dict containing the captions confidence scores. Each axis can be accessed with the axis name and it contains a list of captions. Confidence scores are not provided for the object axis (COCO captions).t
  • purity score: Dict containing the captions purity scores. The purity score measures the semantic similarity of the captions within the same axis (Bleurt-based).
  • diversity score: Dict containing the captions diversity scores. The diversity score measures the lexical diversity of the captions within the same axis (Self-BLEU-based).

Data Splits

There are 14997 images and 134973 high-level captions split into:

  • Train-val: 13498 images and 121482 high-level captions
  • Test: 1499 images and 13491 high-level captions

Dataset Creation

The dataset has been crowdsourced on Amazon Mechanical Turk. From the paper:

We randomly select 14997 images from the COCO 2014 train-val split. In order to answer questions related to actions and rationales we need to ensure the presence of a subject in the image. Therefore, we leverage the entity annotation provided in COCO to select images containing at least one person. The whole annotation is conducted on Amazon Mechanical Turk (AMT). We split the workload into batches in order to ease the monitoring of the quality of the data collected. Each image is annotated by three different annotators, therefore we collect three annotations per axis.

Curation Rationale

From the paper:

In this work, we tackle the issue of grounding high-level linguistic concepts in the visual modality, proposing the High-Level (HL) Dataset: a
V&L resource aligning existing object-centric captions with human-collected high-level descriptions of images along three different axes: scenes, actions and rationales. The high-level captions capture the human interpretation of the scene, providing abstract linguistic concepts complementary to object-centric captions used in current V&L datasets, e.g. in COCO. We take a step further, and we collect confidence scores to distinguish commonsense assumptions from subjective interpretations and we characterize our data under a variety of semantic and lexical aspects.

Source Data

  • Images: COCO
  • object axis annotations: COCO
  • scene, action, rationale annotations: crowdsourced
  • confidence scores: crowdsourced
  • purity score and diversity score: automatically computed

Annotation process

From the paper:

Pilot: We run a pilot study with the double goal of collecting feedback and defining the task instructions. With the results from the pilot we design a beta version of the task and we run a small batch of cases on the crowd-sourcing platform. We manually inspect the results and we further refine the instructions and the formulation of the task before finally proceeding with the annotation in bulk. The final annotation form is shown in Appendix D.

Procedure: The participants are shown an image and three questions regarding three aspects or axes: scene, actions and rationales i,e. Where is the picture taken?, What is the subject doing?, Why is the subject doing it?. We explicitly ask the participants to use their personal interpretation of the scene and add examples and suggestions in the instructions to further guide the annotators. Moreover, differently from other VQA datasets like (Antol et al., 2015) and (Zhu et al., 2016), where each question can refer to different entities in the image, we systematically ask the same three questions about the same subject for each image. The full instructions are reported in Figure 1. For details regarding the annotation costs see Appendix A.

Who are the annotators?

Turkers from Amazon Mechanical Turk

Personal and Sensitive Information

There is no personal or sensitive information

Considerations for Using the Data

[More Information Needed]

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

From the paper:

Quantitying grammatical errors: We ask two expert annotators to correct grammatical errors in a sample of 9900 captions, 900 of which are shared between the two annotators. The annotators are shown the image caption pairs and they are asked to edit the caption whenever they identify a grammatical error. The most common errors reported by the annotators are:

  • Misuse of prepositions
  • Wrong verb conjugation
  • Pronoun omissions

In order to quantify the extent to which the corrected captions differ from the original ones, we compute the Levenshtein distance (Levenshtein, 1966) between them. We observe that 22.5% of the sample has been edited and only 5% with a Levenshtein distance greater than 10. This suggests a reasonable level of grammatical quality overall, with no substantial grammatical problems. This can also be observed from the Levenshtein distance distribution reported in Figure 2. Moreover, the human evaluation is quite reliable as we observe a moderate inter-annotator agreement (alpha = 0.507, (Krippendorff, 2018) computed over the shared sample.

Dataset Curators

Michele Cafagna

Licensing Information

The Images and the object-centric captions follow the COCO terms of Use The remaining annotations are licensed under Apache-2.0 license.

Citation Information

  title={{HL} {D}ataset: {V}isually-grounded {D}escription of {S}cenes, {A}ctions and
  author={Cafagna, Michele and van Deemter, Kees and Gatt, Albert},
  booktitle={Proceedings of the 16th International Natural Language Generation Conference (INLG'23)},
address = {Prague, Czech Republic},
Downloads last month
Edit dataset card

Models trained or fine-tuned on michelecafagna26/hl