metadata

annotations_creators:
  - machine-generated
  - crowdsourced
language_creators:
  - machine-generated
  - crowdsourced
language:
  - ca
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: CAT ManyNames
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - image-classification
task_ids:
  - multi-label-image-classification

Dataset Card for CAT ManyNames

Table of Contents
Dataset Description
Dataset Structure
Dataset Creation
Considerations for Using the Data
Additional Information

Dataset Description

Repository: https://github.com/MarDominguezOrfila/CAT-ManyNames
Point of Contact: mar.dominguez01@estudiant.upf.edu

Dataset Summary

CAT ManyNames is the Catalan version of the ManyNames dataset suitable for training Language & Vision models in the task of object naming. The corpus consists of more than 23K images and their corresponding annotations.

The human-annotated test set has been built to evaluate the quality of the CAT ManyNames dataset. Its corpus consists of 1,072 images and their corresponding annotations (ca. 10 annotations per image).

CAT ManyNames is licensed under a Creative Commons Attribution 4.0 International License.

Supported Tasks and Leaderboards

Object Naming, Language & Vision Model.

Languages

The dataset is in Catalan (ca-ES).

Dataset Structure

Data Instances

'responses' : {"guepard": 27, "lleopard": 3, "animal": 2, "gat": 2, "tigre": 2}

Data Fields

Both CAT ManyNames and its human-annotated test set are provided in a tab-separated text file (.tsv). The first rows contain the column labels. Nested data is stored as Python dictionaries (i.e., "{key: value}"). The columns are labelled as follows (the most important columns are listed first):

responses: Correct responses and their counts
topname: The most frequent name of the object in the largest cluster
domain: The MN domain of the object
incorrect (not available for the human-annotated test set): Incorrect responses and their counts
singletons (not available for the human-annotated test set): All responses which were given only once and are not synonyms or - hypernyms of the top name (these are included in responses)
total_responses: Sum count of correct responses
split: Use of the image in training vs. test vs. validation
vg_object_id: The VisualGenome id of the object
vg_image_id: The VisualGenome id of the image
topname_agreement (only available for the test split): The number of responses for the top name divided by the number of total responses
jaccard_similarity (only available for the test split): Jaccard similarity index of the responses column in CAT ManyNames and its human-annotated test set

Data Splits

Test: 1,072 images

```
Val: 1,110 images
```
- ```
Train: 21,503 images
```

Dataset Creation

Curation Rationale

We created this corpus to contribute to the development of multimodal models in Catalan, a low-resource language.

Source Data

Initial Data Collection and Normalization

The original visual data comes from VisualGenome. The objects were categorized in the categories of people, animals_plants, vehicles, food, home, buildings, and clothing.

Who are the source language producers?

The original ManyNames dataset.

Annotations

Annotations for the CAT ManyNames were obtained by performing a machine translation of the ManyNames dataset, originally in English. The test set was humanly annotated.

Annotation process

[N/A]

Who are the annotators?

The human-annotated test set gathered 220 Catalan native volunteer participants.

Personal and Sensitive Information

There is no sensitive information in this dataset.

Considerations for Using the Data

Social Impact of Dataset

We hope this corpus contributes to the development of multimodal models in Catalan, a low-resource language.

Discussion of Biases

We have not applied any steps to reduce the impact of biases possibly present in the data.

Other Known Limitations