dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
splits:
- name: train
num_bytes: 18061981.698
num_examples: 6413
download_size: 18154825
dataset_size: 18061981.698
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
Dataset Card for English to Kanji Diffusion
This dataset enables the conversion of English meanings into their corresponding Kanji characters, including generating SVG representations and converting these to various image formats for use in machine learning frameworks.
Dataset Details
Dataset Description
The English to Kanji Diffusion dataset is designed for image generation tasks where the input is an English word and the output is the corresponding Kanji character image. It includes parsed data from XML sources and processed images in SVG, PNG, and JPG formats.
- Curated by: Sylvain Lapeyrade
- Language(s) (NLP): English, Japanese (Kanji)
- License: MIT License
Dataset Sources
- Repository: GitHub Repository
- Demo: HuggingFace Model with APLI
Uses
Direct Use
This dataset is suitable for developing machine learning models specifically for text-to-image translation tasks, educational tools, and design applications that require the visual representation of Kanji characters based on English descriptions.
Out-of-Scope Use
The dataset is not designed for natural language processing tasks that involve understanding or generating human language text beyond the scope of translating single words to Kanji characters.
Dataset Structure
Dataset Fields
image: Contains the image file of the Kanji character in JPG format (can easily be made ot other format through the GitHub code).text: Contains the English word corresponding to the Kanji character.
Splits
- Train Split: Contains 6,413 examples, focusing on a diverse set of Kanji characters and their English meanings.
Dataset Creation
Curation Rationale
The dataset was created to facilitate the development of models capable of understanding the symbolic representation of Kanji characters based on English descriptions, aiding in educational and graphical applications.
Source Data
Data Collection and Processing
The data were collected from the KANJIDIC and KanjiVG projects, processed into SVG files, and then converted into image formats suitable for machine learning applications. The entire process emphasizes accuracy in the representation of Kanji characters.
Who are the source data producers?
The data is sourced from public Kanji character databases and projects dedicated to the dissemination and educational use of Kanji characters.
Annotations
Annotation process
The dataset does not contain annotations beyond the automated generation and labeling of images with corresponding English descriptions. Descriptions in other languages could have been possible as they are also in the XML file.
Who are the annotators?
The data and images are generated and labeled automatically using scripts designed for this purpose.
Bias, Risks, and Limitations
The dataset does not comprehensively cover all Kanji characters and their nuances. Users should be aware of the dataset's focus on basic and commonly used Kanji characters. Also the phrasing "a kanji meaning..." is quite specific.