CarakaBinary: A Binary-Labelled Dataset from Early Āyurvedic Texts for the Identification of Compound Words in Classical Sanskrit.

Charaka-Binary, the presented domain-specific Ayurvedic textual dataset in this article, is intended to support natural language processing (NLP) tasks in the Sanskrit language, particularly in analyzing compound words. Charaka-Binary includes 31,179 instances comprising 23,224 non-compound words and 7,955 compound words in Devanāgarī script along with the relevant numerical labels, including “0” indicating non-compound words and “1” denoting compound words to support future computational tasks. The dataset provides a resource for extracting compound words from Sanskrit text. We utilized the publicly accessible ancient Ayurvedic literature, Carakasaṃhitā. Among the 7 chapters of Carakasaṃhitā, we employed the first chapter along with its 30 lessons and fourth chapters’ two lessons as the foundation for dataset collection. The first phase of the dataset collection was to convert 228 pages of Carakasaṃhitā into individual .png files suitable for text extraction, while the second phase included the use of the Optical Character Recognition (OCR) technique to extract the text, followed by proofreading. Then the sentence segmentation step was carried out by word-level labeling into compound and non-compound word categories. The preprocessing steps, including proofreading, segmentation, and labeling, are carried out with the help of domain experts. Caraka-Binary can be used in both computational tasks, which include training models to identify Sanskrit compound words, as well as non-computational purposes, which encompass classroom teaching on Sanskrit compound words, comprehending their types, and analyzing compound patterns.