Code, the Karaka-VG resource (2,937 silver-labeled, image-grounded Hindi karaka role frames), and result files supporting the paper "Karaka-Grounded Visual Semantic Role Labeling for a Case-Marking Language," submitted to Language Resources and Evaluation. Includes the Part I Hindi karaka tagger, the Part II corpus-scale extraction and round-trip consistency filter that builds Karaka-VG, the POS-tagging audit, the Part IV classical- and CNN-feature pixels-only visual classifiers, and the GSR-translation baseline built on a real GSRTR checkpoint. Hindi Visual Genome, the underlying corpus and image archive, is not redistributed here; see the README for how the release keys back to it.
Keywords: add each of these as a separate tag: semantic role labeling, karaka theory, Hindi, low-resource language processing, visual grounding, grounded situation recognition, Panini, computational linguistics