MUMBAI, India, Aug. 10 -- Intellectual Property India has published a patent application (202621073888 A) filed by Symbiosis International Deemed University on June 12, 2026, for A System For Akkadian Cuneiform Translation Using Multilingual Transfer Learning.
Inventors include Dr. Sandeep Kumar; Aaditi Menon; Ayush Mahajan; and Arya Jain.
The application for the patent was published on July 31, 2026, under issue no. 31/2026.
Abstract: ABSTRACT A SYSTEM FOR AKKADIAN CUNEIFORM TRANSLATION USING MULTILINGUAL TRANSFER LEARNING The present invention provides a neural machine translation system (100) for translating Akkadian cuneiform transliterations into English. The system comprises a data ingestion module (110) for receiving heterogeneous parallel corpora, a text normalization pipeline (200) that applies Unicode NFC normalization, subscript digit mapping, lacuna marker replacement, scribal annotation stripping, and whitespace normalization, and a dataset staging module (210) that organizes data into three curriculum stages with document-order splitting. A pretrained mBART-50 multilingual encoder-decoder transformer model (300) with 680 million parameters is fine-tuned through three-stage curriculum training managed by a training engine (310), incorporating genre conditioning (350) through prefix tokens. An inference engine (400) employs beam search decoding with a no-repeat trigram constraint and a proper-name glossary post-processing module (410). The system achieves a geometric mean score of 29.99 combining BLEU and chrF++ on the validation set. [
Disclaimer: Curated by HT Syndication.
