Efficient Neural Machine Translation for Lost Epic Text Reconstruction: A Case Study of a Chapter-based Gesar Epic

The Epic of King Gesar is a multilingual cultural treasure widely transmitted in Tibetan, Chinese, Mongolian, and other ethnic languages. In fact, a substantial portion of the epic texts remains lost, creating an urgent need for rescue, preservation, systematic organization, collation, and publication. Efficient cross-lingual translation can help compensate for missing content and clarify the epic’s overall structure and developmental trajectory. Taking the chapter-based editions as an example, six distinct versions currently coexist, while the classic Guide chapter-based edition survives only in Chinese translation. To reconstruct this missing version, this study first constructs a large-scale Tibetan–Chinese parallel corpus of the Epic of King Gesar, containing 100,000 aligned sentence pairs. Then an efficient neural machine translation model integrating epic-specific entity information is proposed, and this model is applied to perform Chinese-to-Tibetan translation of the Guide edition. A comparison with complete existing chapter-based editions shows that the reconstructed text achieves a similarity score of 82.96%, resulting in a reference draft of the Tibetan source manuscript. In addition, we collate the reconstructed edition against other extant versions, further exploring their textual differences, primary-secondary relationships, and evolutionary pathways.