الذكاء الاصطناعي في فك رموز اللغة القديمة

Ithaca predicts missing text and geographic and chronological attribution for ancient Greek inscriptions. It uses textual context to produce candidate restorations, which scholars assess against the inscription and historical evidence. A separate 2019 neural decipherment paper evaluates Ugaritic with Hebrew and Linear B with Greek; these are separate language-pair experiments, not translation from Ugaritic into Linear B. Its input includes a non-parallel corpus in a known related language, so it does not require aligned bilingual sentences. A character-based sequence-to-sequence model learns correspondences between cognates. Monotonic alignment regularization encourages order-preserving character mappings, while minimum-cost flow constrains the global assignment of words across vocabularies. These assumptions and benchmark results do not prove that an unrelated, insufficiently documented script has been deciphered. Scholars must test proposed readings against independent contextual and linguistic evidence.