Apollo Chatbot Takes On Ancient Greek Papyrus Puzzles

Hundreds of thousands of Ancient Greek papyrus fragments sit in academic libraries worldwide, many too damaged for their meaning to survive. Scholars can restore some by methodically filling in missing words, and researchers have now turned to artificial intelligence to speed up that laborious work.

On Wednesday, the Austrian Academy of Science will release what it calls the world's first advanced large language model for Ancient Greek, developed with French AI lab Mistral and technology services firm Sail Reply. The model, Apollo, is trained on roughly 600 million historical Greek words from manuscripts, papyri, and inscriptions.

Academics will get free access through a chatbot interface. The goal is to help them quickly find fragments relevant to their sub-disciplines and spot promising research avenues. Where documents are tattered, Apollo fills in blanks with the most statistically likely words or passages, potentially revealing hidden details about historical events and practices.

Dimitris Vlitas, a partner at Sail Reply, tells WIRED that unlocking knowledge this way was unthinkable a year ago. Restoring a tattered papyrus has required a skilled academic to identify word divisions-Ancient Greek writing has no gaps-date the document, weigh socio-political contexts, and consult references for suitable words. Stephen Colvin, a professor of classics and historical linguistics at University College London, says very few people are that good at Greek history.

That specialized knowledge is baked into Apollo. Anna Dolganov, a historian and papyrologist at the Austrian Academy of Science, says that when it sees Homer, it supplements Homeric Greek, and when it sees an inscription in Doric dialect, it uses Doric dialect.

Academics bogged down in reconstruction expect Apollo to let them focus on what documents mean rather than what they say. Armand D'Angour, a professor of classical languages and literature at the University of Oxford, which holds the world's largest ancient papyrus collection, calls it very exciting and says a machine offering three possible words for a gap would speed things up considerably.

Apollo is unlikely to change broad understanding of the ancient world; many papyri remain unrestored because they are mundane-personal letters, marital contracts, civil service papers. Colvin says a layperson might think new plays by Sophocles will suddenly appear, but that will not happen. Still, the model could uncover details about life in antiquity and substantiate existing scholarly assumptions. D'Angour says every produced item adds a tiny element of knowledge about the ancient world.

If Apollo succeeds, Vlitas says the same technique could apply to other ancient languages, such as Latin or Egyptian, or any academic discipline that benefits from distilling and indexing a large corpus. AI has had notable successes: OpenAI recently said its models solved a 200-year-old math problem, while Google DeepMind released a vast dataset mapping how genetic mutations affect molecular biology, compiled using AI.

One concern is that relying on a probability-based model to fill gaps could pollute the historical record with errors. To head that off, Apollo proposes a selection of word options for a scholar to choose between. Dolganov says the crucial point is that human competence must remain, and that total reliance on AI transcriptions and interpretations of historical material is when problems start.

Source: Wired