Austria's Ancient Greek model proposes missing words for scholars to choose between

Ancient Greek was written without spaces between words. So before anyone can guess at a missing word in a torn papyrus, the word divisions have to be found, the document dated, its socio-political context weighed, and reference works consulted ^[6]. Four separate acts of specialist judgement come before the guess itself ^[7]. Stephen Colvin is a professor of classics and historical linguistics at University College London. He said of the people who can do all of it: "There are very few people in the world who are that good at Greek history" ^[8].

Anna Dolganov, a historian and papyrologist at the Austrian Academy of Science, said the model matches the register of whatever it is shown: "When it sees Homer, it supplements Homeric Greek. When it sees an inscription in Doric dialect, it uses Doric dialect" ^[9]. The WIRED report does not include an accuracy rate for those suggestions or a comparison against a general-purpose model ^[18]. The case for a narrow model here therefore rests on the corpus that went in and on the shape of what comes back out.

What comes back out is deliberately unfinished. Apollo offers a selection of candidate words and the scholar selects one, a choice the project made to keep probabilistic guesses from polluting the historical record with errors ^[10]. "The crucial point is that human competence needs to remain," Dolganov said ^[11]. Armand D'Angour, a professor of classical languages and literature at the University of Oxford, put the same design from the user's side. "If I had a machine telling me, 'Here are the three possible words that could fit into that gap,' it would speed up matters considerably," he said ^[13].

The user is a specialist working through ordinary paperwork. Academic libraries hold hundreds of thousands of Ancient Greek papyrus fragments, many of them so damaged that their meaning is probably lost ^[5]. The ones still waiting are personal letters, marital contracts and civil service papers ^[15]. Colvin said: "If you were a layperson, you might think suddenly we'll get a few new plays by Sophocles, but that's not going to happen" ^[14]. D'Angour said: "Every time something is produced, it adds a tiny element of knowledge about the ancient world" ^[20].

Two things decide whether this shape transfers to a domain model in your own field. Do you hold a corpus nobody else has indexed, and does a named person accept or reject each suggestion before anyone else sees it? Dimitris Vlitas, a partner at Sail Reply, said the same technique could be applied to Latin or Egyptian, or any academic discipline that would benefit from the distillation and indexing of a large corpus ^[17]. Those disciplines come with editors and peer reviewers already in the loop. A team whose model output goes to a customer with nobody in between cannot copy the shortlist design. There is no one there to choose.