AI, Open Scholarship and Access to the Human Record

This panel reports on work exploring how a range of technologies has begun to transform our relationship to the human record. Translation can be viewed not as an impenetrable and authoritative record but as a gateway into the original source, whether or not we have been able to study the language. This new environment raises its own challenges and demands that readers cultivate a new set of critical skills so that they can properly assess what they encounter. Those critical skills are, however, an extension of the critical skills that we have cultivated in print culture, with the challenges posed by new digital environments reminding us that those skills are as important as ever. We begin by discussing how non-specialists have used Large Language Models (LLMs) to work directly with untranslated sources in both pre-modern and modern languages. We now have access to multiple LLMS, both commercial and open source, all of which can provide not only translation but also grammatical explanations and translations for each word in a source. In the base model, careful readers compare the output from multiple systems to identify significant divergences and to see how effectively they can assess which (if any) provides the most useful interpretation. We report on performance for Ancient Greek, Latin, Classical Persian, Classical Arabic, and Classical Japanese. AI can also contextualize a vast range of cultural artifacts now available to a global audience: films by Kurosawa at the Internet Archive, a 1940s Indian film about Alexander the Great in India, YouTube videos about Latvian history from a Latvian Pagan metal band. Students of classical works in various traditions around the world have developed, over thousands of years, methods by which to explain culturally alien sources to contemporary audiences. We report about research into these practices and about what we can learn from them as we train new AI-based systems to provide cultural context at scale to millions of sources in far more languages than any human can learn and than any university can currently teach. We report also on the current status of work to aggregate the growing range of services and open data. This includes linguistic annotations, translations and word and phrase level alignments between translation and source texts, named entity classification and linking to authority lists, metrical analyses, links between transcriptions and images of manuscripts and other written sources, and time based alignments between transcriptions and performances. Content includes materials created by human beings, purpose built natural language processing pipelines (like spaCy), and LLMs. We also report on progress towards integrating credits for individual annotations produced by multiple sources from different projects and countries and weaving these into portfolios for individuals. Finally, we report on how this work has shaped the development of, and now benefits from students within, a new Digital Humanities curriculum at Tufts University. Gregory Crane is Professor and Chair of Classical Studies; Sarah Abowitz is a PhD student in Computer Science; Alison Babeu is Digital Librarian at Perseus; Peter Nadel is NLP and DH Specialist for Tufts Computing; Charles Pletcher is a developer and part time lecturer at Tufts; James Tauber is a preceport and scholar developer at Signum University; Cliff Wulfman works at the Princeton Library.