I think I have a genuine need for an # LLM . Can someone tell me if this is possible? @ openben
| Source: Mastodon | Original article
A user on the open‑source research platform OpenBenches has posted a concrete request: a corpus of roughly 40 000 cemetery inscriptions needs to be split by the gender of the honoree, but many entries list only initials or ambiguous names. The post, titled “I think I have a genuine need for an #LLM. Can someone tell me if this is possible?”, sparked a rapid response from the community, which began testing large language models for name‑gender inference on historical data.
The experiment hinges on prompting an LLM to reason through ambiguous cases—e.g., “To R Smith” versus the obvious “To Grandma Sylvia”—and to output a confidence score for each prediction. Early trials with OpenAI’s GPT‑4 and the locally hosted SGLang‑based model released last week showed that while the models can correctly classify clear‑cut names, they stumble on initials, gender‑neutral surnames, and culturally specific naming conventions. Researchers also flagged systematic bias: male‑associated names were identified with higher confidence than female‑associated ones, echoing concerns raised in recent analyses of LLM reasoning capabilities.
Why this matters is twofold. First, it demonstrates a practical, low‑cost avenue for digital‑humanities projects that lack dedicated linguistic expertise, potentially accelerating the cataloguing of heritage data across the Nordic region. Second, the bias patterns expose the risk of propagating historical gender imbalances when AI is used for archival work, underscoring the need for transparent evaluation frameworks.
The next steps will involve fine‑tuning a domain‑specific model on a curated list of Nordic names, integrating external gender‑lookup databases, and publishing a benchmark of accuracy versus traditional rule‑based methods. Observers will watch whether the community can produce an open‑source pipeline that balances performance with ethical safeguards, a development that could set a template for AI‑assisted scholarship beyond epigraphy.
Sources
Back to AIPULSEN