New AI Can Now Identify Obscure People, Places, and Things

When AI reads a sentence or looks at a photo, it has to figure out who or what is being mentioned — and then match it to the right entry in a giant digital encyclopedia. Researchers call this "entity linking." The problem is that today's systems are excellent at famous names like Taylor Swift but often fail on obscure ones, like a small-town mayor or a regional dish. Graham Neubig's team at Carnegie Mellon wanted to know exactly how bad that blind spot is.

First, they redefined what "rare" means. Instead of counting how many people have looked something up, they measured how well-documented and well-connected an entry is in a knowledge graph (a web of facts linked together). That surfaced many obscure subjects the old popularity measure completely missed. On these harder cases, even the best existing AI systems lost between 15% and 40% of their accuracy — a striking drop.

Their fix is refreshingly simple and requires no retraining. They let a vision-and-language AI — one that can see images and read text — search Wikipedia step by step and reason about what it finds, like a person Googling while thinking out loud. Controlled tests showed the two halves need each other: thinking alone barely helped, looking things up alone helped rare cases but sometimes hurt common ones, and doing both together worked best.

Across five languages — Hindi, Indonesian, Japanese, Tamil, and Vietnamese — their system beat the previous best by 6.9% overall and up to 23.3% on rare subjects. The team also released a special test set so others can measure progress on obscure cases. The catch: this is research presented at a conference, not a product you can use today, and step-by-step searching is slower than answering instantly from memory.