Dixon Jones talks with Jason Barnard about the bias in the Knowledge Graph.
Jason and Dixon Jones are sitting in two comfy armchairs, looking at the sea in Brighton. We start with a chat about machine learning in Google’s search algorithm, PageRank and then onto the Knowledge Graph. There are less entities in the world than webpages. So Google’s job is easier. But the Knowledge Graph is biased – the seed set for google’s understanding is a bunch of librarians (aka Wikipedia editors) who have little in depth knowledge on the topics they edit, especially in anything that is not within their culture. We happily grab examples from the surrounding environment. Piers become a central point, and piers in Ethiopia in particular. We move onto fan sites, that are not necessarily accurate, and perhaps people believe that William Shatner is a space man. Errors such as that at the start of a seed set will mean learning is biased and perhaps inaccurate… and can quickly spiral out of control. They are building on what Dixon calls ‘areas of light’, but that is biased too. One problem is that genuinely good new ideas are going to have trouble surfacing because of the bias against ideas that are not popularly held belief. We move onto loops of truth and self-fulfilling prophecies. Fake news gets a look in (of course). As does bad fact checking. Then we finish off with InLinks – Dixon’s super new SaaS for automatically building internal knowledge graphs and writing scheme.org structured data on the fly. I ask a trick question, and Dixon deals with it rather well. And we end by coining the phrase ‘The Wikipedia model’.