Wikimedia Research/Showcase/Archive/2018/08

August 2018

13 August 2018 Video: YouTube

Quicksilver

Training an ML system to generate draft Wikipedia articles and Wikidata entries simultaneously

slides

By John Bohannon and Vedant Dharnidharka, Primer

The automatic generation and updating of Wikipedia articles is usually approached as a multi-document summarization task: Given a set of source documents containing information about an entity, summarize the entity. Purely sequence-to-sequence neural models can pull that off, but getting enough data to train them is a challenge. Wikipedia articles and their reference documents can be used for training, as was recently done by a team at Google AI. But how do you find new source documents for new entities? And besides having humans read all of the source documents, how do you fact-check the output? What is needed is a self-updating knowledge base that learns jointly with a summarization model, keeping track of data provenance. Lucky for us, the world’s most comprehensive public encyclopedia is tightly coupled with Wikidata, the world’s most comprehensive public knowledge base. We have built a system called Quicksilver uses them both.