Wikimedia Research/Showcase/Archive/2016/08

August 2016

August 17, 2016 Video: YouTube

Computational Fact Checking from Knowledge Networks

By Giovanni Luca Ciampaglia

Traditional fact checking by expert journalists cannot keep up with the enormous volume of information that is now generated online. Fact checking is often a tedious and repetitive task and even simple automation opportunities may result in significant improvements to human fact checkers. In this talk I will describe how we are trying to approximate the complexities of human fact checking by exploring a knowledge graph under a properly defined proximity measure. Framed as a network traversal problem, this approach is feasible with efficient computational techniques. We evaluate this approach by examining tens of thousands of claims related to history, entertainment, geography, and biographical information using the public knowledge graph extracted from Wikipedia by the DBPedia project, showing that the method does indeed assign higher confidence to true statements than to false ones. One advantage of this approach is that, together with a numerical evaluation, it also provides a sequence of statements that can be easily inspected by a human fact checker.

Deploying and maintaining AI in a socio-technical system. Lessons learned

slides

By Aaron Halfaker

We should exercise great caution when deploying AI into our social spaces. The algorithms that make counter-vandalism in Wikipedia orders of magnitude more efficient also have the potential to perpetuate biases and silence whole classes of contributors. This presentation will describe the system efficiency characteristics that make AI so attractive for supporting quality control activities in Wikipedia. Then, Aaron will tell two stories of how the algorithms brought new, problematic biases to quality control processes in Wikipedia and how the Revision Scoring team learned about and addressed these issues in ORES, a production-level AI service for Wikimedia Wikis. He'll also make an overdue call to action toward leveraging human-review of AIs biases in the practice of AI development.