We are pleased to announce that Zaid Harchaoui, Associate Professor of Statistics, and his UW team received an Outstanding Paper Award at the NeurIPS conference in December 2021 for their paper on “MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers”. This is a joint project with UW students, Krishna Pillutla, John Thickstun and Rowan Zellers, UW-AI2 postdocs Swabha Swayamdipta and Sean Welleck, and UW Allen School Professor, Yejin Choi.

The paper introduces a statistical measure, called Mauve, allowing one to compare machine-generated text to human-written text. Mauve scales up to large language generation models such as UW’s Grover or OpenAI’s GPT-2 by computing information divergences in an embedding space.

In a companion paper with also UW Statistics student Lang Liu and UW Allen School Associate Professor Sewoong Oh, a statistical theory of divergence frontiers, a general class of operating characteristic (OC) curves whose summary statistics include Mauve, is developed with non-asymptotic guarantees.

You can read the Mauve paper here. The companion paper can be found here.

You can also see the full announcement here.