III: Small: Mirador: Explainable Computational Models for Recognizing and Understanding Controversial Topics Encountered Online

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

This project aims to develop algorithms and tools that allow a person to recognize that a web page or other document discusses one or more topics that are controversial -- that is, about which there is strong disagreement within some sizeable group of people. The project will develop algorithms and tools that explain the controversy surrounding the topic, identifying the populations that disagree, the stances that they take, and how those stances conflict with each other. The advances in these algorithms will broaden the research community's understanding of how discussions and disagreements on topics can be modeled computationally and how that resulting information can be conveyed to a general user. The project will assist people in critical evaluation of on-line material and help them understand why a page is educative or why it is not. The aim of this project is to provide users with tools that illuminate the broader context of the topic or topics of a single page or document that someone finds. Previous work has shown that it is possible to recognize with reasonable accuracy that a document is part of a controversial topic, but that work is fragmented across different genres, demands more robust modeling and more thorough evaluation, and lacks explanatory power that can help a reader understand why and how a text is contentious. In this project, the researchers explore fundamental questions about how controversy can be modeled computationally so that it can be recognized "in the wild". The project also explores model variations that allow an algorithm to extract an explanation of the nature of the controversy. The project applies and extends text analysis and comparison techniques. It leverages powerful statistical language modeling methods as well as recent neural network (deep learning) approaches to represent text, its controversial nature, its stances, and their relationships, all extracted from Web pages and other documents. The modeling will be initially used offline to identify collections of topics known to be controversial and then adapt that collection by monitoring slowly-changing news sources and blog postings as well as ephemeral microblog sources of data to capture rapid changes in controversy. The researchers will make the resulting techniques available by providing an open-source example server. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.