Transcript ppt
The Liberal Media and Right-Wing Conspiracies: Using Cocitation Information to Estimate Political Orientation in Web Documents Miles Efron School of Information, University of Texas ACM CIKM 2004 (19% acceptance) Objective • Identify the political orientation of web documents in the US • Left Versus Right Approach • Pure link-based approach • Web pages that are cited frequently with left/right web pages have left/right orientation • Use PMI-IR model on links PMI for Links • Pointwise Mutual Information: • p(di) = probability that a document in the community contains a link to di • p(di,dj) = probability that a document in the community contains links to both di and dj Political Orientation Score • PO(di) = PMI(di, R) - PMI(di, L) • R : set of exemplar right-leaning web pages • L : set of exemplar left-leaning web pages • PO(di) > 0 right orientation • PO(di) = 0 neutral • PO(di) < 0 left orientation Political Score Estimation • After some simple math, we get • “link: di” is the number of links to di given by Altavista • “link: R” is the query “link:R1 OR link:R2 OR … OR link:R|R|” to Altavista Exemplar Partisan Documents • 19 web pages per side • Use Open Directory Project (www.dmoz.org) • Left Side: Use web pages in Politics:Liberalism:Social Liberalism category with at least 1000 incoming links • Right Side: Use Politics:Conservatism category Experiment 1: Web Data • Use Yahoo! and Open Directory Project to build a corpus of 695 testing documents • Baseline: SVM and Naïve Bayes using lexical features – Training corpus of 2412 documents is constructed from the first 2 levels of exemplar web pages Experiment 1 Result Experiment 2: Blogs • 119 left-wing blogs • 43 right-wing blogs Experiment 2 Result Method %Accuracy Cocitation 98.8 (2 mistakes) SVM 60.5 (64 mistakes) Experiment 3: Non-Partisan Documents • Select 72 non-partisan web pages from Open Directory Project • Test whether the orientation score is close to 0 Experiment 3 Result Future Work • Combine lexical information with links