Inferring Advertiser Sentiment in Online Articles using Wikipedia Footnotes

Abstract

Online advertising platforms in partnerships with media companies typically have access to an online user's history of viewed articles. If a concerned brand (advertiser) plans to run advertisement campaigns on users exposed to negative articles, it is essential to first identify articles with negative sentiment about the brand. For an advertising platform, scalable identification of such articles with little human-annotation effort is necessary for launching campaigns soon after an advertiser signs up. In this context, generic sentiment analysis tools suffer from the lack of contextual world knowledge associated with the advertiser. Human annotation of articles for supervised approaches is laborious and painstaking. To address these problems, we propose the use of publicly available Wikipedia footnote references for an advertiser, and propagate their sentiment to several articles related to the advertiser. In particular, our proposed approach has three components: (i) automatically find Wikipedia references which have negative sentiment about an advertiser, (ii) learn distributed representations (doc2vec) of article texts referred in footnotes and other unlabeled articles, and (iii) inferring sentiment in unlabeled articles using label propagation (from references) in the doc2vec space. Our experiments spanning three real brands, and data from a major advertising platform (Yahoo Gemini) show significant lifts in sentiment inference compared to existing baselines. In addition, we share valuable insights on how article sentiment influences the online activities of a user with respect to a brand. CCS CONCEPTS• Information systems → Online advertising.

A graph-based semi-supervised approach to classification learning in digital geographies

“…Doc2Vec + Label Propagation: We use Doc2Vec (Le and Mikolov, 2014) to extract text representation and a traditional semi-supervised machine learning approach Label Propagation (LP) (Zhu and Ghahramani, 2002) to classify tweets. Such a method has been widely adopted on online content analysis (e.g., sentiment analysis (Mishra et al, 2019;Wadawadagi and Pagi, 2020)). As a semi-supervised machine learning approach, LP is used to assess the performance of the GCN framework, which is also a semi-supervised learning framework.…”