Leveraging Wikipedia article evolution for promotional tone detection

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5601 - 5613 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Leveraging Wikipedia article evolution for promotional tone detection Christine de Kock Department of Computer Science and Technology University of Cambridge∗ cd700@cam.ac.uk Andreas Vlachos Department of Computer Science and Technology University of Cambridge av308@cam.ac.uk Abstract Detecting biased language is useful for a variety of applications, such as identifying hyperpartisan news sources or flagging one- sided rhetoric. In this work we introduce WikiEvolve, a dataset for document-level pro- motional tone detection in English. Unlike previously proposed datasets, it contains seven versions of the same article from Wikipedia, from different points in its revision history; one with promotional tone, and six without it. We adapt the gradient reversal layer frame- work to encode two article versions simulta- neously, and thus leverage the training signal present in the multiple versions. In our ex- periments, our proposed adaptation of gradient reversal improves the accuracy of four differ- ent architectures on both in-domain and out- of-domain evaluation. 1 Introduction Maintaining a neutral point of view is a desidera- tum in many communication channels, e.g. news articles, scientific writing, and encyclopaedias. Bi- ased writing detection can help reduce the distri- bution of content which contains unfair represen- tations of a topic. For this reason, datasets and methods have been developed to automate it. A number of studies have approached biased writing detection in the context of news media (e.g. Fan et al. (2019); Chen et al. (2020); Färber et al. (2020)), primarily considering political stance and partisanship. However, biased writing also arises in other settings. In Wikipedia, the online ency- clopaedia, it manifests itself in the form of promo- tional tone which violates the cornerstone “neutral point of view” policy of the platform. The lat- ter allows users to flag articles with such policy violations by adding tags to the article mark-up, which are retained in its edit history. Leveraging ∗This work was initiated during an internship at the Wiki- media Foundation. Alain Connes Alain Connes (born April 1, 1947) is a French mathematician. Although his work in physics was not very convincing he tried to connect the planckian scales with what he called a "2-brane" Universe, model which was largely rejected by string theorists so far. Alain Connes (born April 1, 1947) is a French mathematician. [He] is one of the leading specialists on operator algebras. He made substantial contributions in operator K-theory and index theory, which culminated in the celebrated Baum-Connes conjecture. Alain Connes (born 1 April 1947) is a French mathematician. [He] is one of the leading specialists on operator algebras. He made contributions in operator K-theory and index theory, which culminated in the Baum-Connes conjecture. Version 2 2007-10-25 pos Version 3 2010-03-12 neg_post Version 1 2006-02-12 neg_pre Figure 1: A sample from the dataset. The middle ver- sion was tagged as having a promotional tone problem. The first and third versions were sampled respectively before …

Content truncated to 500 words