{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/adversarial-removal-of-demographic-attributes","title":"Adversarial Removal of Demographic Attributes from Text Data","arxiv_id":"1808.06640","date":"2018-08-20","proceeding":"EMNLP 2018 10","authors":["Yanai Elazar","Yoav Goldberg"],"abstract":"Recent advances in Representation Learning and Adversarial Training seem to\nsucceed in removing unwanted features from the learned representation. We show\nthat demographic information of authors is encoded in -- and can be recovered\nfrom -- the intermediate representations learned by text-based neural\nclassifiers. The implication is that decisions of classifiers trained on\ntextual data are not agnostic to -- and likely condition on -- demographic\nattributes. When attempting to remove such demographic information using\nadversarial training, we find that while the adversarial component achieves\nchance-level development-set accuracy during training, a post-hoc classifier,\ntrained on the encoded sentences from the first part, still manages to reach\nsubstantially higher classification accuracies on the same data. This behavior\nis consistent across several tasks, demographic properties and datasets. We\nexplore several techniques to improve the effectiveness of the adversarial\ncomponent. Our main conclusion is a cautionary one: do not rely on the\nadversarial training to achieve invariant representation to sensitive features.","url_abs":"http://arxiv.org/abs/1808.06640v2","url_pdf":"http://arxiv.org/pdf/1808.06640v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"adversarial-removal-of-demographic-attributes","repo_url":"https://github.com/yanaiela/demog-text-removal","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"none","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"representation-learning","task_name":"Representation Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1808.06640","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}