{"id":903717,"date":"2022-12-02T10:44:43","date_gmt":"2022-12-02T18:44:43","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/"},"modified":"2022-12-02T10:48:27","modified_gmt":"2022-12-02T18:48:27","slug":"invariant-language-modeling-2","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/invariant-language-modeling-2\/","title":{"rendered":"Invariant Language Modeling"},"content":{"rendered":"<p>Large pretrained language models are critical components of modern NLP pipelines.\u00a0 Yet, they suffer from spurious correlations, poor out-of-domain generalization, and biases.\u00a0 Inspired by recent progress in causal machine learning, in particular the invariant risk minimization (IRM) paradigm, we propose \\emph{invariant language modeling}, a framework for learning invariant representations that generalize better across multiple environments.\u00a0 In particular, we adapt a game-theoretic formulation of IRM (IRM-games) to language models, where the invariance emerges from a specific training schedule in which all the environments compete to optimize their own environment-specific loss by updating subsets of the model in a round-robin fashion.\u00a0 We focus on controlled experiments to precisely demonstrate the ability of our method to (i) remove structured noise, (ii) ignore specific spurious correlations without affecting global performance, and (iii) achieve better out-of-domain generalization.\u00a0 These benefits come with a negligible computational overhead compared to standard training, do not require changing the local loss, and can be applied to any language model.\u00a0 We believe this framework is promising to help mitigate spurious correlations and biases in language models.<\/p>\n<span id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\">Opens in a new tab<\/span>","protected":false},"excerpt":{"rendered":"<p>Large pretrained language models are critical components of modern NLP pipelines.\u00a0 Yet, they suffer from spurious correlations, poor out-of-domain generalization, and biases.\u00a0 Inspired by recent progress in causal machine learning, in particular the invariant risk minimization (IRM) paradigm, we propose \\emph{invariant language modeling}, a framework for learning invariant representations that generalize better across multiple environments.\u00a0 [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556,13545],"msr-publication-type":[193716],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-field-of-study":[],"msr-conference":[260143],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2022-12-7","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"Microsoft","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":1,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"file","viewUrl":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2022\/12\/Invariant-Language-Modeling.pdf","id":"903735","title":"invariant-language-modeling-2","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[{"id":903735,"url":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2022\/12\/Invariant-Language-Modeling.pdf"}],"msr-author-ordering":[{"type":"text","value":"Maxim Peyrard","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Sarvjeet Ghotra","user_id":40090,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Sarvjeet Ghotra"},{"type":"text","value":"Martin Josifoski","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Vidhan Agarwal","user_id":40135,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Vidhan Agarwal"},{"type":"user_nicename","value":"Barun Patra","user_id":39099,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Barun Patra"},{"type":"user_nicename","value":"Dean Carignan","user_id":40087,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Dean Carignan"},{"type":"user_nicename","value":"Emre Kiciman","user_id":31739,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Emre Kiciman"},{"type":"user_nicename","value":"Saurabh Tiwary","user_id":39603,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Saurabh Tiwary"},{"type":"text","value":"Robert West","user_id":0,"rest_url":false}],"msr_impact_theme":[],"msr_research_lab":[199565],"msr_event":[],"msr_group":[470706,722851],"msr_project":[],"publication":[],"video":[],"download":[],"msr_publication_type":"inproceedings","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/903717"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":1,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/903717\/revisions"}],"predecessor-version":[{"id":903753,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/903717\/revisions\/903753"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=903717"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=903717"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=903717"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=903717"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=903717"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=903717"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=903717"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=903717"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=903717"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=903717"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=903717"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=903717"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=903717"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=903717"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}