{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:WQNXSGA64KYKW4TVYCYDPM2RMS","short_pith_number":"pith:WQNXSGA6","schema_version":"1.0","canonical_sha256":"b41b79181ee2b0ab7275c0b037b35164bc48cce4569790dc3ecdfe09469280ee","source":{"kind":"arxiv","id":"2205.12586","version":2},"attestation_state":"computed","paper":{"title":"Perturbation Augmentation for Fairer NLP","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Adina Williams, Candace Ross, Douwe Kiela, Eric Smith, Jude Fernandes, Rebecca Qian","submitted_at":"2022-05-25T09:00:29Z","abstract_excerpt":"Unwanted and often harmful social biases are becoming ever more salient in NLP research, affecting both models and datasets. In this work, we ask whether training on demographically perturbed data leads to fairer language models. We collect a large dataset of human annotated text perturbations and train a neural perturbation model, which we show outperforms heuristic alternatives. We find that (i) language models (LMs) pre-trained on demographically perturbed corpora are typically more fair, and (ii) LMs finetuned on perturbed GLUE datasets exhibit less demographic bias on downstream tasks, an"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2205.12586","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2022-05-25T09:00:29Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"08e69b9d7958035470377d08b14b23dbff5dc96dfa0c9265cbccac708f522e61","abstract_canon_sha256":"9175f7e8005f7216fa718b8a062e135cde5dfa940f364d8eec0f52ffc69b7f42"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:06:12.183746Z","signature_b64":"GeIIxcTwOQOCd4Jr/CNDLvleTbb2oyZp3v6Y1fhJzn3gJ57cka96hwc0qwm9FMC31NdcQjPjidO7pmeHhrEUCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b41b79181ee2b0ab7275c0b037b35164bc48cce4569790dc3ecdfe09469280ee","last_reissued_at":"2026-07-05T05:06:12.183234Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:06:12.183234Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Perturbation Augmentation for Fairer NLP","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Adina Williams, Candace Ross, Douwe Kiela, Eric Smith, Jude Fernandes, Rebecca Qian","submitted_at":"2022-05-25T09:00:29Z","abstract_excerpt":"Unwanted and often harmful social biases are becoming ever more salient in NLP research, affecting both models and datasets. In this work, we ask whether training on demographically perturbed data leads to fairer language models. We collect a large dataset of human annotated text perturbations and train a neural perturbation model, which we show outperforms heuristic alternatives. We find that (i) language models (LMs) pre-trained on demographically perturbed corpora are typically more fair, and (ii) LMs finetuned on perturbed GLUE datasets exhibit less demographic bias on downstream tasks, an"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2205.12586","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2205.12586/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2205.12586","created_at":"2026-07-05T05:06:12.183292+00:00"},{"alias_kind":"arxiv_version","alias_value":"2205.12586v2","created_at":"2026-07-05T05:06:12.183292+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2205.12586","created_at":"2026-07-05T05:06:12.183292+00:00"},{"alias_kind":"pith_short_12","alias_value":"WQNXSGA64KYK","created_at":"2026-07-05T05:06:12.183292+00:00"},{"alias_kind":"pith_short_16","alias_value":"WQNXSGA64KYKW4TV","created_at":"2026-07-05T05:06:12.183292+00:00"},{"alias_kind":"pith_short_8","alias_value":"WQNXSGA6","created_at":"2026-07-05T05:06:12.183292+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2508.07111","citing_title":"Investigating Intersectional Bias in Large Language Models using Confidence Disparities in Coreference Resolution","ref_index":54,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS","json":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS.json","graph_json":"https://pith.science/api/pith-number/WQNXSGA64KYKW4TVYCYDPM2RMS/graph.json","events_json":"https://pith.science/api/pith-number/WQNXSGA64KYKW4TVYCYDPM2RMS/events.json","paper":"https://pith.science/paper/WQNXSGA6"},"agent_actions":{"view_html":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS","download_json":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS.json","view_paper":"https://pith.science/paper/WQNXSGA6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2205.12586&json=true","fetch_graph":"https://pith.science/api/pith-number/WQNXSGA64KYKW4TVYCYDPM2RMS/graph.json","fetch_events":"https://pith.science/api/pith-number/WQNXSGA64KYKW4TVYCYDPM2RMS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS/action/storage_attestation","attest_author":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS/action/author_attestation","sign_citation":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS/action/citation_signature","submit_replication":"https://pith.science/pith/WQNXSGA64KYKW4TVYCYDPM2RMS/action/replication_record"}},"created_at":"2026-07-05T05:06:12.183292+00:00","updated_at":"2026-07-05T05:06:12.183292+00:00"}