{"created":"2026-02-19T10:21:22.706668+00:00","metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:02007757","sets":["1164:4179:1771495534039:1771495585049"]},"path":["1771495585049"],"owner":"80578","recid":"2007757","title":["社会的属性ごとに異なる内部表現を考慮した活性化ステアリングによるLLMの社会的バイアス緩和"],"pubdate":{"attribute_name":"PubDate","attribute_value":"2026-02-28"},"_buckets":{"deposit":"43cfcacb-7de6-499e-b567-561644ee0d18"},"_deposit":{"id":"2007757","pid":{"type":"depid","value":"2007757","revision_id":0},"owners":[80578],"status":"published","created_by":80578},"item_title":"社会的属性ごとに異なる内部表現を考慮した活性化ステアリングによるLLMの社会的バイアス緩和","author_link":[],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"社会的属性ごとに異なる内部表現を考慮した活性化ステアリングによるLLMの社会的バイアス緩和","subitem_title_language":"ja"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"LLMの信頼性","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2026-02-28","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"同志社大学文化情報学部"},{"subitem_text_value":"同志社大学大学院文化情報学研究科"},{"subitem_text_value":"同志社大学文化情報学部"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/2007757/files/IPSJ-NL26267031.pdf","label":"IPSJ-NL26267031.pdf"},"date":[{"dateType":"Available","dateValue":"2028-02-28"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-NL26267031.pdf","filesize":[{"value":"1.1 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"23"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"37172b99-df9a-4c0e-bb0a-516b80804365","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2026 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"前川,知歩"}]},{"creatorNames":[{"creatorName":"木村,優介"}]},{"creatorNames":[{"creatorName":"波多野,賢治"}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10115061","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8779","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"大規模言語モデルは学習データに含まれる偏りを反映し，人種や性別などの社会的属性に関して不当な差別や偏見といった社会的バイアスを含む挙動を示すことが知られている．近年，推論時にモデルの内部表現へ介入することで挙動を制御する活性化ステアリングに基づくバイアス緩和手法が提案されているが，多くの既存手法は社会的バイアスをモデル全体で共通の単一現象として扱っている．しかし，社会的バイアスは属性ごとに異なる文脈や内部表現に基づいて発現すると考えられ，単一の介入方向では不要な副作用を引き起こす可能性がある．そこで本研究では，社会的属性ごとにバイアスに関与する内部表現の部分空間が異なるという仮定のもと，内部表現を表現分解手法により高解像度な特徴に分解し，各属性に特化したステアリングベクトルを学習する手法を提案する．","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"9","bibliographic_titles":[{"bibliographic_title":"研究報告自然言語処理（NL）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2026-02-28","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"31","bibliographicVolumeNumber":"2026-NL-267"}]},"relation_version_is_last":true,"weko_creator_id":"80578"},"id":2007757,"updated":"2026-02-19T10:21:27.018563+00:00","links":{}}