{"id":969162,"date":"2023-09-19T17:41:02","date_gmt":"2023-09-20T00:41:02","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&p=969162"},"modified":"2024-05-07T16:43:47","modified_gmt":"2024-05-07T23:43:47","slug":"large-language-models-can-accurately-predict-searcher-preferences","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/large-language-models-can-accurately-predict-searcher-preferences\/","title":{"rendered":"Large Language Models Can Accurately Predict Searcher Preferences"},"content":{"rendered":"

Much of the evaluation and tuning of a search system relies on rele<\/span>vance labels\u2014annotations that say whether a document is useful for <\/span>a given search and searcher. Ideally these come from real searchers, <\/span>but it is hard to collect this data at scale, so typical experiments rely <\/span>on third-party labellers who may or may not produce accurate an<\/span>notations. Label quality is managed with ongoing auditing, training, <\/span>and monitoring.<\/span><\/p>\n

We discuss an alternative approach. We take careful feedback <\/span>from real searchers and use this to select a large language model <\/span>(LLM), and prompt, that agrees with this feedback; the LLM can <\/span>then produce labels at scale. Our experiments show LLMs are as <\/span>accurate as human\u00a0 labellers and as useful for finding the best sys<\/span>tems and hardest queries. LLM performance varies with prompt <\/span>features, but also varies unpredictably with simple paraphrases. This <\/span>unpredictability reinforces the need for high-quality \u201cgold\u201d labels.<\/span><\/p>\nOpens in a new tab<\/span>","protected":false},"excerpt":{"rendered":"

Much of the evaluation and tuning of a search system relies on relevance labels\u2014annotations that say whether a document is useful for a given search and searcher. Ideally these come from real searchers, but it is hard to collect this data at scale, so typical experiments rely on third-party labellers who may or may not […]<\/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":[13556,13555],"msr-publication-type":[193716],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-field-of-study":[246694,248503,267222,246685],"msr-conference":[260209],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"msr_publishername":"ACM","msr_edition":"","msr_affiliation":"","msr_published_date":"2024-7-14","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":"","msr_how_published":"","msr_notes":"An earlier version of this paper appeared as arXiv preprint arXiv:2309.10621v1 [cs.IR].","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":"doi","viewUrl":"false","id":"false","title":"10.1145\/3626772.3657707","label_id":"243106","label":0},{"type":"file","viewUrl":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prodnew\/2023\/09\/LLMs_for_relevance_labelling__SIGIR_24_-2.pdf","id":"1031610","title":"llms_for_relevance_labelling__sigir_24_-2","label_id":"243132","label":0}],"msr_related_uploader":"","msr_attachments":[{"id":1031610,"url":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prodnew\/2024\/05\/LLMs_for_relevance_labelling__SIGIR_24_-2.pdf"},{"id":1030353,"url":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prodnew\/2024\/05\/LLMs_for_relevance_labelling__SIGIR_24_.pdf"}],"msr-author-ordering":[{"type":"user_nicename","value":"Paul Thomas","user_id":36042,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Paul Thomas"},{"type":"user_nicename","value":"Seth Spielman","user_id":43314,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Seth Spielman"},{"type":"user_nicename","value":"Nick Craswell","user_id":33088,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Nick Craswell"},{"type":"user_nicename","value":"Bhaskar Mitra","user_id":31257,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Bhaskar Mitra"}],"msr_impact_theme":[],"msr_research_lab":[199561,437514],"msr_event":[],"msr_group":[913161,267093],"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\/969162"}],"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":5,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/969162\/revisions"}],"predecessor-version":[{"id":1032150,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/969162\/revisions\/1032150"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=969162"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=969162"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=969162"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=969162"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=969162"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=969162"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=969162"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=969162"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=969162"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=969162"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=969162"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=969162"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=969162"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=969162"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}