Deprecated: The each() function is deprecated. This message will be suppressed on further calls in /home/zhenxiangba/zhenxiangba.com/public_html/phproxy-improved-master/index.php on line 456 Paper page - NeKo: Toward Post Recognition Generative Correction Large Language
Models with Task-Oriented Experts
Please give a thumbs up to this comment if you found it helpful!
\n
If you want recommendations for any Paper on Hugging Face checkout this Space
\n
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend
\n","updatedAt":"2024-11-13T01:33:07.305Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7479429841041565},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2411.05945","authors":[{"_id":"67335d4cd40c698f6d001c1f","name":"Yen-Ting Lin","hidden":false},{"_id":"67335d4cd40c698f6d001c20","user":{"_id":"629e1b71bb6419817ed7566c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/629e1b71bb6419817ed7566c/0ZCt-11eQtRDCOk9AozOp.jpeg","isPro":false,"fullname":"Huck Yang","user":"huckiyang","type":"user"},"name":"Chao-Han Huck Yang","status":"extracted_confirmed","statusLastChangedAt":"2025-10-20T05:09:55.802Z","hidden":false},{"_id":"67335d4cd40c698f6d001c21","name":"Zhehuai Chen","hidden":false},{"_id":"67335d4cd40c698f6d001c22","name":"Piotr Zelasko","hidden":false},{"_id":"67335d4cd40c698f6d001c23","user":{"_id":"64b708fa53d91a364aa32bbf","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b708fa53d91a364aa32bbf/f8gnDFShkI3VBhadko8qr.jpeg","isPro":false,"fullname":"Xuesong Yang","user":"magicyoung8","type":"user"},"name":"Xuesong Yang","status":"claimed_verified","statusLastChangedAt":"2025-06-13T07:43:42.121Z","hidden":false},{"_id":"67335d4cd40c698f6d001c24","name":"Zih-Ching Chen","hidden":false},{"_id":"67335d4cd40c698f6d001c25","name":"Krishna C Puvvada","hidden":false},{"_id":"67335d4cd40c698f6d001c26","name":"Szu-Wei Fu","hidden":false},{"_id":"67335d4cd40c698f6d001c27","name":"Ke Hu","hidden":false},{"_id":"67335d4cd40c698f6d001c28","name":"Jun Wei Chiu","hidden":false},{"_id":"67335d4cd40c698f6d001c29","name":"Jagadeesh Balam","hidden":false},{"_id":"67335d4cd40c698f6d001c2a","name":"Boris Ginsburg","hidden":false},{"_id":"67335d4cd40c698f6d001c2b","name":"Yu-Chiang Frank Wang","hidden":false}],"publishedAt":"2024-11-08T20:11:24.000Z","submittedOnDailyAt":"2024-11-12T11:24:23.351Z","title":"NeKo: Toward Post Recognition Generative Correction Large Language\n Models with Task-Oriented Experts","submittedOnDailyBy":{"_id":"629e1b71bb6419817ed7566c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/629e1b71bb6419817ed7566c/0ZCt-11eQtRDCOk9AozOp.jpeg","isPro":false,"fullname":"Huck Yang","user":"huckiyang","type":"user"},"summary":"Construction of a general-purpose post-recognition error corrector poses a\ncrucial question: how can we most effectively train a model on a large mixture\nof domain datasets? The answer would lie in learning dataset-specific features\nand digesting their knowledge in a single model. Previous methods achieve this\nby having separate correction language models, resulting in a significant\nincrease in parameters. In this work, we present Mixture-of-Experts as a\nsolution, highlighting that MoEs are much more than a scalability tool. We\npropose a Multi-Task Correction MoE, where we train the experts to become an\n``expert'' of speech-to-text, language-to-text and vision-to-text datasets by\nlearning to route each dataset's tokens to its mapped expert. Experiments on\nthe Open ASR Leaderboard show that we explore a new state-of-the-art\nperformance by achieving an average relative 5.0% WER reduction and\nsubstantial improvements in BLEU scores for speech and translation tasks. On\nzero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with 15.5% to\n27.6% relative WER reduction in the Hyporadise benchmark. NeKo performs\ncompetitively on grammar and post-OCR correction as a multi-task model.","upvotes":4,"discussionId":"67335d4dd40c698f6d001c63","ai_summary":"A Multi-Task Correction MoE enhances error correction across different domains, achieving state-of-the-art performance with improved WER and BLEU scores.","ai_keywords":["Mixture-of-Experts","Multi-Task Correction MoE","speech-to-text","language-to-text","vision-to-text","Open ASR Leaderboard","WER","BLEU","zero-shot evaluation","NeKo","GPT-3.5","Claude-Opus","Hyporadise","post-OCR correction"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"629e1b71bb6419817ed7566c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/629e1b71bb6419817ed7566c/0ZCt-11eQtRDCOk9AozOp.jpeg","isPro":false,"fullname":"Huck Yang","user":"huckiyang","type":"user"},{"_id":"643b19f8a856622f978df30f","avatarUrl":"/avatars/c82779fdf94f80cdb5020504f83c818b.svg","isPro":false,"fullname":"Yatharth Sharma","user":"YaTharThShaRma999","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"648eb1eb59c4e5c87dc116e0","avatarUrl":"/avatars/c636cea39c2c0937f01398c94ead5dad.svg","isPro":false,"fullname":"fdsqefsgergd","user":"T-representer","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">
A Multi-Task Correction MoE enhances error correction across different domains, achieving state-of-the-art performance with improved WER and BLEU scores.
AI-generated summary
Construction of a general-purpose post-recognition error corrector poses a
crucial question: how can we most effectively train a model on a large mixture
of domain datasets? The answer would lie in learning dataset-specific features
and digesting their knowledge in a single model. Previous methods achieve this
by having separate correction language models, resulting in a significant
increase in parameters. In this work, we present Mixture-of-Experts as a
solution, highlighting that MoEs are much more than a scalability tool. We
propose a Multi-Task Correction MoE, where we train the experts to become an
``expert'' of speech-to-text, language-to-text and vision-to-text datasets by
learning to route each dataset's tokens to its mapped expert. Experiments on
the Open ASR Leaderboard show that we explore a new state-of-the-art
performance by achieving an average relative 5.0% WER reduction and
substantial improvements in BLEU scores for speech and translation tasks. On
zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with 15.5% to
27.6% relative WER reduction in the Hyporadise benchmark. NeKo performs
competitively on grammar and post-OCR correction as a multi-task model.
NeKo is a new series of post recognition correction LLMs for post ASR, MT, OCR correction with average relative 5.0% WER reduction in the open ASR leaderboard. On zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with 15.5% to 27.6% relative WER reduction in the Hyporadise benchmark.