https://huggingface.co/collections/artefactory/translation-alignment-analysis-66f3e56669bed67108c309ea

\n","updatedAt":"2024-10-03T08:00:16.372Z","author":{"_id":"65fa95405355a52c784633fc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65fa95405355a52c784633fc/rSfBUHPa7eSAsLd8DuOq4.png","fullname":"Hippolyte Gisserot-Boukhlef","name":"hgissbkh","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":20,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7220274806022644},"editors":["hgissbkh"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/65fa95405355a52c784633fc/rSfBUHPa7eSAsLd8DuOq4.png"],"reactions":[],"isReport":false}},{"id":"66ff46099c32082ec49c3252","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},"createdAt":"2024-10-04T01:34:01.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [Cross-lingual Human-Preference Alignment for Neural Machine Translation with Direct Quality Optimization](https://huggingface.co/papers/2409.17673) (2024)\n* [Model-based Preference Optimization in Abstractive Summarization without Human Feedback](https://huggingface.co/papers/2409.18618) (2024)\n* [Instruction-tuned Large Language Models for Machine Translation in the Medical Domain](https://huggingface.co/papers/2408.16440) (2024)\n* [ASFT: Aligned Supervised Fine-Tuning through Absolute Likelihood](https://huggingface.co/papers/2409.10571) (2024)\n* [Progressively Selective Label Enhancement for Language Model Alignment](https://huggingface.co/papers/2408.02599) (2024)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend

\n","updatedAt":"2024-10-04T01:34:01.827Z","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":0.7558488845825195},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2409.20059","authors":[{"_id":"66fd33a20cde4879f9ab4654","user":{"_id":"65fa95405355a52c784633fc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65fa95405355a52c784633fc/rSfBUHPa7eSAsLd8DuOq4.png","isPro":false,"fullname":"Hippolyte Gisserot-Boukhlef","user":"hgissbkh","type":"user"},"name":"Hippolyte Gisserot-Boukhlef","status":"claimed_verified","statusLastChangedAt":"2024-10-02T12:44:47.000Z","hidden":false},{"_id":"66fd33a20cde4879f9ab4655","user":{"_id":"60d9a9791fa5d458da77754d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1624887555363-60d9a9791fa5d458da77754d.jpeg","isPro":false,"fullname":"Ricardo Rei","user":"RicardoRei","type":"user"},"name":"Ricardo Rei","status":"claimed_verified","statusLastChangedAt":"2025-03-16T21:23:48.804Z","hidden":false},{"_id":"66fd33a20cde4879f9ab4656","user":{"_id":"66f2d6a684a241caac8e16dc","avatarUrl":"/avatars/81acb87c2b07bea938251b40a2139911.svg","isPro":false,"fullname":"Emmanuel Malherbe","user":"emmanuelmalherbe","type":"user"},"name":"Emmanuel Malherbe","status":"admin_assigned","statusLastChangedAt":"2024-10-03T08:53:40.295Z","hidden":false},{"_id":"66fd33a20cde4879f9ab4657","user":{"_id":"61efea03a57920a251ec19b8","avatarUrl":"/avatars/f47c8e3cb17a2bf7d43f2c152bb86885.svg","isPro":false,"fullname":"Celine Hudelot","user":"CelineH","type":"user"},"name":"Céline Hudelot","status":"admin_assigned","statusLastChangedAt":"2024-10-03T08:53:47.051Z","hidden":false},{"_id":"66fd33a20cde4879f9ab4658","user":{"_id":"644a900e3a619fe72b14af0f","avatarUrl":"/avatars/e2d5dac3d92757ed48e37e126a3464a3.svg","isPro":false,"fullname":"Colombo","user":"PierreColombo","type":"user"},"name":"Pierre Colombo","status":"admin_assigned","statusLastChangedAt":"2024-10-03T08:53:53.386Z","hidden":false},{"_id":"66fd33a20cde4879f9ab4659","user":{"_id":"61f06c8ba57920a251ec1a5c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/61f06c8ba57920a251ec1a5c/TmmSN9xWm1noLeAtsUSRk.jpeg","isPro":false,"fullname":"Nuno M. Guerreiro","user":"nunonmg","type":"user"},"name":"Nuno M. Guerreiro","status":"admin_assigned","statusLastChangedAt":"2024-10-03T08:54:03.564Z","hidden":false}],"publishedAt":"2024-09-30T08:01:44.000Z","submittedOnDailyAt":"2024-10-03T06:30:16.351Z","title":"Is Preference Alignment Always the Best Option to Enhance LLM-Based\n Translation? An Empirical Analysis","submittedOnDailyBy":{"_id":"65fa95405355a52c784633fc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65fa95405355a52c784633fc/rSfBUHPa7eSAsLd8DuOq4.png","isPro":false,"fullname":"Hippolyte Gisserot-Boukhlef","user":"hgissbkh","type":"user"},"summary":"Neural metrics for machine translation (MT) evaluation have become\nincreasingly prominent due to their superior correlation with human judgments\ncompared to traditional lexical metrics. Researchers have therefore utilized\nneural metrics through quality-informed decoding strategies, achieving better\nresults than likelihood-based methods. With the rise of Large Language Models\n(LLMs), preference-based alignment techniques have gained attention for their\npotential to enhance translation quality by optimizing model weights directly\non preferences induced by quality estimators. This study focuses on Contrastive\nPreference Optimization (CPO) and conducts extensive experiments to evaluate\nthe impact of preference-based alignment on translation quality. Our findings\nindicate that while CPO consistently outperforms Supervised Fine-Tuning (SFT)\non high-quality data with regard to the alignment metric, it may lead to\ninstability across downstream evaluation metrics, particularly between neural\nand lexical ones. Additionally, we demonstrate that relying solely on the base\nmodel for generating candidate translations achieves performance comparable to\nusing multiple external systems, while ensuring better consistency across\ndownstream metrics.","upvotes":16,"discussionId":"66fd33a30cde4879f9ab46ad","ai_summary":"Contrastive Preference Optimization enhances translation quality on high-quality data but may cause instability across different evaluation metrics, while the base model's performance is comparable to using multiple external systems.","ai_keywords":["neural metrics","machine translation","quality-informed decoding","likelihood-based methods","Large Language Models","preference-based alignment","Contrastive Preference Optimization","Supervised Fine-Tuning","alignment metric","lexical metrics"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66fd7539cba6a6c958f4cdb1","avatarUrl":"/avatars/91f66eb9e63c7d9c8018b61361d518ad.svg","isPro":false,"fullname":"Vincent Auriau","user":"VincentAuriau","type":"user"},{"_id":"649aa2322e34c690156205c0","avatarUrl":"/avatars/5eadb3e0bdd301b12d9128765a8659d9.svg","isPro":false,"fullname":"Veronika Shilova","user":"VeronikaShilova","type":"user"},{"_id":"66fd75e04a81f2d3ee0becbf","avatarUrl":"/avatars/e83fecf5af04d9ecdc7da04d31505ca8.svg","isPro":false,"fullname":"Abdoulaye","user":"abd-sakho","type":"user"},{"_id":"66f2d6a684a241caac8e16dc","avatarUrl":"/avatars/81acb87c2b07bea938251b40a2139911.svg","isPro":false,"fullname":"Emmanuel Malherbe","user":"emmanuelmalherbe","type":"user"},{"_id":"653a661e5cd715dafdafbea5","avatarUrl":"/avatars/c828e6120140d2754667489aa6833feb.svg","isPro":false,"fullname":"Jules","user":"JulesHF","type":"user"},{"_id":"65fa95405355a52c784633fc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65fa95405355a52c784633fc/rSfBUHPa7eSAsLd8DuOq4.png","isPro":false,"fullname":"Hippolyte Gisserot-Boukhlef","user":"hgissbkh","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":"6489609284f4f8799340b28d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6489609284f4f8799340b28d/nxME-CTWj8woeQsn45g02.png","isPro":false,"fullname":"Eyel","user":"Eyel","type":"user"},{"_id":"625921d05f80a3c1aad0bae3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/625921d05f80a3c1aad0bae3/ElN3-6V5nGId2fzI3Dqlr.jpeg","isPro":true,"fullname":"Phi","user":"Xalphinions","type":"user"},{"_id":"648eb1eb59c4e5c87dc116e0","avatarUrl":"/avatars/c636cea39c2c0937f01398c94ead5dad.svg","isPro":false,"fullname":"fdsqefsgergd","user":"T-representer","type":"user"},{"_id":"6555125a4f361968f0e3aad7","avatarUrl":"/avatars/e7692d82804338f21ecdc6e731f5c5ea.svg","isPro":false,"fullname":"marinaretikof","user":"marinaretik","type":"user"},{"_id":"6270324ebecab9e2dcf245de","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6270324ebecab9e2dcf245de/cMbtWSasyNlYc9hvsEEzt.jpeg","isPro":false,"fullname":"Kye Gomez","user":"kye","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">

Papers

arxiv:2409.20059

Is Preference Alignment Always the Best Option to Enhance LLM-Based Translation? An Empirical Analysis

Published on Sep 30, 2024

· Submitted by

Hippolyte Gisserot-Boukhlef on Oct 3, 2024

Upvote

Authors:

Hippolyte Gisserot-Boukhlef ,

Ricardo Rei ,

Emmanuel Malherbe ,

Céline Hudelot ,

Pierre Colombo ,

Nuno M. Guerreiro

Abstract

Contrastive Preference Optimization enhances translation quality on high-quality data but may cause instability across different evaluation metrics, while the base model's performance is comparable to using multiple external systems.

AI-generated summary

Neural metrics for machine translation (MT) evaluation have become increasingly prominent due to their superior correlation with human judgments compared to traditional lexical metrics. Researchers have therefore utilized neural metrics through quality-informed decoding strategies, achieving better results than likelihood-based methods. With the rise of Large Language Models (LLMs), preference-based alignment techniques have gained attention for their potential to enhance translation quality by optimizing model weights directly on preferences induced by quality estimators. This study focuses on Contrastive Preference Optimization (CPO) and conducts extensive experiments to evaluate the impact of preference-based alignment on translation quality. Our findings indicate that while CPO consistently outperforms Supervised Fine-Tuning (SFT) on high-quality data with regard to the alignment metric, it may lead to instability across downstream evaluation metrics, particularly between neural and lexical ones. Additionally, we demonstrate that relying solely on the base model for generating candidate translations achieves performance comparable to using multiple external systems, while ensuring better consistency across downstream metrics.