Soft Alignment Objectives for Robust Adaptation of Language
Generation

D 2023

Soft Alignment Objectives for Robust Adaptation of Language Generation

ŠTEFÁNIK, Michal, Marek KADLČÍK a Petr SOJKA

Základní údaje

Originální název

Soft Alignment Objectives for Robust Adaptation of Language Generation

Autoři

ŠTEFÁNIK, Michal (703 Slovensko, domácí), Marek KADLČÍK (203 Česká republika, domácí) a Petr SOJKA (203 Česká republika, domácí)

Vydání

Toronto, Canada, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), od s. 8837-8853, 17 s. 2023

Nakladatel

Association for Computational Linguistics

Další údaje

Jazyk

angličtina

Typ výsledku

Stať ve sborníku

Obor

10201 Computer sciences, information science, bioinformatics

Stát vydavatele

Spojené státy

Utajení

není předmětem státního či obchodního tajemství

Forma vydání

elektronická verze "online"

Odkazy

URL

Kód RIV

RIV/00216224:14330/23:00130912

Organizační jednotka

Fakulta informatiky

ISBN

978-1-959429-72-2

ISSN

DOI

http://dx.doi.org/10.18653/v1/2023.acl-long.492

UT WoS

001190962500024

Klíčová slova anglicky

generation; robustness; machine translation; adaptation

Štítky

core_A, firank_1

Příznaky

Mezinárodní význam, Recenzováno

Změněno: 27. 6. 2024 11:07, Mgr. Michal Petr

Anotace

V originále

Domain adaptation allows generative language models to address specific flaws caused by the domain shift of their application. However, the traditional adaptation by further training on in-domain data rapidly weakens the model's ability to generalize to other domains, making the open-ended deployments of the adapted models prone to errors. This work introduces novel training objectives built upon a semantic similarity of the predicted tokens to the reference. Our results show that (1) avoiding the common assumption of a single correct prediction by constructing the training target from tokens' semantic similarity can largely mitigate catastrophic forgetting of adaptation, while (2) preserving the adaptation in-domain quality, (3) with negligible additions to compute costs. In the broader context, the objectives grounded in a continuous token similarity pioneer the exploration of the middle ground between the efficient but na\"{\i}ve exact-match token-level objectives and expressive but computationally- and resource-intensive sequential objectives.

Návaznosti

MUNI/A/1339/2022, interní kód MU

Název: Rozvoj technik pro zpracování dat pro podporu vyhledávání, analýz a vizualizací rozsáhlých datových souborů s využitím umělé inteligence

Investor: Masarykova univerzita, Rozvoj technik pro zpracování dat pro podporu vyhledávání, analýz a vizualizací rozsáhlých datových souborů s využitím umělé inteligence

MUNI/A/1433/2022, interní kód MU

Název: Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 23

Investor: Masarykova univerzita, Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 23

90129, velká výzkumná infrastruktura

Název: Czech-BioImaging II

Citovat

ŠTEFÁNIK, Michal, Marek KADLČÍK a Petr SOJKA. Soft Alignment Objectives for Robust Adaptation of Language Generation. Online. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Linguistics, 2023, s. 8837-8853. ISBN 978-1-959429-72-2. Dostupné z: https://dx.doi.org/10.18653/v1/2023.acl-long.492.

@inproceedings{2288099,
   author = {Štefánik, Michal and Kadlčík, Marek and Sojka, Petr},
   address = {Toronto, Canada},
   booktitle = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
   doi = {http://dx.doi.org/10.18653/v1/2023.acl-long.492},
   keywords = {generation; robustness; machine translation; adaptation},
   howpublished = {elektronická verze "online"},
   language = {eng},
   location = {Toronto, Canada},
   isbn = {978-1-959429-72-2},
   note = {To appear},
   pages = {8837-8853},
   publisher = {Association for Computational Linguistics},
   title = {Soft Alignment Objectives for Robust Adaptation of Language Generation},
   url = {https://aclanthology.org/2023.acl-long.492},
   year = {2023}
}

TY  - JOUR
ID  - 2288099
AU  - Štefánik, Michal - Kadlčík, Marek - Sojka, Petr
PY  - 2023
TI  - Soft Alignment Objectives for Robust Adaptation of Language Generation
PB  - Association for Computational Linguistics
CY  - Toronto, Canada
SN  - 9781959429722
N1  - To appear
KW  - generation
KW  - robustness
KW  - machine translation
KW  - adaptation
UR  - https://aclanthology.org/2023.acl-long.492
N2  - Domain adaptation allows generative language models to address specific flaws caused by the domain shift of their application. However, the traditional adaptation by further training on in-domain data rapidly weakens the model's ability to generalize to other domains, making the open-ended deployments of the adapted models prone to errors. This work introduces novel training objectives built upon a semantic similarity of the predicted tokens to the reference. Our results show that (1) avoiding the common assumption of a single correct prediction by constructing the training target from tokens' semantic similarity can largely mitigate catastrophic forgetting of adaptation, while (2) preserving the adaptation in-domain quality, (3) with negligible additions to compute costs. In the broader context, the objectives grounded in a continuous token similarity pioneer the exploration of the middle ground between the efficient but na\"{\i}ve exact-match token-level objectives and expressive but computationally- and resource-intensive sequential objectives.
ER  -

ŠTEFÁNIK, Michal, Marek KADLČÍK a Petr SOJKA. Soft Alignment Objectives for Robust Adaptation of Language Generation. Online. In \textit{Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}. Toronto, Canada: Association for Computational Linguistics, 2023, s.~8837-8853. ISBN~978-1-959429-72-2. Dostupné z: https://dx.doi.org/10.18653/v1/2023.acl-long.492.

Podrobný výpis o publikaci