Adversarial Fine-Tuning of Pretrained Language Models

Βερνίκος, Γεώργιος; Vernikos, Georgios

dc.contributor.author	Βερνίκος, Γεώργιος	el
dc.contributor.author	Vernikos, Georgios	en
dc.date.accessioned	2020-12-03T07:05:45Z
dc.date.available	2020-12-03T07:05:45Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/52190
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.19888
dc.description	Εθνικό Μετσόβιο Πολυτεχνείο--Μεταπτυχιακή Εργασία. Διεπιστημονικό-Διατμηματικό Πρόγραμμα Μεταπτυχιακών Σπουδών (Δ.Π.Μ.Σ.) “Επιστήμη Δεδομένων και Μηχανική Μάθηση”	el
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Adversarial	en
dc.subject	Fine-tuning	en
dc.subject	Language models	en
dc.subject	Transformers	en
dc.subject	Transfer learning	en
dc.subject	Αντιπαλικός όρος ποινής	el
dc.subject	Προσαρμογή	el
dc.subject	Γλωσσικά μοντέλα	el
dc.subject	Μεταφορά μάθησης	el
dc.subject	Εξομαλυντής	el
dc.title	Adversarial Fine-Tuning of Pretrained Language Models	en
dc.title	Τεχνικές Μεταφοράς Μάθησης για την Προσαρμογή Προεκπαιδευμένων Γλωσσικών Μοντέλων	el
dc.contributor.department	Artificial Intelligence and Learning Systems Laboratory (AILS Lab)	el
heal.type	masterThesis
heal.classification	Artificial Intelligence	en
heal.classification	Machine Learning	en
heal.classification	Deep Learning	en
heal.classification	Τεχνητή Νοημοσύνη	el
heal.classification	Μηχανική Μάθηση	el
heal.classification	Βαθιά Μάθηση	el
heal.language	el
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2020-07-07
heal.abstract	In this work, we investigate methods in order to effectively transfer knowledge from a pretrained model to downstream tasks. Our goal is to improve the performance of pretrained language models on natural language processing tasks. We evaluate our approach on four natural language understanding tasks: sentence acceptability, sentiment analysis, paraphrase detection and textual entailment. Pretrained language models have achieved state-of-the-art results on most natural language processing tasks. Part of this success lies in their pretraining on large unlabeled corpora. The transferring process of language models consists of further training (fine-tuning) on the task-specific dataset. We argue that this process can hinder the knowledge captured during pretraining, a phenomenon known as \textit{catastrophic forgetting}. In our work, we identify the emergence of too domain-specific features during fine-tuning as a form of catastrophic forgetting. We aim to effectively transfer the knowledge gained during the pretraining of language models to the adaptation process. In order to preserve most of the knowledge captured during pretraining and exploit the capabilities of pretrained language models, we introduce \textsc{after}, short for domain adversarial fine-tuning as an effective regularizer. We leverage unlabeled data from a different domain than the task-specific dataset and we constrain the extent to which the model representations are allowed to differ across different domains. This constraint is realized through a classifier that tries to distinguish between domains. The parameters of the pretrained language model are trained adversarially to maximize the loss of this classifier. The addition of this domain adversarial loss has a regularizing effect on the fine-tuning process, encouraging domain-invariant representations and subsequently leading to improved performance on downstream tasks. We experiment with two top-performing pretrained language models (BERT and XLNet), although our approach is equally applicable to any language model. The proposed adversarial fine-tuning method, \textsc{after}, outperforms standard fine-tuning on four natural language understanding tasks using two different pretrained language models. We additionally conduct an ablation study regarding the effect of the domain of origin of the unlabeled corpus and the similarity between the domain of the pretraining and the task-specific data. Our adversarial fine-tuning approach requires minimal changes on the fine-tuning process and leverages unlabeled data.	en
heal.advisorName	Σταφυλοπάτης, Ανδρέας-Γεώργιος	el
heal.advisorName	Stafylopatis, Andreas-Georgios	en
heal.committeeMemberName	Στάμου, Γιώργος	el
heal.committeeMemberName	Σιόλας, Γεώργιος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	126 σ.	el
heal.fullTextAvailability	false