On the Interplay Between Fine-tuning and Sentence-level Probing for Linguistic Knowledge in Pre-trained Transformers

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2502–2516 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2502 On the Interplay Between Fine-tuning and Sentence-level Probing for Linguistic Knowledge in Pre-trained Transformers Marius Mosbach Anna Khokhlova Michael A. Hedderich Dietrich Klakow Spoken Language Systems (LSV) Department of Language Science and Technology Saarland Informatics Campus, Saarland University, Germany {mmosbach,akhokhlova,mhedderich,dklakow}@lsv.uni-saarland.de Abstract Fine-tuning pre-trained contextualized embed- ding models has become an integral part of the NLP pipeline. At the same time, prob- ing has emerged as a way to investigate the linguistic knowledge captured by pre-trained models. Very little is, however, understood about how fine-tuning affects the represen- tations of pre-trained models and thereby the linguistic knowledge they encode. This paper contributes towards closing this gap. We study three different pre-trained models: BERT, RoBERTa, and ALBERT, and investi- gate through sentence-level probing how fine- tuning affects their representations. We find that for some probing tasks fine-tuning leads to substantial changes in accuracy, possibly sug- gesting that fine-tuning introduces or even re- moves linguistic knowledge from a pre-trained model. These changes, however, vary greatly across different models, fine-tuning and prob- ing tasks. Our analysis reveals that while fine- tuning indeed changes the representations of a pre-trained model and these changes are typ- ically larger for higher layers, only in very few cases, fine-tuning has a positive effect on probing accuracy that is larger than just us- ing the pre-trained model with a strong pool- ing method. Based on our findings, we argue that both positive and negative effects of fine- tuning on probing require a careful interpreta- tion. 1 Introduction Transformer-based contextual embeddings like BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019b) and ALBERT (Lan et al., 2020) recently be- came the state-of-the-art on a variety of NLP down- stream tasks. These models are pre-trained on large amounts of text and subsequently fine-tuned on task-specific, supervised downstream tasks. Their strong empirical performance triggered questions concerning the linguistic knowledge they encode in their representations and how it is affected by the training objective and model architecture (Kim et al., 2019; Wang et al., 2019a). One promi- nent technique to gain insights about the linguis- tic knowledge encoded in pre-trained models is probing (Rogers et al., 2020). However, works on probing have so far focused mostly on pre-trained models. It is still unclear how the representations of a pre-trained model change when fine-tuning on a downstream task. Further, little is known about whether and to what extent this process adds or removes linguistic knowledge from a pre-trained model. Addressing these issues, we are investigat- ing the following questions: 1. How and where does fine-tuning affect the representations of a pre-trained model? 2. To which extent (if at all) can changes in prob- ing accuracy be attributed to a change in lin- guistic knowledge encoded by the model? To answer these questions, we investigate three different pre-trained encoder models, BERT, RoBERTa, and ALBERT. We fine-tune them on sentence-level classification tasks from …

Content truncated to 500 words