A Closer Look at How Fine-tuning Changes BERT

A Closer Look at How Fine-tuning Changes BERT Yichu Zhou School of Computing University of Utah flyaway@cs.utah.edu Vivek Srikumar School of Computing University of Utah svivek@cs.utah.edu Abstract Given the prevalence of pre-trained contextual- ized representations in today’s NLP, there have been many efforts to understand what infor- mation they contain, and why they seem to be universally successful. The most common ap- proach to use these representations involves fine-tuning them for an end task. Yet, how fine-tuning changes the underlying embedding space is less studied. In this work, we study the English BERT family and use two probing tech- niques to analyze how fine-tuning changes the space. We hypothesize that fine-tuning affects classification performance by increasing the distances between examples associated with different labels. We confirm this hypothesis with carefully designed experiments on five dif- ferent NLP tasks. Via these experiments, we also discover an exception to the prevailing wis- dom that “fine-tuning always improves perfor- mance”. Finally, by comparing the representa- tions before and after fine-tuning, we discover that fine-tuning does not introduce arbitrary changes to representations; instead, it adjusts the representations to downstream tasks while largely preserving the original spatial structure of the data points. 1 Introduction Pre-trained transformer-based language mod- els (e.g., Devlin et al., 2019) form the basis of state-of-the-art results across NLP. The relative opacity of these models has prompted the devel- opment of many probes to investigate linguistic regularities captured in them (e.g., Kovaleva et al., 2019; Conneau et al., 2018; Jawahar et al., 2019). Broadly speaking, there are two ways to use a pre-trained representation (Peters et al., 2019): as a fixed feature extractor (where the pre-trained weights are frozen), or by fine-tuning it for a task. The probing literature has largely focused on the former (e.g., Kassner and Schütze, 2020; Perone et al., 2018; Yaghoobzadeh et al., 2019; Krasnowska-Kiera´s and Wróblewska, 2019; Wal- lace et al., 2019; Pruksachatkun et al., 2020; Agha- janyan et al., 2021). Some previous work (Mer- chant et al., 2020; Mosbach et al., 2020b; Hao et al., 2020) does provide insights about fine-tuning: fine-tuning changes higher layers more than lower ones and linguistic information is not lost during fine-tuning. However, relatively less is understood about how the representation changes during the process of fine-tuning and why fine-tuning invari- ably seems to improve task performance. In this work, we investigate the process of fine- tuning of representations using the English BERT family (Devlin et al., 2019). Specifically, we ask: 1. Does fine-tuning always improve perfor- mance? 2. How does fine-tuning alter the representation to adjust for downstream tasks? 3. How does fine-tuning change the geometric structure of different layers? We apply two probing techniques—classifier- based probing (Kim et al., 2019; Tenney et al., 2019) and DIRECTPROBE (Zhou and Srikumar, 2021)—on variants of BERT representations that are fine-tuned on five tasks: part-of-speech tag- ging, dependency head prediction, preposition su- persense role & function prediction and text clas- sification. Beyond confirming previous findings about fine-tuning, our analysis reveals several new findings, …

Content truncated to 500 words