BERTnesia: Investigating the capture and forgetting of knowledge in BERT

TL;DR

This paper investigates how BERT captures and forgets relational knowledge across its layers, revealing that intermediate layers hold significant knowledge and that fine-tuning leads to forgetting, with implications for understanding model representations.

Contribution

It provides a detailed analysis of relational knowledge distribution in BERT layers and how fine-tuning affects knowledge retention and forgetting.

Findings

Intermediate layers contain 17-60% of total relational knowledge.

Fine-tuning causes BERT to forget relational knowledge, especially in certain tasks.

Ranking models retain more knowledge and forget less during fine-tuning.

Abstract

Probing complex language models has recently revealed several insights into linguistic and semantic patterns found in the learned representations. In this paper, we probe BERT specifically to understand and measure the relational knowledge it captures. We utilize knowledge base completion tasks to probe every layer of pre-trained as well as fine-tuned BERT (ranking, question answering, NER). Our findings show that knowledge is not just contained in BERT's final layers. Intermediate layers contribute a significant amount (17-60%) to the total knowledge found. Probing intermediate layers also reveals how different types of knowledge emerge at varying rates. When BERT is fine-tuned, relational knowledge is forgotten but the extent of forgetting is impacted by the fine-tuning objective but not the size of the dataset. We found that ranking models forget the least and retain more knowledge in their final layer. We release our code on github to repeat the experiments.