[2101.11492] On the Evolution of Syntactic Information Encoded by BERT's Contextualized Representations

The adaptation of pretrained language models to solve supervised tasks has become a baseline in NLP, and many recent works have focused on studying how linguistic information is encoded in the pretrained sentence representations. Among other information, it has been shown that entire syntax trees are implicitly embedded in the geometry of such models. As these models are often fine-tuned, it becomes increasingly important to understand how the encoded knowledge evolves along the fine-tuning. In this paper, we analyze the evolution of the embedded syntax trees along the fine-tuning process of BERT for six different tasks, covering all levels of the linguistic structure. Experimental results show that the encoded syntactic information is forgotten (PoS tagging), reinforced (dependency and constituency parsing) or preserved (semantics-related tasks) in different ways along the fine-tuning process depending on the task.

1 mentions: @stanfordnlp
Keywords: bert
Date: 2021/02/21 02:21

Related Entries

Read more Stanford CS 224N (NLP with Deep Learning) Poster Session
0 users, 2 mentions 2019/03/19 15:47
Read more The Brains Behind AI
0 users, 1 mentions 2020/02/01 08:21
Read more The Stanford Natural Language Processing Group
1 users, 1 mentions 2020/04/07 17:21
Read more 1st Workshop on NLI
0 users, 2 mentions 2020/04/13 06:52
Read more Postdoctoral Fellowship in Knowledge Graphs and Natural Language Processing
0 users, 1 mentions 2020/04/15 17:21