Published Version
Open DOI LinkAbstract
Transfer learning from pre-trained neural language models towards downstream tasks has been a predominant theme in NLP recently. Several researchers have shown that deep NLP models learn non-trivial amount of linguistic knowledge, captured at different layers of the model. We investigate how fine-tuning towards downstream NLP tasks impacts the learned linguistic knowledge. We carry out a study across popular pre-trained models BERT, RoBERTa and XLNet using layer and neuron-level diagnostic classifiers. We found that for some GLUE tasks, the network relies on the core linguistic information and preserve it deeper in the network, while for others it forgets. Linguistic information is distributed in the pre-trained language models but becomes localized to the lower layers post fine-tuning, reserving higher layers for the task specific knowledge. The pattern varies across architectures, with BERT retaining linguistic information relatively deeper in the network compared to RoBERTa and XLNet, where it is predominantly delegated to the lower layers.
Highlights
Contextualized word representations learned in transformer-based language models capture rich linguistic knowledge, making them ubiquitous for transfer learning towards downstream NLP problems such as Natural Language Understanding tasks e.g. GLUE (Wang et al, 2018)
We investigate i) if the fine-tuned models retain the same amount of linguistic information, ii) how this information is redistributed across different layers and individual neurons
Comparing GLUE tasks: We found that linguistic phenomena are more important for certain downstream tasks, for example STS, RTE and MRPC where they are preserved in the higher layers post fine-tuning, as opposed to others, for example SST, QNLI and MNLI where they are forgotten in the higher layers
Read more 
Summary
Introduction
Contextualized word representations learned in transformer-based language models capture rich linguistic knowledge, making them ubiquitous for transfer learning towards downstream NLP problems such as Natural Language Understanding tasks e.g. GLUE (Wang et al, 2018). Descriptive methods in neural interpretability investigate what knowledge is learned within the representations through relevant extrinsic phenomenon varying from word morphology (Vylomova et al, 2016; Belinkov et al, 2017a; Dalvi et al, 2017) to high level concepts such as structure (Shi et al, 2016; Linzen et al, 2016) and semantics (Qian et al, 2016; Belinkov et al, 2017b) or more generic properties such as sentence length (Adi et al, 2016; Bau et al, 2019) These studies are carried towards analyzing representations from pre-trained models. Our work complements their findings while extending the layer-wise analysis to core-linguistic tasks and looking at the distribution and relocation of neurons after fine-tuning
Methodology
Experimental Setup
Layer-wise Probing
Neuron-wise Probing
Network Pruning
Conclusion
Data and Representations
Top versus Bottom Neurons
Pruning Layers
Control Tasks
Infrastructure and Run Time
Findings
Hyperparameters
Sign-in/Register to view highlights & summary Full Text
Open DOI Link Sign-in/Register to access full text options Topics from this Paper
Pre-trained Language Models- Downstream Tasks
- Linguistic Knowledge
- Pre-trained Models
- Linguistic Information + Show 5 more
Create a personalized feed of these topics
Get StartedTalk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callSimilar Papers
How transfer learning impacts linguistic knowledge in deep NLP models? 
Relevant
Not Relevant
Aug 1, 2021
RelevantNot Relevant StaResGRU-CNN with CMedLMs: A stacked residual GRU-CNN with pre-trained biomedical language models for predictive intelligence RelevantNot Relevant
Applied Soft Computing- Dec 1, 2021
RelevantNot Relevant Robust Transfer Learning with Pretrained Language Models through Adapters RelevantNot Relevant
- Aug 1, 2021
RelevantNot Relevant Neural Transfer Learning For Vietnamese Sentiment Analysis Using Pre-trained Contextual Language Models RelevantNot Relevant
- Dec 16, 2021
RelevantNot Relevant Towards an Enhanced Understanding of Bias in Pre-trained Neural Language Models: A Survey with Special Emphasis on Affective Bias RelevantNot Relevant
- Jan 1, 2022
RelevantNot Relevant Chinese-Korean Weibo Sentiment Classification Based on Pre-trained Language Model and Transfer Learning RelevantNot Relevant
- May 6, 2022
RelevantNot Relevant Prompt Learning with Structured Semantic Knowledge Makes Pre-Trained Language Models Better RelevantNot Relevant
- Electronics
- Jul 30, 2023
RelevantNot Relevant Multi-Faceted Knowledge-Driven Pre-Training for Product Representation Learning RelevantNot Relevant
- IEEE Transactions on Knowledge and Data Engineering
- Jan 1, 2022
RelevantNot Relevant A Multi-tasking and Multi-stage Chinese Minority Pre-trained Language Model RelevantNot Relevant
- Jan 1, 2022
RelevantNot Relevant NoisyTune: A Little Noise Can Help You Finetune Pretrained Language Models Better RelevantNot Relevant
- May 11, 2022
RelevantNot Relevant PANLP at MEDIQA 2019: Pre-trained Language Models, Transfer Learning and Knowledge Distillation RelevantNot Relevant
- Jan 1, 2019
RelevantNot Relevant PPT: Backdoor Attacks on Pre-trained Models via Poisoned Prompt Tuning RelevantNot Relevant
- Jul 1, 2022
RelevantNot Relevant NoisyTune: A Little Noise Can Help You Finetune Pretrained Language Models Better RelevantNot Relevant
- Jan 1, 2022
RelevantNot Relevant