Back to all available

Layerwise attention for finetuning language models

artificial-intelligence
computer-education
data-mining-and-machine-learning
natural-language-and-speech

Abstract

Background. Encoder-based language models have shown successful performance in natural language understanding tasks. Many studies took the standard fine-tuning method that exploits embeddings of the last layer. However, previous studies found that different layers convey different linguistic knowledge, suggesting that the last layer might not be the best for downstream tasks.

Methods. In this paper, as a new fine-tuning method, we propose a layer-wise attention mechanism using a pivot layer. The pivot layer is used to compute attention scores of encoder layers, and we defined three types of pivot layers. We also examined four attention functions and demonstrated that the attention function plays an important role in the layer-wise attention for fine-tuning by experiments.

Results. Our proposed mechanism outperformed the standard fine-tuning method and other recent method in General Language Understanding Evaluation benchmark. By visualizing the attention distributions, we found that the last layer is always not preferable for every General Language Understanding Evaluation benchmark task, and also found that the difference in attention distribution affects the task performance.

Ask to review this manuscript

Notes for potential reviewers

  • Volunteering is not a guarantee that you will be asked to review. There are many reasons: reviewers must be qualified, there should be no conflicts of interest, a minimum of two reviewers have already accepted an invitation, etc.
  • This is NOT OPEN peer review. The review is single-blind, and all recommendations are sent privately to the Academic Editor handling the manuscript. All reviews are published and reviewers can choose to sign their reviews.
  • What happens after volunteering? It may be a few days before you receive an invitation to review with further instructions. You will need to accept the invitation to then become an official referee for the manuscript. If you do not receive an invitation it is for one of many possible reasons as noted above.
  • PeerJ Computer Science does not judge submissions based on subjective measures such as novelty, impact or degree of advance. Effectively, reviewers are asked to comment on whether or not the submission is scientifically and technically sound and therefore deserves to join the scientific literature. Our Peer Review criteria can be found on the "Editorial Criteria" page - reviewers are specifically asked to comment on 3 broad areas: "Basic Reporting", "Experimental Design" and "Validity of the Findings".
  • Reviewers are expected to comment in a timely, professional, and constructive manner.
  • Until the article is published, reviewers must regard all information relating to the submission as strictly confidential.
  • When submitting a review, reviewers are given the option to "sign" their review (i.e. to associate their name with their comments). Otherwise, all review comments remain anonymous.
  • All reviews of published articles are published. This includes manuscript files, peer review comments, author rebuttals and revised materials.
  • Each time a decision is made by the Academic Editor, each reviewer will receive a copy of the Decision Letter (which will include the comments of all reviewers).

If you have any questions about submitting your review, please email us at [email protected].