Don't Stop Pretraining? Make Prompt-based Fine-tuning Powerful Learner

Language models (LMs) trained on vast quantities of unlabelled data have greatly advanced the field of natural language processing (NLP). In this study, we re-visit the widely accepted notion in NLP that continued pre-training LMs on task-related texts improves the performance of fine-tuning (FT) in downstream tasks. Through experiments on eight single-sentence tasks and eight sentence-pair tasks in both semi-supervised and fully-supervised settings, we find that conventional continued pre-training does not consistently provide benefits and can even be detrimental for sentence-pair tasks or when prompt-based FT is used. To tackle these issues, we propose Prompt-based Continued Pre-training (PCP), which combines the idea of instruction tuning with conventional continued pre-training. Our approach aims to improve the performance of prompt-based FT by presenting both task-related texts and prompt templates to LMs through unsupervised pre-training objectives before fine-tuning for the target task. Our empirical evaluations on 21 benchmarks demonstrate that the PCP consistently improves the performance of state-of-the-art prompt-based FT approaches (up to 20.1% absolute) in both semi-supervised and fully-supervised settings, even with only hundreds of unlabelled examples. Additionally, prompt-based FT with the PCP outperforms state-of-the-art semi-supervised approaches with greater simplicity, eliminating the need for an iterative process and extra data augmentation. Our further analysis explores the performance lower bound of the PCP and reveals that the advantages of PCP persist across different sizes of models and datasets.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer bM1j7/10 · confidence 5/52023-06-29

Summary

The paper proposes a new approach to pre-training language models called Prompt-based Continued Pre-training (PCP). PCP combines the idea of instruction tuning with conventional continued pre-training. The authors argue that PCP can improve the performance of prompt-based fine-tuning on a variety of natural language processing tasks.

Strengths

1. The paper presents a clear comparison of the proposed PCP with the conventional continued pre-training. The evidence on benchmark datasets is strong to support the claim that PCP performs better on these tasks. 2. The paper is well-written and easy to understand. The authors do a good job of explaining the technical details of the proposed approach, and they provide clear and concise summaries of the experimental results. 3. The approach is simple and easy to adopt.

Weaknesses

1. It is not well understood why conventional continued pretraining (TAPT) is doing so poorly on sentence-pair tasks. 2. Because PCP makes use of templated prompts to align continued pretraining and fine-tuning, with additional pseudo-labels. It is unclear whether the gain is due to prompt alignment or the additional data augmentation from pseudo-labels. Further ablation is necessary. 3. It is unclear how well the proposed approach generalizes to other types of pretraining objectives such as language modeling.

Questions

1. More ablations and analysis are needed to understand why TAPT is not working for sentence-pair tasks. 2. More experiments to ablate the effect of template alignment vs data augmentation via pseudo-labels.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Generalization of the approach to other pretraining objectives is unclear.

Reviewer fyTk6/10 · confidence 4/52023-07-03

Summary

This work re-examines a well-known technique in the NLP literature, called continued pre-training, that can be utilized to enhance the performance of language models on downstream tasks (in this case, for classification and regression tasks). The authors have revealed that the conventional approach to continued pre-training is ineffective when applied to sentence-pair tasks and prompt-based fine-tuning settings. Based on their findings, the authors propose a simple method that involves training on the masking language modeling objective while augmenting the text input with prompts. This proposed approach has demonstrated its effectiveness across a range of tasks and configurations, suggesting that it can be an effective choice for natural language understanding tasks when using Transformer encoder-based models.

Strengths

- The proposed method is technically simple and effective. - The work begins with a compelling finding that challenges the effectiveness of the conventional approach, which was previously regarded as effective. - The authors made efforts to provide analysis from multiple perspectives, aiming to convince readers of the effectiveness of the proposed approach.

Weaknesses

- Although the paper offers a detailed explanation and analysis, there are still ambiguous points that need to be clarified for a better understanding by the readers. Let me ask about this point in the following Questions section. - There is room for improving the grammar and fluency of the paper's writing. - I would like to see a thorough analysis of the factors that potentially contribute to the success of the proposed method. For instance, it would be valuable to explore the significance of including labels in continued fine-tuning (although partially considered in Section 4.4) and investigate why sentence-pair tasks do not benefit from the conventional TAPT approach.

Questions

To the best of my understanding, it appears that your method incorporates prompts and their corresponding verbalized labels into the process of continued pre-training. If this is the case, there is a possibility that the tokens for verbalized labels are chosen as the target for masked language modeling. This implies that the method might be performing a similar task during both continued pre-training and fine-tuning. Consequently, there is a potential risk that the effectiveness of the proposed method is primarily attributed to unintentionally prolonged fine-tuning. If this turns out to be true, it could diminish the significance of this work. To alleviate any ambiguity regarding this point, it would be helpful if you could provide concrete examples demonstrating how your method is applied to real data instances. This would offer a clearer understanding of the methodology and help evaluate its effectiveness more accurately.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The evaluation primarily focuses on the utilization of a single model, specifically RoBERTa-large, for relatively straightforward downstream tasks such as classification. These tasks are considered comparatively easier to solve.

Reviewer pe8v7/10 · confidence 3/52023-07-04

Summary

This paper explores the problem of continued pre-training on task-related text. The authors discovered that conventional continued pre-training methods may not be very effective and can even have a negative impact on fine-tuning performance. To address this, they introduce prompt-based continued pre-training. The approach involves generating pseudo labels on unlabeled data using a fine-tuned model and constructing a prompt-based pre-training corpus by applying templates to the pseudo labeled dataset. The researchers then utilize this corpus, which incorporates prompt information, for continued pre-training and prompt-based fine-tuning. Experimental evaluations conducted on various datasets validate the effectiveness of the proposed approach, as it achieves significant improvements over the baseline methods.

Strengths

* The researchers identify the limitations of conventional continued pre-training on task-related text and propose an innovative solution. * The experimental results provide strong evidence for the effectiveness of the proposed approach.

Weaknesses

No major weaknesses have been identified in this paper.

Questions

Is it possible to extend this approach to the task of language modeling fine-tuning?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer TYSB5/10 · confidence 4/52023-07-07

Summary

This paper makes a contribution by studying how to adapt pre-trained models to downstream tasks. The authors identify the limitations of TAPTs and show when they do not work well. They then propose PCP, a better algorithm that can adapt a pre-trained model to a target task. PCP is shown to be more effective than TAPTs on a variety of tasks.

Strengths

The paper provides an intriguing analysis of when TAPT style is effective for fine-tuning. TAPT is not very effective on sentence-pair tasks or prompt-based fine-tuning, particularly.

Weaknesses

The paper does a good job of pointing out the weaknesses of TAPT, but it is less clear why PCP works better. The authors do not need to know exactly why it works, but some hypotheses would be helpful. For example, why does PCP-style pretraining work better for prompt-based fine-tuning? Is it because the "pretraining" is more similar to "fine-tuning" in PCP? If so, how can we understand why PCP also works for sentence pair tasks? The presentation of the paper could also be improved. Figure 2 is not very easy to understand, and the overall flow of the paper is not as clear as it could be. Overall, the paper could be improved by providing more clarity and explanation.

Questions

Cloud the authors provide more intuitions or analysis why PCP works better? Based on the intuitions, what kind of additional analysis can be done here?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The paper only applies on text classifications tasks, ignoring many text generation tasks.

Reviewer CwVD5/10 · confidence 4/52023-07-09

Summary

The proposed method (CPC) is built on top of pseudo-labeling and continued pre-training via masked language modeling. This method provides an alternative way to use pseudo-labeled data before fine-tuning the model on downstream tasks. It improves the TAPT method and other semi-supervised methods for text classification.

Strengths

- The novelty of this paper lies in using pseudo-labeled data for masked language modeling. This idea is simple, yet effective. Even though pseudo-labeling and continued MLM training are not new, the way to combine these two is new. - Extensive experiments on a large number of text classification tasks show the effectiveness of the method.

Weaknesses

- One of the motivations (mentioned in the introduction) is that TAPT does not work well on sentence pair classification. However, it’s still unclear why TAPT does not work well on sentence pair classification while CPC can address this issue. - Although CPC works empirically better than TAPT, the intuition of CPC is still unclear. Why should we continue pre-training on pseudo-labeled data instead of unlabeled data in TAPT?

Questions

1. What masking strategy is used in Step 2? Are you using the random masking strategy as used in Robert? 2. If you have pseudo labels in CPC, is random masking in Step 2 the optimal strategy? How about some selective masking strategies (e.g., PMI-masking [1])? 3. The paper is missing a self-training baseline, where you only mask the pseudo labels for MLM pre-training in Step 2. 4. It’d be good to include the scores of TAPT in Fig 3, representing “no label + FT”. It will be helpful to compare “wrong label + FT” and “no label + FT”. Because the method is sensitive to the pseudo-labeling performance, it'd be good to add a few-shot setting where the base model is fine-tuned on few-shot examples in Step 1. And compare CPC with TAPT in the few-shot setup. 5. To better compare TAPT (“no label +FT”) with CPC (“pseudo-label +FT”), we should have a better understanding of the impact of the “pseudo-label” in the MLM process. Does the addition of “pseudo-label” in the input sentence improve the MLM prediction accuracy? [1] PMI-Masking: Principled masking of correlated spans

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer bM1j2023-08-11

I appreciate the due-diligence the authors took to address some of the concerns the reviewers including myself raised. I don't have further concerns despite that we still don't fully understand why TAPT performs poorly on sentence pair tasks (the authors did a very thorough ablation given the limited time). I would change my Rating to 7: Accept.

Authorsrebuttal2023-08-11

We are very glad to hear that our response is helpful! May we kindly ask the reviewer to consider updating the score officially, as the organizers have now rectified the issue with updating scores? Thank you again for your understanding and assistance.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC