Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GOVERN: Gradient Orientation Vote Ensemble for Multi-Teacher Reinforced Distillation

Wenjie Zhou, Zhenxin Ding, Xiaodong Zhang, Haibo Shi, Junfeng Wang, Dawei Yin

TL;DR

A novel algorithm, GOVERN, is proposed to tackle the issue of how can knowledge from multiple teacher models be effectively ensemble during this stage without the guidance of labels, and has been successfully deployed in a real-world commercial question-answering system.

Abstract

Pre-trained language models have become an integral component of question-answering systems, achieving remarkable performance. However, for practical deployment, it is crucial to perform knowledge distillation to maintain high performance while operating under computational constraints. In this paper, we address a key question: given the importance of unsupervised distillation for student model performance, how can knowledge from multiple teacher models be effectively ensemble during this stage without the guidance of labels? We propose a novel algorithm, GOVERN, to tackle this issue. GOVERN has demonstrated significant improvements in both offline and online experiments, enabling the student model to achieve results comparable to that of teacher ensembles. Our experiments show that GOVERN remarkably requires a mere 1\% of the ensemble method's inference budget to achieve 99.5\% of performance. The proposed algorithm has been successfully deployed in a real-world commercial question-answering system, demonstrating its real-world applicability.

GOVERN: Gradient Orientation Vote Ensemble for Multi-Teacher Reinforced Distillation

TL;DR

A novel algorithm, GOVERN, is proposed to tackle the issue of how can knowledge from multiple teacher models be effectively ensemble during this stage without the guidance of labels, and has been successfully deployed in a real-world commercial question-answering system.

Abstract

Pre-trained language models have become an integral component of question-answering systems, achieving remarkable performance. However, for practical deployment, it is crucial to perform knowledge distillation to maintain high performance while operating under computational constraints. In this paper, we address a key question: given the importance of unsupervised distillation for student model performance, how can knowledge from multiple teacher models be effectively ensemble during this stage without the guidance of labels? We propose a novel algorithm, GOVERN, to tackle this issue. GOVERN has demonstrated significant improvements in both offline and online experiments, enabling the student model to achieve results comparable to that of teacher ensembles. Our experiments show that GOVERN remarkably requires a mere 1\% of the ensemble method's inference budget to achieve 99.5\% of performance. The proposed algorithm has been successfully deployed in a real-world commercial question-answering system, demonstrating its real-world applicability.
Paper Structure (18 sections, 12 equations, 5 figures, 4 tables)

This paper contains 18 sections, 12 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Answer Selection Task
  3. Methodology
  4. Unsupervised Distillation
  5. GOVERN
  6. Supervised Distillation: GOVERN-CA
  7. Experiments and Results
  8. Dataset
  9. Experiment Details
  10. Evaluation Metrics
  11. Main Results
  12. Online Experiment
  13. Ablation Study
  14. Related Work
  15. Conclusion
  16. ...and 3 more sections

Figures (5)

  • Figure 1: Procedures of Gradient Orientation Vote Ensemble Reinforced Distillation
  • Figure 2: The Answer Card is retrieve by the question answering system. Web pages below are not display in answer card format.
  • Figure 3: The effect of the number of teachers.
  • Figure 4: Left part shows the distribution of our model's output on test set, and right part shows the distribution of $Beta(19.0, 3.0)$. We can see that the model's output keep similar distribution with Beta function.
  • Figure 5: