Open Access

ARTICLE

DPAL-BERT: A Faster and Lighter Question Answering Model

Lirong Yin¹, Lei Wang¹, Zhuohang Cai², Siyu Lu^2,*, Ruiyang Wang², Ahmed AlSanad³, Salman A. AlQahtani³, Xiaobing Chen⁴, Zhengtong Yin⁵, Xiaolu Li⁶, Wenfeng Zheng^2,3,*

1 Department of Geography and Anthropology, Louisiana State University, Baton Rouge, LA 70803, USA
2 School of Automation, University of Electronic Science and Technology of China, Chengdu, 610054, China
3 College of Computer and Information Sciences, King Saud University, Riyadh, 11574, Saudi Arabia
4 School of Electrical and Computer Engineering, Louisiana State University, Baton Rouge, LA 70803, USA
5 College of Resources and Environmental Engineering, Guizhou University, Guiyang, 550025, China
6 School of Geographical Sciences, Southwest University, Chongqing, 400715, China

* Corresponding Authors: Siyu Lu. Email: ; Wenfeng Zheng. Email:

(This article belongs to the Special Issue: Emerging Artificial Intelligence Technologies and Applications)

Computer Modeling in Engineering & Sciences 2024, 141(1), 771-786. https://doi.org/10.32604/cmes.2024.052622

Received 09 April 2024; Accepted 02 July 2024; Issue published 20 August 2024

Abstract

Recent advancements in natural language processing have given rise to numerous pre-training language models in question-answering systems. However, with the constant evolution of algorithms, data, and computing power, the increasing size and complexity of these models have led to increased training costs and reduced efficiency. This study aims to minimize the inference time of such models while maintaining computational performance. It also proposes a novel Distillation model for PAL-BERT (DPAL-BERT), specifically, employs knowledge distillation, using the PAL-BERT model as the teacher model to train two student models: DPAL-BERT-Bi and DPAL-BERT-C. This research enhances the dataset through techniques such as masking, replacement, and n-gram sampling to optimize knowledge transfer. The experimental results showed that the distilled models greatly outperform models trained from scratch. In addition, although the distilled models exhibit a slight decrease in performance compared to PAL-BERT, they significantly reduce inference time to just 0.25% of the original. This demonstrates the effectiveness of the proposed approach in balancing model performance and efficiency.

---

Keywords

---