Improving Lifelong Learning Models for Classification in Natural Language Processing

Please enable javascript in your browser.

Sadraei Javaheri, Mohammad Ali | 2024

0 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 57266 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Rabiee, Hamid Reza; Asgari, Ehsaneddin
Abstract:
Fine-tuning language models has become a prevalent approach to tackling various natural language processing tasks. However, fine-tuning a model sequentially on a series of tasks leads to catastrophic forgetting: as the model learns a new task, it tends to forget the previously learned tasks. This research aims to address this issue by proposing a soft prompt tuning approach, a parameter-efficient fine-tuning method introduced in recent years. We introduce an improved version called SuperPositon Prompts, which is evaluated on the T5 model using tasks from the GLUE and SuperGLUE benchmarks. The results demonstrate the superior performance of the proposed method compared to previous approaches. Additionally, we empirically show that the removal of dropout layer improves the stability and performance of soft prompt tuning methods, particularly for "SuperPositon Prompts". Finally, we design a framework that leverages "SuperPositon Prompts" to enhance lifelong learning. We explore the possibility of knowledge transfer in this tuning method and conclude that forward knowledge transfer is feasible, making it a suitable approach for lifelong learning
Keywords:
Lifelong Learning ; Soft Prompt Tuning ; Parameter Efficient Fine-Tuning ; Natural Language Processing ; Language Model