Lai, Weixi (2024) Parameter-Efficient Fine-Tuning for Sarcasm Detection in Speech Using the Self-Supervised Pre-Trained Model WavLM. Master thesis, Voice Technology (VT).
|
PDF
MasterThesisWeixiLaifinally12-2.pdf Download (874kB) | Preview |
Abstract
As an integral part of human language and culture, sarcasm has naturally attracted great interest from researchers across various fields, including artificial intelligence. While much attention has been devoted to sarcasm detection in textual data, the realm of speech has remained relatively unexplored. Leveraging recent advancements in self-supervised learning (SSL) in speech processing, I aim to explore Low-rank Adaptation (LoRA), one of the parameter-efficient fine-tuning (PEFT) techniques, with the self-supervised pre-trained model WavLM. To my knowledge, this study represents a pioneering effort in utilizing PEFT and WavLM for this specific task. By leveraging recent advancements in WavLM, the effectiveness of LoRA is rigorously evaluated through extensive analysis and comparison with the traditional fine-tuning method and other PEFT approaches. The results demonstrate LoRA’s superiority in F1 score, recall, and precision metrics for sarcasm speech detection, while also highlighting its capability to significantly reduce parameter requirements. These findings provide valuable insights into the potential and challenges of employing LoRA with SSL in sarcasm speech detection, offering critical guidance for future research in advancing natural language understanding and enhancing human-computer interaction.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Gao, X. |
| Date Deposited: | 13 Aug 2024 08:54 |
| Last Modified: | 13 Aug 2024 08:54 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/549 |
Actions (login required)
 |
View Item |