Parameter-efficient Fine-tuning in Reading Comprehension
Name
Rustam Abdumalikov
Abstract
Question Answering is an important task in Natural Language Processing. There are different approaches to answering questions, such as using the knowledge learned during pre-training or extracting an answer from a given context, which is commonly known as reading comprehension. One problem with the knowledge learned during pre-trained is that it can become outdated because we train it only once. Instead of replacing outdated information in the model, an alternative approach is to add updated information to the model input. However, there is a risk that the model may rely too much on its memorized knowledge and ignore new information, which can cause errors. Our study aims to analyze whether parameter-efficient fine-tuning methods would improve the model’s ability to handle new information. We assess the effectiveness of these techniques in comparison to traditional fine-tuning for reading comprehension on an augmented NaturalQuestions dataset. Our findings indicate that parameter-efficient fine-tuning leads to a marginal improvement in performance compared to fine-tuning. Furthermore, we observed that data augmentations contributed the most substantial performance enhancements.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Yova Kementchedjhieva, Kairit Sirts
Defence year
2023