Magooda, Ahmed
(2022)
Techniques to Enhance Abstractive Summarization Model Training for Low Resource Domains.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Nowadays, the amount of information is growing exponentially, and it is challenging to digest even the information for a particular topic. Summarization can reduce the information into a handful of paragraphs, helping human readers digest information easier. Automatic summarization spans different techniques (abstractive, extractive, phrase extractive, etc.). Abstractive summarization specially aims to mimic how humans summarize, as it aims to summarize a large amount of text into a readable, comprehensive summary. Abstractive summarization has benefited from recent advances in Machine learning and Natural Language Processing. However, the majority of prior studies focus on data-rich domains, where large datasets are available.
On the other hand, very few studies focus on data scarce domains. A typical practical issue that is rendered in such domains is model overfitting. Training complex models using a few samples can easily lead to overfitting. As a step towards remedying these shortcomings, this thesis aims to enhance abstractive summarization models in low-resource settings by tackling three challenges.
1-Can we adapt widely used data augmentation/synthesis techniques to abstractive summarization to remedy the scarceness issue?
2- How can we benefit from domain transfer or pretraining, and what can be a helpful strategy to do it more efficiently?
3- Can we extract additional information from the data and to use it more effectively?
This thesis first proposes new data synthesis (augmentation) models, novel techniques to synthesize new data for model training. We then introduced a variant of a recent data augmentation technique to be used in generative tasks. Additionally, we explored the utility of using curriculum learning to both improve pretraining and fine tuning processes. Finally, to overcome the third challenge, we propose integrating the summarization model into a multitask learning setting. We also show that some auxiliary tasks can consistently improve abstractive summarization in a low resource setting. We finally combine multitask learning and data augmentation to observe if the combination would be more helpful than each approach in isolation. We ultimately showed that combining more than one technique can introduce some improvements compared to a single technique. However, overall, using techniques in isolation leads to more consistent improvements.
Share
Citation/Export: |
|
Social Networking: |
|
Details
Item Type: |
University of Pittsburgh ETD
|
Status: |
Unpublished |
Creators/Authors: |
|
ETD Committee: |
|
Date: |
6 June 2022 |
Date Type: |
Publication |
Defense Date: |
4 March 2022 |
Approval Date: |
6 June 2022 |
Submission Date: |
4 April 2022 |
Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
Number of Pages: |
157 |
Institution: |
University of Pittsburgh |
Schools and Programs: |
Dietrich School of Arts and Sciences > Computer Science |
Degree: |
PhD - Doctor of Philosophy |
Thesis Type: |
Doctoral Dissertation |
Refereed: |
Yes |
Uncontrolled Keywords: |
NLP
ML
Text Summarization
Text Synthesis
Data Augmentation
Multitask Learning |
Date Deposited: |
06 Jun 2022 15:56 |
Last Modified: |
06 Jun 2022 15:56 |
URI: |
http://d-scholarship.pitt.edu/id/eprint/42259 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |