Friday, September 23, 2022 - 11:30 to 13:00
SCHOOL OF COMPUTER SCIENCE
The School of Computer Science is pleased to present…
MSc Thesis Proposal by: Vishakha Gautam
Date: September 23rd, 2022
Time: 11:30 AM – 1:00 PM
Meeting URL: https://us06web.zoom.us/j/82360865407?from=addon
Passcode: If interested in attending this event, contact the Graduate Secretary at firstname.lastname@example.org with sufficient notice before the event to obtain the passcode.
Being an NLP pre-training model built with transformer and attention mechanism, BERT has proven to be highly efficient in developing various applications like text classification, machine translation, question-answering systems etc. Despite the recent advancement, however, the robustness of the predicted models remains a challenging issue. Considering BERT-based question-answering systems, we propose in this thesis, two data augmentation techniques to construct adversarial examples to strengthen the robustness. The first technique follows the generative approach to label-preserving paraphrasing of a given question. The paraphrased question is obtained from a BERT-based generative model where a chosen keyword in the question is replaced by a similar word in its associated context. We propose to define this similarity by the word embedding obtained from word2vec. The second technique follows the rule-based approach to construct an unanswerable context to a given question. The context is selected from existing unanswerable passages according to its relevance to the question, defined by the highest ranking of this context paired with a keyword in the question. We propose to use TF-IDF to define the relevance score of a keyword-context pair. To compare with the related work in the literature, the proposed method will be exercised on the benchmark Stanford Question Answering Dataset (SQuAD) and evaluated with the exact match and F1 score.
Keywords - BERT, Paraphrasing, Data augmentation, QA system, Robustness, Adversarial Example
MSc Thesis Committee:
Internal Reader: Dr Jianguo Lu
External Reader: Dr Ning Zhang
Advisor: Dr Jessica Chen
MSc Thesis Proposal Announcement
5113 Lambton Tower 401 Sunset Ave. Windsor ON, N9B 3P4 (519) 253-3000 Ext. 3716 email@example.com