Abstract:
"
What if a Machine Learning (ML) model could have training data generated based on
its classification uncertainty and query a human-in-the-loop for their label? In doing so,
would we able to construct a better classifier while reducing data acquisition costs for
practical, real-world Text Classification applications? This is the premise of Active
Learning based on the Membership Query Synthesis scenario, which is coined as Query
Synthesis Active Learning in this research. It is an emerging area of research in Active
Learning where generative methods and augmentation methods have been proposed for
generating new textual training data (Query Synthesis). However, there are several gaps
that exist in previous work such as support for Deep Learning architectures in NLP,
lack of interpretability in estimation of model uncertainty and query synthesis, and the
need to train and finetune a separate model for query synthesis. Therefore, this research
aims to design, develop, and test a novel Query Synthesis Active Learning method to
address these gaps. To demonstrate the real world, practical application of the Query
Synthesis Active Learning method, an interactive data labelling GUI prototype is
developed. In addition, another contribution is the source code is designed and
developed to be reusable as a Python package that NLP researchers and practitioners
can use out-of-the-box for Query Synthesis Active Learning."
