A Survey on Visualization for Explainable Deep Learning in Natural Language Processing

PhD Qualifying Examination


Title: "A Survey on Visualization for Explainable Deep Learning in Natural
Language Processing"

by

Mr. Zhihua JIN


Abstract:

Natural language processing (NLP) enables computers to analyze and synthesize
natural language. Recently, deep learning models have become increasingly
popular in NLP. The deep learning models can learn powerful representations for
natural language and greatly improve performance in NLP. However, with the
models becoming large and complex, it is difficult for users to understand
their inner mechanisms and brings much trouble in diagnosing the models. To
handle these issues, visualization techniques have been applied to improve the
explainability of deep learning models for NLP tasks. Visualization can work as
a proxy between models and users, and help users understand, debug, and refine
deep learning models for NLP tasks.

In this survey, we extensively study existing research on using visualization
techniques to improve the explainability of deep-learning-based NLP models. We
first introduce the concepts and deep learning methods in NLP. Then, we
categorize existing work based on their goals, i.e., understanding NLP models,
debugging NLP models, and refining NLP models, and further summarize their
techniques and major pros and cons of the related work in each category.
Finally, we conclude the survey and discuss future research directions.


Date:                   Friday, 9 October 2020

Time:                   3:00pm - 5:00pm

Zoom meeting:
https://hkust.zoom.com.cn/j/92183909755?pwd=czB3VFVvMHJYdEVVaVpUNnpRSkxpdz09

Committee Members:      Prof. Huamin Qu (Supervisor)
                        Dr. Xiaojuan Ma (Chairperson)
                        Dr. Qifeng Chen
                        Dr. Yangqiu Song


**** ALL are Welcome ****