Yi Zhao
- Harbin
I’m a graduate student in Research Center for Social Computing and Information Retrieval(SCIR), at Harbin Institute of Technology (HIT, China). I am advised by Prof. Ting Liu. My research interests include cognition-inspired natural language processing and brain-computer interface.
Publications
Text Difficulty Study: Do machines behave the same as humans regarding text difficulty? Machine Intelligence Research
Bowen Chen, Xiao Ding, Yi Zhao, Bo Fu, Tingmao Lin, Bing Qin and Ting Liu
We explored text complexity's impact on model training. Using different models, tasks, and complexity measures, we divided the training set into subsets and studied their effect on accuracy. Our result shows that (1) LSTM is more sensitive to complexity than BERT; (2) UID is a promising text complexity measure; (3) Training on low-complexity data generalizes well to high-complexity tests, but not vice versa; (4) Training from easy to hard reduces steps by 30%-50% without accuracy loss, compressing training time effectively.