Yikang Liu

About me

Hi, I'm Yikang, a second-year graduate student of computational linguistics at Shanghai Jiao Tong University (SJTU). I am currently advised by Hai Hu. Before that, I received my B.A. in English and Minor B.Eng. in CS from Huazhong University of Science and Technology (HUST).

For now, I am trying to understand the behavior of language models by analyzing how they process human-understandable linguistic features.

My story with language started with The Lord of the Rings, which I was fascinated by as a child. Then I came to know that J.R.R Tolkien was a philologist, which inspired me to learn language. During my first year as an undergraduate, I realized that language can go beyond literature and philology. One can look into language by building computational models, from perspectives of science and engineering. Thus I started a minor program in computer science, and later I decided to pursue a master's degree in computational linguistics.

For my daily hobbies, I like to play the violin and listen to classical music, especially symphonies and violin concerto. I was once a violin player in the student ochestra of HUST, performing a New Year Concert in the magnificant Qintai Concert Hall in Wuhan!

Experience

Tongyi Lab, Alibaba Cloud Hangzhou, China
Research Intern 2024.07 - Now
Shanghai Jiao Tong University Shanghai, China
M.A. in Computational Linguistics 2023.09 - Now
Huazhong University of Science and Technology Wuhan, China
B.A. in English Language and Literature 2019.09 - 2023.06
B.Eng. in Computer Science (Minor) 2021.02 - 2023.06

Publications

ZhoBLiMP: a Systematic Assessment of Language Models with Linguistic Minimal Pairs in Chinese
Yikang Liu, Yetin Shen, Hongao Zhu, Lilong Xu, Zhiheng Qian, Siyuan Song, Kejia Zhang, Jialong Tang, Pei Zhang, Baosong Yang, Rui Wang, Hai Hu
[preprint, 2024] [paper] [code]

This is a minimal pair paradigm benchmark for Chinese, inspired by BLiMP. ZhoBLiMP includes 118 linguistic paradigms, covering 15 linguistic phenomena, with 35k minimal pairs in total. With ZhoBLiMP, we kind of investigate the "scaling law" of syntactic acquisition of language models. We also find a U-shaped learning curve of LMs on grammar acquisition, which we cannot interpret yet.

MELA: multilingual evaluation of linguistic acceptability
Ziyin Zhang*, Yikang Liu*, Weifang Huang, Junyu Mao, Rui Wang, Hai Hu (*Equal contribution)
[ACL 2024] [paper] [data]

This is my very first paper submitted to and accepted by ACL. We proposed a multilingual benchmark on linguistic acceptability, covering 10 languages. I probed XLM-R models fine-tuned on MELA and ran LLM evaluations. It is a good start of my academic career!

Can local ecological knowledge establish conservation baselines for the Critically Endangered Blue-crowned Laughingthrush?
Rosalind A. Gleave, Sarah K. Papworth, David Bauman, Steven J. Portugal, Weiwei Zhang, Yikang Liu, Zhiming Cao, Xiaojin Cheng, Samuel T. Turvey
[People and Nature, 2024] [paper]

I took part in this project at the age of 17, doing field work in Wuyuan, Jiangxi, China, collecting ecological data, something like doing questionnaires with local farmers and measuring the height and width of trees. By the way, my father is a bird keeper of BCLT in the zoo...

ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models
Yikang Liu, Ziyin Zhang, Wanyang Zhang, Shisen Yue, Xiaojing Zhao, Xinyuan Cheng, Yiwen Zhang, Hai Hu
[preprint, 2023] [paper] [data] [demo]

This was a timely work beginning just two months after the release of ChatGPT. We collected argumentative essays generated by GPT models, evaluated these essays from various aspects, and trained classifiers to detect GPT-generated essays. This paper earned me an Honor Undergraduate Thesis!