Tianze (Etha) Hua

Research Interests

I am interested in natural language understanding and its relation with the symbol grounding problem. I believe for us to attribute cognitive properties (m knows p; m understands English; m is conscious; etc.) to AI systems, two components are required: (a) operationalzied definitions of those cognitive properties, and (b) sufficient knowledge of the inner workings of the AI systems of interest. While the first component has been the subject of lengthy debates in the realms of philosophy of mind and epistemology and may continue to be unresolved, the second component offers a domain (interpretability) where concrete answers are achievable.

To get a better understanding of the internal mechanisms in LLMs, I am currently investigating the shared representations across similar tasks in transformer models.

Besides interpretability, I am interested in experimenting and analyzing language understanding through multi-agent interactions.

Updates

• March 2024: Our work "mOthello: When Do Cross-Lingual Representation Alignment and Cross-Lingual Transfer Emerge in Multilingual Models?" got accepted into Findings of NAACL 2024!
• August 2023: I attended the Cornell, Maryland, Max Planck Pre-doctoral Research School 2023 in Saarbrücken, Germany.

Publications

• mOthello: When Do Cross-Lingual Representation Alignment and Cross-Lingual Transfer Emerge in Multilingual Models?
  Tianze Hua*, Tian Yun*, and Ellie Pavlick
  Findings of NAACL 2024 [Code] [Project Page]

Education

2023-now

Sc.M. Computer Science, Brown University

2019-2023

B.S. Computer Science & Philosophy, Tufts University