🧑‍🔬 Wee Sun Lee

Wee Sun Lee works on machine learning for natural language. The papers here follow two lines. One is aspect-based sentiment analysis, where a system must find the aspects a review mentions and judge the sentiment expressed toward each. The other is the reinforcement-learning training of large language models for mathematical reasoning. Both studies take an established pipeline apart. They ask how its parts interact and which of them actually account for the observed results.

The most cited of these papers, An Interactive Multi-Task Learning Network for End-to-End Aspect-Based Sentiment Analysis (2019), moves away from handling aspect extraction and sentiment classification as consecutive steps, an arrangement in which mistakes propagate and the link between the two subtasks is ignored. It instead builds one network of CNN encoders whose shared latent representation feeds several task-specific components, so that the subtasks can inform each other. The newer direction appears in Understanding R1-Zero-Like Training: A Critical Perspective (2025). This paper examines the idea of applying reinforcement learning straight to a base model without supervised fine-tuning, and it separates the contribution of the base model from that of the RL algorithm. Testing several base models on MATH training questions, it looks at the effect of prompt templates, at exploration and at self-reflection. It reports that self-reflection is already present before any RL. It also finds that Qwen2.5 models perform best with no template, which hints that question–answer text was part of their pretraining.

Lee is based at the National University of Singapore, with bylines there from 2002 onward. Earlier papers list the Australian National University, the Australian Defence Force Academy, UNSW Canberra and UNSW Sydney. Co-authors on the papers here include Ruidan He, Hwee Tou Ng, Daniel Dahlmeier and Min Lin.

Wee Sun Lee has 2 papers in our library, in 2019–2025, most of them on Sentiment & emotion; often with Daniel Dahlmeier, Hwee Tou Ng, Min Lin.