KEWEI LIU
Research Profile | The University of Tokyo · EEIS
Kewei Liu • 劉 可惟 • 刘可惟 • Tokyo, Japan
CURRENT RESEARCH
At the Minematsu–Saito Laboratory, I study how visual scene information, speech and language processing, and staged hints can help Japanese learners begin speaking in their own words. The current design examines support in terms of speaking-onset latency, amount of speech, content specificity, naturalness, and user burden.
UNDERGRADUATE RESEARCH
I formulated video summarization in which the user specifies a time-varying emotional trajectory in valence–arousal space together with a summary ratio. The system estimates segment-level visual, acoustic, and textual emotion cues, fuses them with path-conditioned reliability-aware directed cross-attention and routing, and selects temporally ordered segments with dynamic programming under a cost that balances target-path proximity and content importance.
On TVSum, the module-level multimodal fuser improved internal VA-path coherence from CCCmean 0.427 ± 0.038 for confidence-based fusion to 0.531 ± 0.025 for the reliability-aware attention-and-router setting under 5-fold cross-validation. The paper is accepted at APSIPA ASC 2026.