Liejun Wang

Linear scalarization in multi-task radiology report generation is fundamentally flawed due to a "Double Dilemma" in gradient dynamics, but a new optimizer can fix it.

Emotion recognition can be significantly improved by adapting to individual expressive traits, with ML-SAN outperforming static models in capturing nuanced emotional expressions.

Unlock cross-lingual speech emotion recognition for low-resource languages with a new method that needs only 5 labeled examples.