Decoupled Rationalization with Asymmetric Learning Rates: A Flexible Lipschitz Restraint
Published in KDD, 2023
- **Task:** Self-explaining rationalization in NLP - **Problem:** Degeneration. That's to say, in a cooperative game, the predictor and the generator (i.e., rationalizer) may collude to use uninformative rationale candidates to get the right label. - **Insights:** We first link the degeneration problem to the predictor's Lipschitz continuity. And then we link the predictor's Lipschitz continuity to the coordination of the predictor and the generator. - **Solution:** Slow down the predictor, which is very simple and is compatible with many variants of this kind of two-player rationaliser/classifier games.