- 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: If the whole model achieves high prediction accuracy, the generator can always learn the true semantic.
- Solution: Sharing the encoders between the generator and the predictor, which is very simple and is compatible with many variants of this kind of two-player rationaliser/classifier games.
- 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.
MGR: Multi-generator Based Rationalization
- Task: Causal discovery from a perspective of causal feature selection
- Problem: Previous causal discovery methods usually involve many assumptions for causal inference (e.g., the conditional independence assumed by a graphical model, and the ignorability of a SCM).
- Insights: We provide some insights into how to select causal features from a purely probabilistic perspective, which doesn’t involve the above assumptions.
