Interpretability

Turn the rationales of single examples into a global explanation, such as a knowledge base.

Study whether a mixture of experts model can address interlocking in selective rationalization.

Evaluate whether large language models can perform selective rationalization through prompting.

Extract rationales from structured representations of text, such as trees, under structural constraints.

memBERT gives a transformer a memory of plain-text expert explanations, so each prediction points to the rationale behind it while keeping classification performance.