Extractive Explanations for Interpretable Text Ranking

Neural document ranking models perform impressively well due to superior language understanding gained from pre-training tasks. However, their complexity and large number of parameters these (typically transformer-based) are often non-interpretable in that decisions can not be clearly attributed specific parts the input documents. In this article, we propose inherently interpretable by generating explanations as a by-product prediction decision. We introduce Select-And-Rank paradigm for ranking, where first output an explanation selected subset sentences document. Thereafter, solely use or selection make prediction, making first-class citizens process. Technically, treat sentence latent variable trained jointly with ranker final output. To end, end-to-end training technique utilizing reparameterizable sampling using Gumbel-max trick . conduct extensive experiments demonstrate our approach is competitive state-of-the-art methods. Our broadly applicable numerous tasks furthers goal building design Finally, present real-world applications benefit method.