Extractive Explanations for Interpretable Text Ranking

Abstract

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

Explaining Search Result Stances to Opinionated People

“…Previous XAI literature has contributed to explaining information retrieval systems, focusing on the interpretability of document-retrieval mechanisms [25,26,52]. For example, the authors of [52] propose a listwise explanation generator, which provides an explanation that covers all the documents contained in the page (e.g., by describing which query aspects were covered by each document).…”

Supervised Contrastive Learning Approach for Contextual Ranking

“…Note that a common problem in both approaches is due to an upper bound on the acceptable input length of contextual models that restricts its applicability to shorter documents. When documents do not fit into the model the documents are chunked into passages/sentences to fit within token limit either by using transformer-kernels [18,19], truncation [7], or careful pre-selection of relevant text [26,53].…”