Beyond Topicality: Including Multidimensional Relevance in Cross-encoder Re-ranking

In this paper, we propose a novel approach to consider multiple dimensions of relevance in cross-encoder re-ranking. On the one hand, cross-encoders constitute an effective solution for re-ranking when considering a single relevance dimension such as topicality, but are not designed to straightforwardly account for additional relevance dimensions. On the other hand, the majority of re-ranking models accounting for multdimensional relevance are often based on the aggregation of multiple relevance scores at the re-ranking stage, leading to potential compensatory effects. To address these issues, in the proposed solution we enhance the candidate documents retrieved by a first-stage lexical retrieval model with suitable relevance statements related to distinct relevance dimensions, and then perform a re-ranking on them with cross-encoders. In this work we focus, in particular, on an extra dimension of relevance beyond topicality, namely, credibility, to address health misinformation in the Consumer Health Search task. Experimental evaluations are performed by considering publicly available datasets; our results show that the proposed approach statistically outperforms state-of-the-art aggregation-based and cross-encoder re-rankers.