A Multi-Task Learning Model for Multidimensional Relevance Assessment

In recent years, deep learning models have been successfully applied to Information Retrieval (IR), mainly for assessing the topical relevance of documents with respect to queries. However, relevance is a multidimensional concept, which can be assessed based on several criteria, depending on the document type, the domain considered, the search task performed, etc. Given that recent advancements in deep neural networks enable several learning tasks to be solved simultaneously, in this paper we examine the possibility of modeling multidimensional relevance by jointly solving a retrieval task, to learn topical relevance, and a classification task, to learn additional relevance dimensions. To instantiate and evaluate the proposed model, we consider three query-independent relevance dimensions beyond topicality, i.e., readability, trustworthiness, and credibility. The reported findings show that the proposed joint modeling can improve the performance of the retrieval task.