Learning to Rank with Capsule Neural Networks

Anna Nesterenko, Anastasia Ianina

Abstract: Exploratory search scenarios are becoming common in recommender systems. Exploratory search is aimed at self-education and knowledge acquisition, its query often cannot be formulated in a short phrase but can be represented as one or several topically coherent documents. The results of exploratory search can be very extensive, but the ranking algorithms mostly use simple statistics which are not sufficient to rank documents in the convenient for further investigation and refinement order. In this work, we study capsule neural networks in combination with various textual representations applied to document ranking. The proposed model is tested on ArXiv triplets and Microsoft News Datasets and compared to classical approaches. Experiments show that capsule networks can significantly improve the quality of rankings in terms of search Presicion and Recall.