Towards Controllable Explanation Generation for Recommender Systems via Neural Template

Abstract

It has been commonly agreed that the explanation associated with recommendation can be effective in increasing the recommender systems (RS)’s transparency and thus users’ satisfaction and acceptance. Among the various types of explanation in RS, the commonly used textual explanation can be roughly classified into two categories, i.e., template-based and generation-based. As for the former, the fixed template may lose flexibility, while, though the latter may enrich the explanation, it may produce less useful content due to the lack of controllability. In this work, we combine the advantages of the two types of method by developing a neural generation approach named Neural Template (NETE) whose explanations are not only flexible but also controllable and useful. Our human evaluation results confirm that the explanations from our model are perceived helpful by users. Furthermore, our case study illustrates that the explanation generation process is controllable. To demonstrate the controllability of our model, we present a demo that can be easily viewed on a Web browser.