On the Importance of Subtext in Recommender Systems

Abstract

Conversational recommender systems have been shown capable of allowing users to navigate even complex and unknown application domains effectively. However, optimizing preference elicitation remains a largely unsolved problem. In this paper we introduce SPEECHREC, a speech-enabled, knowledge-based recommender system, that engages the user in a natural-language dialog, identifying not only purely factual constraints from the users’ input, but also integrating nuanced lexical qualifiers and paralinguistic information into the recommendation strategy. In order to assess the viability of this concept, we present the results of an empirical study where we compare SPEECHREC to a traditional knowledge-based recommender system and show how incorporating more granular user preferences in the recommendation strategy can increase recommendation quality, while reducing median session length by 46 %.