User Memory Reasoning for Conversational Recommendation

Proceedings of the 28th International Conference on Computational Linguistics, pages 5288–5308 Barcelona, Spain (Online), December 8-13, 2020 5288 User Memory Reasoning for Conversational Recommendation Hu Xu1,2∗, Seungwhan Moon1, Honglei Liu1, Bing Liu1, Pararth Shah1 Bing Liu2 and Philip S. Yu2 1Facebook 2Department of Computer Science, University of Illinois at Chicago {huxu,shanemoon,honglei,bingl,pararths}@fb.com {liub,psyu}@uic.edu Abstract We study an end-to-end approach for conversational recommendation that dynamically manages and reasons over users’ past (offline) preferences and current (online) requests through a struc- tured and cumulative user memory knowledge graph. This formulation extends existing state tracking beyond the boundary of a single dialog to user state tracking (UST). For this study, we create a new Memory Graph (MG) ↔ Conversational Recommendation parallel corpus called MGConvRex with 7K+ human-to-human role-playing dialogs, grounded on a large-scale user memory bootstrapped from real-world user scenarios. MGConvRex captures human-level rea- soning over user memory and has disjoint training/testing sets of users for zero-shot (cold-start) reasoning for recommendation. We propose a simple yet expandable formulation for construct- ing and updating the MG, and an end-to-end graph-based reasoning model that updates MG from unstructured utterances and predicts optimal dialog policies (e.g. recommendation) based on updated MG. The prediction of our proposed model inherits the graph structure, providing a natural way to explain policies. Experiments are conducted for both offline metrics and online simulation, showing competitive results.1 1 Introduction Conversational recommendation system has recently gained traction in the dialog community, in which the model aims to learn up-to-date (online) user preferences, instead of using static (offline) preferences as in the traditional, e.g. collaborative filtering (CF). Most existing works focus on combining a static recommender system with a dialog system, by updating user preferences via asking relevant questions (often referred to as “System Ask User Respond (SAUR)” (Zhang et al., 2018). However, this “short- term” update in the model unnaturally isolates users’ history and their preference in the current dialog (often forgotten after the dialog is finished). An intelligent system should be able to dynamically maintain and reason over users’ knowledge for current (and possibly future) recommendations. To this end, we introduce a novel concept called user memory graph to holistically represent the knowledge about users and associated items. This user memory graph extends the notion of dialog state tracking (DST) beyond the boundary of a single dialog to user state tracking (UST). As in Figure 1, such a graph contains static knowledge obtained offline (e.g. items, attributes, the history of users and past dialogs) and users’ knowledge online (e.g. from the current and possibly future dialogs). User memory graph naturally has the following benefits. (1) Holistic reasoning considers available knowledge about users and items all together (not just dialog states) to generate policies. We focus on two types of reasoning policies: graph update policy and dialog policy, where the former updates users’ online knowledge and serves as the basis for the later to generate outputs of the agent’s current turn. We believe reasoning is the core problem in conversational recommendation because asking a …

Content truncated to 500 words