Personalization in Goal-Oriented Dialog

TL;DR

The paper develops a synthetic, goal-oriented dialog dataset expanded with user profiles and speech-style variations to study personalization in end-to-end models. It introduces a split memory architecture for Memory Networks and demonstrates that separating profile information from dialog history improves personalized reasoning on KB-driven tasks, though training remains challenging. It further shows that a single multi-profile model in a multi-task setting outperforms profile-specific models by leveraging shared patterns across profiles. The work provides a structured evaluation framework to analyze personalization components in dialog systems and points toward practical applications in restaurant reservations and customer support. Overall, the study advances understanding of how to integrate user attributes and speech style into goal-oriented dialog, highlighting architectures and learning strategies that support personalization.

Abstract

The main goal of modeling human conversation is to create agents which can interact with people in both open-ended and goal-oriented scenarios. End-to-end trained neural dialog systems are an important line of research for such generalized dialog models as they do not resort to any situation-specific handcrafting of rules. However, incorporating personalization into such systems is a largely unexplored topic as there are no existing corpora to facilitate such work. In this paper, we present a new dataset of goal-oriented dialogs which are influenced by speaker profiles attached to them. We analyze the shortcomings of an existing end-to-end dialog system based on Memory Networks and propose modifications to the architecture which enable personalization. We also investigate personalization in dialog as a multi-task learning problem, and show that a single model which shares features among various profiles outperforms separate models for each profile.

Figures (2)

• Figure 1: Personalized Restaurant Reservation System. The user (in green or yellow) conducts a dialog with the bot (in blue) to reserve a table at a restaurant. At each turn, a model has access to the user's profile attributes, the conversation history and the outputs from the API call (in light red) and must predict the next bot utterance or API call (in dark red). The horizontal lines between dialog groups signify the separate tasks that are described in the following sections. (Illustration adapted from Figure 1, BordesW16.)
• Figure 2: Split memory architecture for Memory Networks. Profile attributes and conversation history are modeled in two separate memories. The outputs from both memories are summed to get the final response.