Resource Constrained Dialog Policy Learning Via Differentiable Inductive Logic Programming

Abstract

Motivated by the needs of resource constrained dialog policy learning, we introduce dialog policy via differentiable inductive logic (DILOG). We explore the tasks of one-shot learning and zero-shot domain transfer with DILOG on SimDial and MultiWoZ. Using a single representative dialog from the restaurant domain, we train DILOG on the SimDial dataset and obtain 99+% in-domain test accuracy. We also show that the trained DILOG zero-shot transfers to all other domains with 99+% accuracy, proving the suitability of DILOG to slot-filling dialogs. We further extend our study to the MultiWoZ dataset achieving 90+% inform and success metrics. We also observe that these metrics are not capturing some of the shortcomings of DILOG in terms of false positives, prompting us to measure an auxiliary Action F1 score. We show that DILOG is 100x more data efficient than state-of-the-art neural approaches on MultiWoZ while achieving similar performance metrics. We conclude with a discussion on the strengths and weaknesses of DILOG.

Efficient Dialog Policy Learning by Reasoning with Contextual Knowledge

“…Traditionally, ASP reasons about discrete-valued symbols and does not quantify uncertainty explicitly. But in some works such as P-log (Baral, Gelfond, and Rushton 2009), RIL (Zhang et al 2019), NeurASP (Yang, Ishay, and Lee 2020) and DILOG (Zhou et al 2020), the paradigm is extended to allow probability atoms, and integrated with subsymbolic approaches. In this paper, we just use ASP without probabilities, to characterize and reason about some contextual factors that are not modeled in dialog states.…”

Every time I fire a conversational designer, the performance of the dialog system goes down

“…This approach is the more general line of research of merging symbolic knowledge and neural networks. In the context of neural-based dialogue systems, this line is pursued by using constrained rules (Jhunjhunwala et al, 2020), logical rules to be used in inductive logic programming (Zhou et al, 2020) or declarative language (Altszyler et al, 2020). These rules and models can be easily included in the existing dialogue state tracking models to guide the training and predictions phases without additional learning parameters (Hu et al, 2016;van Krieken et al, 2020).…”

“…As far as we aim to inject knowledge in the dialogue system, we need to clarify how dialogue states are represented and how rules can be described. For this reasons, we express both dialogue states and rules in a logical form (as in (Zhou et al, 2020)). By using this logical form, rules can be expressed by using logical constraints and variables.…”