NSSDM 2023

Many task-oriented dialogue systems use deep reinforcement learning (DRL) to learn policies that respond to the user appropriately and complete the tasks successfully. Training DRL agents with diverse dialogue trajectories prepare them well for rare user requests and unseen situations. One effective diversification method is to let the agent interact with a diverse set of learned user models. However, trajectories created by these artificial user models may contain generation errors, which can quickly propagate into the agent’s policy. It is thus important to control the quality of the diversification and resist the noise. In this paper, we propose a novel dialogue diversification method for task-oriented dialogue systems trained in simulators. Our method, Intermittent Short Extension Ensemble (I-SEE), constrains the intensity to interact with an ensemble of diverse user models and effectively controls the quality of the diversification. Evaluations show that I-SEE successfully boosts the performance of several DRL dialogue agents.

Bias in data as well as lack of transparency and fairness in algorithms are not new problems, but with the increasing scale, complexity, and adoption, most AI systems are suffering from these issues at a level unprecedented. Information access systems are not spared since these days, almost all large-scale systems of information access are mediated by algorithms. These algorithms are optimized not only for relevance, which is subjective to begin with, but also for measures of engagement and impressions. They are picking up signals of what may be 'good' from individuals and perpetuating that through learning methods that are opaque and hard to debug. Considering 'fairness' and introducing more transparency can help, but it can also backfire or create other issues. We also need to understand how and why users of these systems engage with content. In this talk, I will share some of our attempts for bringing fairness in ranking systems and then talk about how the solutions are not that simple.

The candidate matching, i.e., retrieving potentially relevant candidate items to users for further ranking, is highly important since even the cleverest housewife cannot bake bread without flour. While many research efforts are devoted to the ranking stage, few works focus on the former one where the space for more sophisticated models seems to be limited due to the requirement of low computation cost, especially for an industrial system. No matter which scenario, the matching stage is highly important since even the cleverest housewife cannot bake bread without flour. In this talk, we will introduce the background knowledge towards this task at first. Then, an overview on the recent efforts of this line will be presented. At last, we will discuss the possible directions in the near future.

Graph Neural Networks (GNNs) are powerful models to exploit the high-order relationship between entities on graphs. Despite the superior performance, we have little knowledge about the explainability of GNNs. In this talk, we will introduce two themes of explainability, (1) Post-hoc explainability: Using an additional explainer method to explain a black-box model post hoc, but explanations could be unfaithful to the decision-making process of a model; (2) Intrinsic Interpretability: Incorporating a rationalization module into the model design, so as to transform a black-box to a white-box. We find causal theory is one promising solution and we will discuss interpretability and generalization.