Task-oriented dialogue systems are essential for enabling natural, efficient interactions between users and machines in practical applications like booking services, customer support, etc. With recent advancements in language models, dialogue systems have started to gain more attention both in academia and industries. They can be broadly categorized into two — open-domain and task-oriented dialogue systems. Unlike open-domain dialogue systems that focus more on making open-ended conversations, task-oriented ones are trained to accomplish goals given by the user.
In this talk, we will begin by discussing recent progress in the field and outlining the two main types of dialogue systems. We will broadly discuss task-oriented dialogue systems, with our primary focus on natural language generation systems. We will go through an overview of the traditional statistical and neural network-based dialogue systems. We will briefly discuss the various modules and the recent shift to end-to-end architectures. We will also discuss the existing evaluation methods and challenges.
Despite significant advancements, current task-oriented systems often produce responses that are dull and monotonous. While they tend to meet user goals, their “naturalness” is still a point of concern, as responses are often unaligned to the user inputs. In the latter part of the talk, we will present our work recently accepted at NAACL Findings 2024, which tackles these challenges. We demonstrate that our methods generate responses that are more “natural” and aligned with the user in a task-oriented setting without hurting the task success.
Additional resources: