Adding Chit-Chat to Enhance Task-Oriented Dialogues

In this work, we propose to integrate both types of systems by Adding Chit-Chat to ENhance Task-ORiented dialogues (ACCENTOR), with the goal of making virtual assistant conversations more engaging and interactive. Specifically, we propose a Human $ AI collaborative data collection approach for generating diverse chitchat responses to augment task-oriented dialogues with minimal annotation effort.

We employ ACUTE-Eval (Li et al., 2019) to compare the augmented versions with the originals along four axes: engagingness, interestingness, knowledge, and humanness. We find that the augmented dialogues are consistently preferred by human judges across the four axes for both datasets (Section 4.1).

In addition, we propose and evaluate three models for adding chit-chat to task-oriented dialogues, including an end-to-end model and two code-switcher models built upon off-the-shelf task oriented and chit-chat systems (Section 3). Compared with the baseline model trained with the original unaugmented data, our models trained with the augmented version can generate significantly higher-rated responses in terms of human preference while maintaining competitive task performance in goal tracking