FaithDial: A Faithful Benchmark for Information-Seeking Dialogue

FAITHDIAL: A Faithful Benchmark for Information-Seeking Dialogue Nouha Dziri† ♦ § Ehsan Kamalloo† Sivan Milton‡ Osmar Zaiane† § Mo Yu¶∗ Edoardo M. Ponti♣ Siva Reddy♦ ‡ †University of Alberta, Canada ♦Mila – Quebec AI Institute, Canada ‡McGill University, Canada ¶WeChat AI, Tencent, USA ♣University of Edinburgh, UK §Alberta Machine Intelligence Institute (Amii), Canada dziri@cs.ualberta.ca Abstract The goal of information-seeking dialogue is to respond to seeker queries with natural language utterances that are grounded on knowledge sources. However, dialogue sys- tems often produce unsupported utterances, a phenomenon known as hallucination. To mit- igate this behavior, we adopt a data-centric solution and create FAITHDIAL, a new bench- mark for hallucination-free dialogues, by editing hallucinated responses in the Wizard of Wikipedia (WOW) benchmark. We observe that FAITHDIAL is more faithful than WoW while also maintaining engaging conversa- tions. We show that FAITHDIAL can serve as training signal for: i) a hallucination critic, which discriminates whether an utterance is faithful or not, and boosts the performance by 12.8 F1 score on the BEGIN benchmark compared to existing datasets for dialogue coherence; ii) high-quality dialogue genera- tion. We benchmark a series of state-of-the-art models and propose an auxiliary contrastive objective that achieves the highest level of faithfulness and abstractiveness based on sev- eral automated metrics. Further, we find that the benefits of FAITHDIAL generalize to zero-shot transfer on other datasets, such as CMU-Dog and TopicalChat. Finally, human evaluation reveals that responses generated by models trained on FAITHDIAL are perceived as more interpretable, cooperative, and engaging. 1 Introduction Despite the recent success of knowledge-grounded neural conversational models (Thoppilan et al., 2022; Prabhumoye et al., 2021; Zhao et al., 2020, inter alia) in generating fluent responses, they also generate unverifiable or factually incorrect statements, a phenomenon known as hallucination (Rashkin et al., 2021b; Dziri et al., 2021; Shuster ∗Work done while at IBM Research. et al., 2021). Ensuring that models are trustworthy is key to deploying them safely in real-world applications, especially in high-stakes domains. In fact, they can unintentionally inflict harm on members of the society with unfounded statements or can be exploited by malicious groups to spread large-scale disinformation. Recently, Dziri et al. (2022a) investigated the underlying roots of this phenomenon and found that the gold-standard conversational datasets (Dinan et al., 2019; Gopalakrishnan et al., 2019; Zhou et al., 2018)—upon which the models are commonly fine-tuned—are rife with hallucina- tions, in more than 60% of the turns. An example of hallucination in Wizard of Wikipedia (WoW; Dinan et al. 2019) is shown in the red box of Figure 1. In WoW, an information SEEKER aims to learn about a topic and a human WIZARD harnesses knowledge (typically a sentence) from Wikipedia to answer. This behavior, where the human WIZ- ARD ignores the knowledge snippet and assumes a fictitious persona, can later reverberate in the dia- logue system trained on this kind of data. Instead, the ideal WIZARD response, highlighted in green, should acknowledge the bot’s nature, and when- ever the knowledge is not sufficient or …

Content truncated to 500 words