Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark

Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark Nouha Dziri∗†♦ Hannah Rashkin∗§ Tal Linzen§♣ David Reitter§ ♦University of Alberta, Canada §Google Research, USA ♣New York University, USA dziri@cs.ualberta.ca {hrashkin,linzen,reitter}@google.com Abstract Knowledge-grounded dialogue systems pow- ered by large language models often generate responses that, while fluent, are not attribut- able to a relevant source of information. Pro- gress towards models that do not exhibit this issue requires evaluation metrics that can quan- tify its prevalence. To this end, we introduce the Benchmark for Evaluation of Grounded INteraction (BEGIN), comprising 12k dialogue turns generated by neural dialogue systems trained on three knowledge-grounded dialogue corpora. We collect human annotations assess- ing the extent to which the models’ responses can be attributed to the given background in- formation. We then use BEGIN to analyze eight evaluation metrics. We find that these metrics rely on spurious correlations, do not reliably distinguish attributable abstractive responses from unattributable ones, and perform sub- stantially worse when the knowledge source is longer. Our findings underscore the need for more sophisticated and robust evaluation metrics for knowledge-grounded dialogue. We make BEGIN publicly available at https:// github.com/google/BEGIN-dataset. 1 Introduction Neural language models (Bengio et al., 2000; Vaswani et al., 2017; Radford et al., 2019, inter alia) often form the backbone of open-ended dia- logue systems (Wolf et al., 2019; Zhang et al., 2020b; Roller et al., 2021; Adiwardana et al., 2020). Utterances sampled from such language models sound natural, as reflected in these systems’ high scores in human evaluations fo- cused on measures such as ‘‘engagingness’’ or ‘‘human-likeness’’ (See et al., 2019). While fluent, however, the responses generated by these systems often contain statements that are not supported by ∗Equal contribution. †Work done while at Google Research. the evidence available to the system; such state- ments are sometimes referred to informally as ‘‘hallucinations’’ (Tian et al., 2019; Maynez et al., 2020a; Dziri et al., 2021; Shuster et al., 2021; see Figure 1 for an example). This issue is particularly salient for knowledge-grounded dialogue systems, which are expected to interact with a user in an open-ended fashion while conveying information that is attributable to external identifiable sources. In this work, we develop a benchmark that can be used to assess attribution in knowledge-based dia- log systems; following Rashkin et al. (2021a), we define an attributable response1 as one connected to textual evidence that supports the entirety of the response. A number of modeling approaches have recently been proposed to increase attribution in knowledge- grounded dialog systems (Rashkin et al., 2021b; Shuster et al., 2021; Dziri et al., 2021, 2022a). Progress in this area crucially relies on metrics that can measure the attribution of the text gen- erated by the system; and indeed, recent work has developed automated metrics with relatively high correlations with human annotations, poten- tially paving the way for alternatives to expen- sive human evaluations (Honovich et al., 2021; Dziri et al., 2021, 2022a). Yet our understand- ing of these recently proposed metrics, as well as more established ones, remains …

Content truncated to 500 words