Investigating Evaluation of Open-Domain Dialogue Systems With Human Generated Multiple References

TL;DR

Open-domain dialogue evaluation suffers from a one-to-many problem when using single references. The authors introduce multi-reference evaluation by expanding the DailyDialog test set with four human-generated references per context and demonstrate that this approach yields stronger correlations between automatic metrics and human judgments for both quality and diversity. They provide a detailed analysis of how many references are needed, and show that referenced diversity metrics align better with human judgments than unreferenced metrics. The work suggests that multi-reference test sets enable more reliable automatic evaluation of open-domain dialogue systems and can be extended to other datasets.

Abstract

The aim of this paper is to mitigate the shortcomings of automatic evaluation of open-domain dialog systems through multi-reference evaluation. Existing metrics have been shown to correlate poorly with human judgement, particularly in open-domain dialog. One alternative is to collect human annotations for evaluation, which can be expensive and time consuming. To demonstrate the effectiveness of multi-reference evaluation, we augment the test set of DailyDialog with multiple references. A series of experiments show that the use of multiple references results in improved correlation between several automatic metrics and human judgement for both the quality and the diversity of system output.

Figures (3)

• Figure 1: System level correlations for BLEU-2 and METEOR metrics. Multi-reference evaluation shows higher correlation with more clear differentiation in model performance.
• Figure 2: Change in correlation with varying number of references. Trend stablizes after 4-5 references
• Figure 3: Interface used for multi-reference data collection.