Investigating Data Variance in Evaluations of Automatic Machine Translation Metrics

Current practices in metric evaluation mainly focus on one single dataset from a targeted domain, e.g., News domain in each year's WMT Metrics Shared Task. In this paper, we qualitatively and quantitatively show that the performances of metrics are sensitive to data even when the data is from the same domain, i.e., the ranking of metrics varies when the evaluation is conducted on different datasets from the same domain. Then this paper further investigates two potential hypotheses, i.e., insignificant data points and the deviation of Independent and Identically Distributed (i.i.d) assumption, which may take responsibility for the issue of data variance. In conclusion, our findings suggest that when evaluating automatic translation metrics, researchers should take data variance into account and be cautious to claim the result on a single dataset even from the same domain, because it may leads to inconsistent results with most of other datasets.