Investigating Data Variance in Evaluations of Automatic Machine Translation Metrics

TL;DR

This paper demonstrates that the evaluation of automatic machine translation metrics is highly sensitive to dataset choice, highlighting the importance of considering data variance for reliable assessment.

Contribution

It provides a comprehensive analysis of data variance effects on metric evaluation and investigates hypotheses explaining this variability, emphasizing cautious interpretation of single-dataset results.

Findings

Metric performance varies across datasets.

Data points and i.i.d assumptions influence evaluation stability.

Single dataset evaluations may lead to inconsistent conclusions.

Abstract

Current practices in metric evaluation focus on one single dataset, e.g., Newstest dataset in each year's WMT Metrics Shared Task. However, in this paper, we qualitatively and quantitatively show that the performances of metrics are sensitive to data. The ranking of metrics varies when the evaluation is conducted on different datasets. 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, because it may leads to inconsistent results with most of other datasets.