Average Is Not Enough: Caveats of Multilingual Evaluation
This addresses methodological flaws in evaluating multilingual AI systems, which is crucial for researchers and practitioners to avoid skewed assessments, though it is incremental in proposing a visualization-based detection approach.
The paper tackles the problem of linguistic bias in multilingual evaluation, showing that using average performance metrics favors dominant language families and demonstrating through a case study that visualization with the URIEL database can detect such biases.
This position paper discusses the problem of multilingual evaluation. Using simple statistics, such as average language performance, might inject linguistic biases in favor of dominant language families into evaluation methodology. We argue that a qualitative analysis informed by comparative linguistics is needed for multilingual results to detect this kind of bias. We show in our case study that results in published works can indeed be linguistically biased and we demonstrate that visualization based on URIEL typological database can detect it.