Towards Understanding Sample Variance in Visually Grounded Language Generation: Evaluations and Observations
This work addresses the reliability of evaluation protocols in vision-and-language tasks, which is crucial for building robust benchmarks and models, though it is incremental in nature.
The paper investigates how sample variance in multi-reference datasets affects model performance in visually grounded language generation, finding that human-generated references vary significantly across tasks and that CIDEr exhibits larger variances than other metrics.
A major challenge in visually grounded language generation is to build robust benchmark datasets and models that can generalize well in real-world settings. To do this, it is critical to ensure that our evaluation protocols are correct, and benchmarks are reliable. In this work, we set forth to design a set of experiments to understand an important but often ignored problem in visually grounded language generation: given that humans have different utilities and visual attention, how will the sample variance in multi-reference datasets affect the models' performance? Empirically, we study several multi-reference datasets and corresponding vision-and-language tasks. We show that it is of paramount importance to report variance in experiments; that human-generated references could vary drastically in different datasets/tasks, revealing the nature of each task; that metric-wise, CIDEr has shown systematically larger variances than others. Our evaluations on reference-per-instance shed light on the design of reliable datasets in the future.