Towards Understanding Sample Variance in Visually Grounded Language Generation: Evaluations and Observations

TL;DR

This paper investigates how sample variance in multi-reference datasets impacts the evaluation of visually grounded language generation models, emphasizing the importance of reporting variance and analyzing dataset reliability.

Contribution

It introduces experimental analyses of sample variance effects in vision-and-language tasks, highlighting the need for variance reporting and dataset reliability considerations.

Findings

CIDEr metric shows larger variance than others

Human references vary significantly across datasets and tasks

Reporting variance is crucial for reliable evaluation

Abstract

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.