Measuring Compositional Consistency for Video Question Answering

TL;DR

This paper introduces a new benchmark and evaluation metrics for assessing the compositional reasoning abilities of video question answering models, revealing their limitations in reasoning correctly and consistently.

Contribution

It develops a question decomposition engine and creates AGQA-Decomp, a large benchmark with question graphs to evaluate compositional reasoning in models.

Findings

Models often fail to reason correctly through compositions.

Models rely on incorrect reasoning or data biases.

High accuracy achieved despite reasoning failures.

Abstract

Recent video question answering benchmarks indicate that state-of-the-art models struggle to answer compositional questions. However, it remains unclear which types of compositional reasoning cause models to mispredict. Furthermore, it is difficult to discern whether models arrive at answers using compositional reasoning or by leveraging data biases. In this paper, we develop a question decomposition engine that programmatically deconstructs a compositional question into a directed acyclic graph of sub-questions. The graph is designed such that each parent question is a composition of its children. We present AGQA-Decomp, a benchmark containing $2.3M$ question graphs, with an average of $11.49$ sub-questions per graph, and $4.55M$ total new sub-questions. Using question graphs, we evaluate three state-of-the-art models with a suite of novel compositional consistency metrics. We find that models either cannot reason correctly through most compositions or are reliant on incorrect reasoning to reach answers, frequently contradicting themselves or achieving high accuracies when failing at intermediate reasoning steps.