VLC-BERT: Visual Question Answering with Contextualized Commonsense Knowledge

TL;DR

VLC-BERT is a novel vision-language model that incorporates contextualized commonsense knowledge from COMET to improve visual question answering, especially for knowledge-intensive questions, outperforming models using static knowledge bases.

Contribution

This work introduces VLC-BERT, a new pre-trained transformer that integrates contextualized commonsense knowledge with visual and textual information for VQA tasks.

Findings

VLC-BERT outperforms static knowledge base models on OK-VQA and A-OKVQA datasets.

Contextualized knowledge benefits certain question types more than others.

Analysis reveals which questions gain from COMET-based knowledge integration.

Abstract

There has been a growing interest in solving Visual Question Answering (VQA) tasks that require the model to reason beyond the content present in the image. In this work, we focus on questions that require commonsense reasoning. In contrast to previous methods which inject knowledge from static knowledge bases, we investigate the incorporation of contextualized knowledge using Commonsense Transformer (COMET), an existing knowledge model trained on human-curated knowledge bases. We propose a method to generate, select, and encode external commonsense knowledge alongside visual and textual cues in a new pre-trained Vision-Language-Commonsense transformer model, VLC-BERT. Through our evaluation on the knowledge-intensive OK-VQA and A-OKVQA datasets, we show that VLC-BERT is capable of outperforming existing models that utilize static knowledge bases. Furthermore, through a detailed analysis, we explain which questions benefit, and which don't, from contextualized commonsense knowledge from COMET.