Synthetic Multimodal Question Generation

TL;DR

This paper introduces SMMQG, a synthetic data generation framework that creates high-quality multimodal question-answer pairs for evaluating multimodal retrieval augmented generation models, addressing dataset scarcity issues.

Contribution

The paper presents SMMQG, a novel framework leveraging retrievers, LLMs, and LMMs to generate style- and modality-specific multimodal QA datasets from documents.

Findings

SMMQG-generated data quality matches crowdsourced benchmarks.

Evaluation on generated data provides new insights into model performance.

Human study confirms the synthetic data's quality is comparable to real datasets.

Abstract

Multimodal Retrieval Augmented Generation (MMRAG) is a powerful approach to question-answering over multimodal documents. A key challenge with evaluating MMRAG is the paucity of high-quality datasets matching the question styles and modalities of interest. In light of this, we propose SMMQG, a synthetic data generation framework. SMMQG leverages interplay between a retriever, large language model (LLM) and large multimodal model (LMM) to generate question and answer pairs directly from multimodal documents, with the questions conforming to specified styles and modalities. We use SMMQG to generate an MMRAG dataset of 1024 questions over Wikipedia documents and evaluate state-of-the-art models using it, revealing insights into model performance that are attainable only through style- and modality-specific evaluation data. Next, we measure the quality of data produced by SMMQG via a human study. We find that the quality of SMMQG-generated synthetic data is on par with the quality of the crowdsourced benchmark MMQA and that downstream evaluation results using both datasets strongly concur.