Sequence-to-Sequence Pre-training with Unified Modality Masking for Visual Document Understanding

TL;DR

GenDoc is a versatile sequence-to-sequence model pre-trained with unified masking across text, image, and layout modalities, enabling improved performance and robustness in diverse document understanding tasks.

Contribution

The paper introduces GenDoc, a novel encoder-decoder pre-training framework with unified modality masking and strategies for effective multi-modal integration in document understanding.

Findings

Achieves superior or competitive results on multiple downstream tasks.

Demonstrates robustness under imperfect OCR conditions.

Extends pre-training to include masked image and layout prediction tasks.

Abstract

This paper presents GenDoc, a general sequence-to-sequence document understanding model pre-trained with unified masking across three modalities: text, image, and layout. The proposed model utilizes an encoder-decoder architecture, which allows for increased adaptability to a wide range of downstream tasks with diverse output formats, in contrast to the encoder-only models commonly employed in document understanding. In addition to the traditional text infilling task used in previous encoder-decoder models, our pre-training extends to include tasks of masked image token prediction and masked layout prediction. We also design modality-specific instruction and adopt both disentangled attention and the mixture-of-modality-experts strategy to effectively capture the information leveraged by each modality. Evaluation of the proposed model through extensive experiments on several downstream tasks in document understanding demonstrates its ability to achieve superior or competitive performance compared to state-of-the-art approaches. Our analysis further suggests that GenDoc is more robust than the encoder-only models in scenarios where the OCR quality is imperfect.