Every Part Matters: Integrity Verification of Scientific Figures Based on Multimodal Large Language Models

TL;DR

This paper introduces a novel task, Figure Integrity Verification, designed to evaluate the precision of technologies in aligning textual knowledge with visual elements in scientific figures, and proposes an innovative framework, Every Part Matters, which leverages Multimodal Large Language Models to not only incrementally improve the alignment and verification of text-figure integrity but also enhance integrity through analogical reasoning.

Abstract

This paper tackles a key issue in the interpretation of scientific figures: the fine-grained alignment of text and figures. It advances beyond prior research that primarily dealt with straightforward, data-driven visualizations such as bar and pie charts and only offered a basic understanding of diagrams through captioning and classification. We introduce a novel task, Figure Integrity Verification, designed to evaluate the precision of technologies in aligning textual knowledge with visual elements in scientific figures. To support this, we develop a semi-automated method for constructing a large-scale dataset, Figure-seg, specifically designed for this task. Additionally, we propose an innovative framework, Every Part Matters (EPM), which leverages Multimodal Large Language Models (MLLMs) to not only incrementally improve the alignment and verification of text-figure integrity but also enhance integrity through analogical reasoning. Our comprehensive experiments show that these innovations substantially improve upon existing methods, allowing for more precise and thorough analysis of complex scientific figures. This progress not only enhances our understanding of multimodal technologies but also stimulates further research and practical applications across fields requiring the accurate interpretation of complex visual data.

Figures (11)

• Figure 1: Comparison of alignment tasks for natural images and scientific figures. The text-figure alignment task requires models to parse each module element within a figure, align the text, and identify unaligned elements (Integrity Verification). Additionally, it requires models to provide supplementary descriptions for unaligned elements through figure understanding (Integrity Augmentation). The natural images shown in the examples are from the CoCo dataset lin2014microsoft, and the scientific figures are from wang2022dataset.
• Figure 2: Overview of the construction process for a fine-grained dataset aligning scientific figures with text, comprising four sequential stages: preparation, filtering, processing, and annotation. The collaboration between automated tools and human annotators facilitates the execution of the filtering, processing, and annotation stages.
• Figure 3: Instances of semantic enhancement based on the figure from the study by zhao2012collocation. In their figure, 'Sentiment analysis' modules appear multiple times. To differentiate these modules more precisely, we incorporate spatial and semantic information.
• Figure 4: The overall framework for implementing the integrity verification task. Figure (a) depicts the key modules and the implementation process required to complete the task. Figure (b) elaborately illustrates the training and inference process of the core module - the Text-figure alignment Model, which is crucial for aligning figure modules with text.
• Figure 5: Chain-of-Attribute (CoA)