Assessing Y-Axis Influence: Bias in Multimodal Language Models on Chart-to-Table Translation
Multimodal Language Models (MLMs) are increasingly used to convert chart images into structured tables, a process essential for answering complex data-driven queries. However, this research highlights that public chart datasets contain significant imbalances in how y-axis information is presented. These imbalances lead to unintended biases, causing MLMs to perform inconsistently depending on the visual characteristics of the chart. This paper introduces a new framework to systematically analyze these biases and understand their impact on model accuracy.
The FairChart2Table Framework
To address the lack of systematic research into these biases, the authors developed a framework called FairChart2Table. This tool is designed to evaluate how y-axis-related factors influence the performance of five state-of-the-art multimodal models. By isolating specific visual and numerical variables, the framework allows researchers to pinpoint exactly where and why these models struggle during the translation process.
Key Drivers of Bias
The study identified several specific factors that contribute to uneven performance across models. The researchers found that MLMs are sensitive to the following y-axis characteristics:
Numerical Complexity: The digit length of major tick values and the overall range of values displayed.
Scale Density: The number of major ticks present on the axis.
Formatting: How the data is presented, such as the use of abbreviations or scientific notation.
Beyond the y-axis, the study also noted that the number of legends or entities depicted within a chart image significantly impacts how well a model can translate the visual data into a table.
Improving Model Performance
A significant finding of the research is that model performance is not fixed; it can be improved through better prompting strategies. The authors discovered that when MLMs are explicitly provided with y-axis information via prompts, their ability to accurately translate charts into tables increases significantly for some models. This suggests that while inherent biases exist, targeted interventions can help bridge the gap in model reliability.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!