Augmenting Inputs using a Novel Figure-to-Text Pipeline to Assist Visual Language Models in Answering Scientific Domain Queries

Abstract

Recent advancements in visual language models (VLMs) have transformed the way we interpret and interact with digital imagery, bridging the gap between visual and textual data. However, these models, like Bard, GPT4-v, and LLava, often struggle with specialized fields, particularly when processing scientific imagery such as plots and graphs in scientific literature.

In this thesis, we discuss the development of a pioneering reconstruction pipeline to extract metadata, regenerate plot data, and filter out extraneous noise like legends from plot images. Ultimately, the collected information is presented to the VLM in structured, textual manner to assist in answering domain specific queries. The efficacy of this pipeline is evaluated using a novel dataset comprised of scientific plots extracted from battery domain literature, alongside the existing benchmark datasets including PlotQA and ChartQA. Results about the component accuracy, task accuracy, and question-answering with augmented inputs to a VLM show promise in the future capabilities of this work.

By assisting VLMs with scientific imagery, we aim to not only enhance the capabilities of VLMs in specialized scientific areas but also to transform the performance of VLMs in domain specific areas as a whole. This thesis provides a detailed overview of the work, encompassing a literature review, methodology, results, and recommendations for future work.