In the field of document understanding, significant advances have been made in the fine-tuning of Multimodal Large Language Models (MLLMs) with instruction-following data. Nevertheless, the potential of text-grounding capability within text-rich scenarios remains underexplored. In this paper, we present a text-grounding document understanding model, termed TGDoc, which addresses this deficiency by enhancing MLLMs with the ability to discern the spatial positioning of text within images. Empirical evidence suggests that text-grounding improves the model's interpretation of textual content, thereby elevating its proficiency in comprehending text-rich images. Specifically, we compile a dataset containing 99K PowerPoint presentations sourced from the internet. We formulate instruction tuning tasks including text detection, recognition, and spotting to facilitate the cohesive alignment between the visual encoder and large language model. Moreover, we curate a collection of text-rich images and prompt the text-only GPT-4 to generate 12K high-quality conversations, featuring textual locations within text-rich scenarios. By integrating text location data into the instructions, TGDoc is adept at discerning text locations during the visual question process. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple text-rich benchmarks, validating the effectiveness of our method.
Practices in the built environment have become more digitalized with the rapid development of modern design and construction technologies. However, the requirement of practitioners or scholars to gather complicated professional knowledge in the built environment has not been satisfied yet. In this paper, more than 80,000 paper abstracts in the built environment field were obtained to build a knowledge graph, a knowledge base storing entities and their connective relations in a graph-structured data model. To ensure the retrieval accuracy of the entities and relations in the knowledge graph, two well-annotated datasets have been created, containing 2,000 instances and 1,450 instances each in 29 relations for the named entity recognition task and relation extraction task respectively. These two tasks were solved by two BERT-based models trained on the proposed dataset. Both models attained an accuracy above 85% on these two tasks. More than 200,000 high-quality relations and entities were obtained using these models to extract all abstract data. Finally, this knowledge graph is presented as a self-developed visualization system to reveal relations between various entities in the domain. Both the source code and the annotated dataset can be found here: https://github.com/HKUST-KnowComp/BEKG.