Abstract:
The paper presents a method for analyzing the layout of PDF documents based on graph neural networks (GNN), which uses words as graph nodes to overcome the limitations of modern approaches based on strings or local areas. The proposed WordGLAM model, based on modified graph convolutional layers, demonstrates the possibility of constructing hierarchical structures through word aggregation, which ensures a balance between the accuracy of element detection and their semantic connectivity. Despite lagging behind state-of-the-art models (for example, Vision Grid Transformer) in accuracy metrics, the study reveals systemic problems of the region: data imbalance, ambiguity in word clustering ("chain links", "bridges" between unrelated regions), as well as controversial criteria selecting classes in the markup. The key contribution of this work is the formulation of new research tasks, including optimization of vector representations of words, consideration of edge embeddings, and development of estimation methods for complex word hierarchies. The results confirm the prospects of the approach for creating adaptable models capable of processing multi-format documents (scientific articles, legal texts). This paper highlights the need for further research in the field of regularization and extension of training data, opening up ways to improve the portability of layout analysis methods to new domains. The code and models were published on GitHub (https://github.com/YRL-AIDA/wordGLAM).
Fulltext:
PDF file (1428 kB)