Doc-GCN: Heterogeneous Graph Convolutional Networks for Document Layout Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 2906–2916 October 12–17, 2022. 2906 Doc-GCN: Heterogeneous Graph Convolutional Networks for Document Layout Analysis Siwen Luo1∗, Yihao Ding 1*, Siqu Long1, Josiah Poon1, Soyeon Caren Han1, 2† 1The University of Sydney, 2 The University of Western Australia {siwen.luo, yihao.ding, josiah.poon, caren.han}@sydney.edu.au, caren.han@uwa.edu.au, slon6753@uni.sydney.edu.au Abstract Recognizing the layout of unstructured digital documents is crucial when parsing the docu- ments into the structured, machine-readable for- mat for downstream applications. Recent stud- ies in Document Layout Analysis usually rely on visual cues to understand documents while ignoring other information, such as contextual information or the relationships between doc- ument layout components, which are vital to boost better layout analysis performance. Our Doc-GCN presents an effective way to harmo- nize and integrate heterogeneous aspects for Document Layout Analysis. We construct dif- ferent graphs to capture the four main features aspects of document layout components, in- cluding syntactic, semantic, density, and ap- pearance features. Then, we apply graph con- volutional networks to enhance each aspect of features and apply the node-level pooling for integration. Finally, we concatenate features of all aspects and feed them into the 2-layer MLPs for document layout component classi- fication. Our Doc-GCN achieves state-of-the- art results on three widely used DLA datasets: PubLayNet, FUNSD, and DocBank. The code will be released at https://github.com/ adlnlp/doc_gcn 1 Introduction Digital documents (incl. Scanned Document Im- ages and PDF files) are popular and convenient for storing written textual information, so almost 2.5 trillion documents worldwide are available in the digital format (Zhong et al., 2019). However, it is challenging to automatically recognize the layout and components of these unstructured digital doc- uments and extract meaningful information using this format. For example, the financial office would require the scanned document image after their client signed. It is then crucial to recognize and ∗Co-First Authors †Corresponding Author (caren.han@sydney.edu.au) extract the form component, such as the form title, person name, and the date the document is signed. This task is widely called Document Layout Anal- ysis (DLA). The DLA task aims at understanding the documents from either 1) the physical analy- sis by detecting the document structure and the boundaries of each layout region or 2) the logical analysis by categorizing the detected layout compo- nents (segments) into the predefined document ele- ment classes, such as Title, Date, Author, and Fig- ure (Binmakhashen and Mahmoud, 2019). In this research, we focus on the logical DLA task to clas- sify the different layout components of PDF docu- ments by understanding the relationships between components. Traditional deep learning-based DLA approaches mainly focus on processing visual fea- tures of layout components (Soto and Yoo, 2019; Augusto Borges Oliveira and Palhares Viana, 2017; Li et al., 2020a) using CNN-based models. Some recent studies started to use texts to solve the prob- lem with the support of semantic information for each layout component (Li et al., 2020b; Xu et al., 2020). However, applying visual and textual fea- tures is not enough to analyze …

Content truncated to 500 words