TableBank: Table Benchmark for Image-based Table Detection and Recognition

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1918–1925 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1918 TableBank: Table Benchmark for Image-based Table Detection and Recognition Minghao Li1, Lei Cui2, Shaohan Huang2, Furu Wei2, Ming Zhou2, Zhoujun Li1 1State Key Lab of Software Development Environment, Beihang University, Beijing, China 2Microsoft Research Asia, Beijing, China {liminghao1630, lizj}@buaa.edu.cn, {lecu, shaohanh, fuwei, mingzhou}@microsoft.com Abstract We present TableBank, a new image-based table detection and recognition dataset built with novel weak supervision from Word and Latex documents on the internet. Existing research for image-based table detection and recognition usually fine-tunes pre-trained models on out-of-domain data with a few thousand human-labeled examples, which is difficult to generalize on real-world applications. With TableBank that contains 417K high quality labeled tables, we build several strong baselines using state-of-the-art models with deep neural networks. We make TableBank publicly available and hope it will empower more deep learning approaches in the table detection and recognition task. The dataset and models can be downloaded from https://github.com/doc-analysis/TableBank. Keywords: TableBank, table detection and recognition, weak supervision, image-based deep learning network 1. Introduction Table detection and recognition is an important task in many document analysis applications as tables often present essential information in a structured way. It is a difficult problem due to varying layouts and formats of the tables as shown in Figure 1. Conventional techniques for table analysis have been proposed based on the layout anal- ysis of documents. Most of these techniques fail to gener- alize because they rely on handcrafted features which are not robust to layout variations. Recently, the rapid develop- ment of deep learning in computer vision has significantly boosted the data-driven image-based approaches for table analysis. The advantage of the image-based table analy- sis lies in its robustness to document types, making no as- sumption of whether scanned images of pages or natively- digital document formats. It can be applied to a wide va- riety of document types to extract tables including PDF, HTML, PowerPoint as well as their scanned copies. Al- though some document types may have structured tabular data, we still need a general approach to detect tables on different kinds of documents. Therefore, large-scale end- to-end deep learning models make it feasible for achieving better performance. Existing deep learning based table analysis models usually fine-tune pre-trained object detection models with several thousand human-labeled training instances, while still diffi- cult to scale up in real-world applications. For instance, we found that models trained on data similar to Figure 1a 1b 1c would not perform well on Figure 1d because the table lay- outs and colors are so different. Therefore, enlarging the training data should be the only way to build open-domain table analysis models with deep learning. Deep learning models are massively more complex than most traditional models, where many standard deep learning models today have hundreds of millions of free parameters and require considerably more labeled training data. In practice, the The paper …

Content truncated to 500 words