GFTE: Graph-Based Financial Table Extraction

Abstract

Tabular data is a crucial form of information expression, which can organize data in a standard structure for easy information retrieval and comparison. However, in financial industry and many other fields tables are often disclosed in unstructured digital files, e.g. Portable Document Format (PDF) and images, which are difficult to be extracted directly. In this paper, to facilitate deep learning based table extraction from unstructured digital files, we publish a standard Chinese dataset named FinTab, which contains more than 1,600 financial tables of diverse kinds and their corresponding structure representation in JSON. In addition, we propose a novel graph-based convolutional neural network model named GFTE as a baseline for future comparison. GFTE integrates image feature, position feature and textual feature together for precise edge prediction and reaches overall good results 1 .

LGPMA: Complicated Table Structure Recognition with Local and Global Pyramid Mask Alignment

“…The early techniques about table detection and recognition can be found in the comprehensive survey [37]. With the great success of deep neural network in computer vision field, works began to focus on the image-based table with more general structures [21,30,24,9,13,36,23,14,33]. According to the basic components granularities, we roughly divide previous methods into two types: global-object-based methods and local-object-based methods.…”

“…After that, a group of methods [36,23,38] tries to recover the cell relations based on some heuristic rules and algorithms. Another type of methods [11,2,14,24,26] treat the detected boxes as nodes in a graph and attempt to predict the relations based on techniques of Graph Neural Networks [29]. [14] predicts the relations between nodes in three classes (the horizontal connection, the vertical connection, no connection) using several features such as visual features, text positions, word embedding, etc.…”

“…Another type of methods [11,2,14,24,26] treat the detected boxes as nodes in a graph and attempt to predict the relations based on techniques of Graph Neural Networks [29]. [14] predicts the relations between nodes in three classes (the horizontal connection, the vertical connection, no connection) using several features such as visual features, text positions, word embedding, etc. [2] adopts graph attention mechanism to enhance the predicting accuracy.…”

“…[24] alleviates the problem of large graph nodes numbers by the pair sampling strategy. The above three works [14,2,24] also published new table datasets for this research area. Since there is no empty cell detected, local-object based-methods usually fall into empty cell ambiguity.…”

“…Another group of methods solves the above problems in a bottom-up way to firstly detect the text blocks' positions and then recover the bounding-boxes' relations by heuristic rules [38] or GNN(Graph Neural Networks) [29,14,2,24,26]. However, rules designed based on bounding boxes of text regions are vulnerable to handling complicated matching situations.…”