融合容错机制的基于Attention-Mask RCNN地质表格信息抽取方法

地质表格信息提取是地质报告从信息转换到知识阶段的重要任务之一，对将非结构化的数据转化为结构化的地学知识具有重要意义，同时还为文本与表格的知识关联提供了技术支撑。然而现有的表格解析方法在地学领域存在局限性，在单元格提取中，地质表格中大量的合并单元格造成了不同单元格间大小差异大，大量小面积单元格无法被提取；在表格解析方面地质表格包含了大量的被斜线分割的特殊表头，难以自动化解析。为解决上述问题，本文提出了一种基于注意力机制的Mask RCNN单元格提取模型及基于OpenCV框架的表格结构解析方法。主要包括两个步骤：1）上下文注意模块（CAM）学习上下文特征以识别不同大小单元格；2）一种标准容错机制的复杂表头解析方法，解析含斜线分割的复杂表头单元格。在构建的地质表格数据集上进行模型性能评估，该方法对于多数地质表格的解析准确率达到95% 以上；相比其他单元格识别和表格结构解析方法，该方法解析效果更优。

Geological table information extraction method based on Attention-Mask RCNN with fault-tolerant mechanism

Extraction of geological table information is one of the important tasks of geological report transformation from information to knowledge, which is of great significance for transforming unstructured data into structured geological knowledge, and also provides technical support for knowledge association between text and table. However, the existing table parsing methods have limitations in the field of geosciences. In the extraction of cells, a large number of combined cells in the geological table result in large size differences among different cells, and a great number of small area cells cannot be extracted. In the aspect of table parsing, the geological table contains a lot of special headers divided by slashes, which is difficult to parse automatically. To solve the above problems, a Mask RCNN cell extraction model based on attention mechanism and a table structure parsing method based on OpenCV framework are proposed. It consists of two main steps: 1)the Context Awareness Module(CAM)learns the context features to identify cells of different sizes; 2)a complex table header parsing method with standard fault tolerance mechanism, which analyzes complex table header cells with slash division. The model performance was evaluated on the data set of the constructed geological tables, and the analytical accuracy of the method for most geological tables reached more than 95%. Compared with other cell recognition and table structure parsing methods, this method has better parsing effect.