为有效利用PDF文献中的非结构化文本数据,面向费托合成催化材料领域文献,设计了关键信息抽取流水线从PDF文献中抽取表格及其相应注释等关键信息.以微分二值化网络(differentiable binarization network,DBNet)为基准模型,通过引入自适应空间注意力(adaptive spatial attention,ASA)模块,提出了 DB-ASA文本检测模型,提高了检测精度.采用单视觉文本识别模型(scene text recognition with a single visual model,SVTR)进行文本识别,结合领域字典文件在自建数据集上对模型进行微调,文本识别准确率可达93.87％.
A literature information extraction method for catalytic materials
In order to effectively utilize the unstructured text data in PDF literature in the Fischer-Tropsch synthesis of catalytic materials,a key information extraction pipeline was designed to extract key information such as tables and corresponding annotations from PDF documents.A DB-ASA text detection model was proposed by using the differentiable binarization network(DBNet)as a benchmark model and introducing an adaptive spatial attention(ASA)module,resulting in improved detection accuracy.Using scene text recognition with a single visual model(SVTR)for text recognition,the model was fine-tuned on a self-built dataset by combining domain dictionary files,achieving a text recognition accuracy of 93.87％.
万方数据
维普
