Extracting Structural Knowledge for Professional Text Inference

Tianyu Xia, Jian Wang, Tianyuan Liu, Hailan Jiang, Yuqing Sun

Abstract: Grading subjective questions of specialty text is a kind of text inference task. Since there are many specialty terms and concepts, it is difficult to judge the knowledge contained in a text as the usual way on inferring a general text. In this paper, we propose a specialty text inference model by extracting the structural knowledge from text. We first propose a knowledge graph construction method for the extraction of knowledge from specialty texts. By combining the constructed knowledge features with the text semantic features, we design the specialty text inference model. Finally, we use real datasets from a national professional exam to validate the soundness of the knowledge graph construction method and the performance of the inference model. The experiments under different training set sizes and network structures are also conducted to detailly analyze the design of our method. The experimental results show the effectiveness and practicality of our approach.