BaKGraSTeC: A Background Knowledge Graph Based Method for Short Text Classification

Xuhui Jiang, Yinghan Shen, Yuanzhuo Wang, Xiaolong Jin, Xueqi Cheng

Abstract: Short text classification is an important task in the area of natural language processing. Recent studies attempt to employ external knowledge to improve classification performance, but they ignore the correlation between external knowledge and have poor interpretability. This paper proposes a novel Background Knowledge Graph based method for Short Text Classification called BaKGraSTeC for short, which can not only employ external knowledge from a knowledge graph to enrich text information, but also utilize its structural information through a graph neural network to promote the understanding of texts. Specifically, we construct a background knowledge graph based on training data, then we propose a novel architecture that integrates background knowledge graph into a graph neural network to model and capture implicit interactions between its concepts and classes. Besides, we propose an attention mechanism considering both similarity and co-occurrence between concepts and classes to identify the informative concepts in texts. Our experimental results demonstrate the effectiveness with good interpretability of BaKGraSTeC through using external knowledge and their structural information for short text classification.