Improving Multi-label Text Classification Models with Knowledge Graphs

Multi-label Text Classification (MLTC) is a variant of classification problem where multiple labels are assigned to each instance. Most existing MLTC methods ignore the relationship between the target labels. Since the hierarchical relationship for addressing these problems is significant, a semantic network approach with the help of knowledge graphs can be used. This paper proposes a knowledge graph-based approach together with GRU (Gated Recurrent Unit) neural network model to solve an MLTC problem on a research text dataset. In particular, we leverage the Tax2Vec approach to extract hypernyms from the WordNet knowledge graph and enrich the dataset. The enrichment results in following a tree-like structure to identify the relationship between the semantic concepts. The result shows that the enriched dataset outperforms the traditional GRU neural network-based model based on different evaluation metrics.