Improving fragment information recommendation via enhanced knowledge graph

Abstract

Fragment information refers to irregular, unorganized and unstructured data scattered across various sources such as texts or images in public newspapers, reports, and other online articles. Due to its small units, large quantity and low value density, it is not easy to discover the related fragments for queries from a specific domain. To make better use of domain knowledges, this paper utilizes the entity co-occurrence relationship in domain to enhance base knowledge graph, followed by fusing the enhanced graph representations to obtain the entity-related knowledge representation via attention mechanism. By enhancing the base knowledge graph with frequency statistics of entity co-occurrence, the neighborhood information can be encoded into entity representations, thereby providing entity-related domain knowledge. We then linearly combine the entity representation with the pretrained representation of whole fragment texts and calculate the similarity between the resulted vectors to retrieve top-k related items for fragment information recommendation. Finally we demonstrate the effectiveness of the proposed method through experiments on the MIND dataset.