Joint Semantic Relevance Learning with Text Data and Graph Knowledge

Inferring semantic relevance among entities (e.g., entries of Wikipedia) is important and challenging. According to the information resources, the inference can be categorized into learning with either raw text data, or labeled text data (e.g., wiki page), or graph knowledge (e.g, WordNet). Although graph knowledge tends to be more reliable, text data is much less costly and offers a better coverage. We show in this paper that different resources are complementary and can be combined to improve semantic learning. Particularly, we present a joint learning approach that learns vectors of entities by leveraging resources of both text data and graph knowledge. The experiments conducted on the semantic relatedness task show that text-based learning works well on general domain tasks, however for tasks in specific domains, joint learning that involves both text data and graph knowledge offers significant improvement.