Improving Knowledge Base Completion by Incorporating Implicit Information

Abstract

Over the past few years, many large Knowledge Bases (KBs) have been constructed through relation extraction technology but they are still often incomplete. As a supplement to training a more powerful extractor, Knowledge Base Completion which aims at learning new facts based on existing ones has recently attracted much attention. Most of the existing methods, however, are only utilizing the explicit facts in a single KB. By analyzing the data, we find that some implicit information should also been captured for a more comprehensive consideration during completion process. These information include the intrinsic properties of KBs (e.g. relational constraints) and potential synergies between various KBs (i.e. semantic similarity). For the former, we distinguish the missing data by using relational constraints to reduce the data sparsity. For the later, we incorporate two semantical regularizations into the learning model to encode the semantic similarity. Experimental results show that our approach is better than the methods that consider only explicit facts or only a single knowledge base, and achieves significant accuracy improvements in binary relation prediction.