Two-Tier Random Walk Based Relational Inference Algorithm

Relational inference is one of the crucial techniques for knowledge base population tasks, typical application scenarios include relationship prediction and entity linking. The challenging problem of relational inference is how to infer new relations between entities from the facts existed in the knowledge bases. The reasoning models adopted in current mainstream knowledge bases can be divided into two categories: the latent factor models and the random walk models. The latent factor models realize the reasoning by mapping the entities and relations into a low dimensional real-valued vector space, and then computing with corresponding vector similarity measures. The random walk models, however, are based on the first-order predicate logic to deduce the reasoning between the entities and reduce the algorithm complexity through stochastic algorithm. In comparison, the efficiency of the latent factor models usually suffer from their computational complexity caused by large-scale matrix computation operations. While the random walk models usually suffer from their low recall rates, due to the fact that it is difficult to fully utilize all of the available structure information provided by the knowledge bases with any random sampling design. This work studied the potential problems of the basic assumptions adopted by the existing random walk models, and proposed two new inference modeling assumptions thereby. Firstly, the random walk models represented by the Path Ranking Algorithm (PRA) adopt the unidimensionality assumption of the relationships in between the entities. In typical random walk models, the entity-relation-entity tuples that existed in the knowledge base are regarded as first-order Horn clauses, in which the relationships are treated as partial ordering relations between the subjects and the objects. Our hypothesis is that although the relation between two entities is literally, syntactically directional, the information conveyed by this relation is equally shared between the connected entities on both side, thus all of the relations are semantically bidirectional, which allows the relational reasoning algorithm to use the inverse relation semantics from the object to the subject for reasoning. Secondly, the PRA algorithm makes use of the first-order predicate logic for relational reasoning, it also introduces a random sampling scheme in order to avoid exhaustive search in the path space and to speed up the calculation process. However, we argue in this paper that this maybe one of the major reasons that explains why the PRA and alike algorithms can not make full use of the existing information in the knowledge base for relational inference tasks, and then we propose an alternative assumption for remedy, which claims that the topology structures of the relation-specific sub-graphs in knowledge bases can be exploited to improve the performance of the random-walk based relational inference algorithms. In order to verify the validity of the above assumptions and algorithm, we propose a novel relational inference algorithm based on a two-tier random walk strategy. Experimental results on benchmark datasets include WN18, FB15K, and FB40K, show that the proposed algorithm can be very effective in promoting the accuracy and recall rate of the random walk models. The proposed algorithm also outperforms other prevalent latent factor models on each data sets. © 2017, Science Press. All right reserved.