Knowledge Representation With Geometric Transformations

Knowledge graph is a set of knowledge items with the relationship between the head and tail entities.On one hand,knowledge has always been the central subject of artificial intelligence,which carries the theoretical and industrial foundations.On the other hand,knowledge as a powerful booster of intelligent systems,can be a critical necessity to Internet applications,resulting in greater social and economic benefits,such as Baidu,Google,etc.However,there are two flaws for the traditional logic-based knowledge representations.First,it could not be jointly combined with most machine learning technologies,which hinders the latent potentials of knowledge graph.Second,logical reasoning is very time-consuming,making it difficult to be applied in large industrial datasets.To fix these issues,this paper presents a series of knowledge graph representation methods based on geometric translations.Each entity corresponds to a point in geometric space,and each relation is presented by a geometric transformation.Thus,the effective algorithms based on statistical learning are designed to solve problems such as knowledge graph completion.In this paper,the graph representations are explored from four aspects: the geometry of data,the geometry of model,the geometry of interaction and the geometry of semantic interpretation.Model characterizes data and data selects model.Due to this fact,this paper explores this task from the two dual aspects,that of data and model.Furthermore,the geometric interactions between data and model are analyzed.(1)At the level of data,this paper analyzes the representation of relation and discovers the multiple relation semantics.This phenomenon,mainly refers that a relation always presents multiple semantics.For example,the "Has Part" relation,which indicates the geographical location semantics in(China,Has Part,Beijing),while it represents the composition semantics in(Table,Has Part,Table Leg).In this paper,a mixture model(TransG)is used to leverage this phenomenon.Each triple is a mixture of different relational semantics.This method can not only improve the accuracy of knowledge related tasks,but could also demonstrate each semantics of this relation.The experimental results show that TransG dominates the traditional single relation semantic model.This work has been published in ACL 2016.(2)At the level of model,this paper analyzes the geometric singularity of the existing models and makes the corresponding algebraic interpretations.The traditional model regards the triple as translation process,from the head point to the tail one via relation vectors.The nature of geometric representations constrains each head corresponding to a specific tail under a concrete predicate,which is over-strict for N-N relations.In this paper,for a specific head and relation,we extend the standard position from one point to a manifold.Our method(ManifoldE)overcomes the geometric singularity and algebraic ill-posed issues,thus being capable of reducing the representation noise.Experimental results show that ManifoldE is much better than other similar models in link prediction tasks.This work has been published in IJCAI 2016.(3)At the level of interaction,this paper introduces the entity descriptions to enhance the knowledge representations.Traditionally,the entity descriptions and knowledge triples are encoded within a relatively independent process.In this paper,the representations of triple are always the main procedure for knowledge representation.Entity description information must interact with triple information to better model the knowledge and textual semantics.Our method(SSP)constructs a semantic hyperplane using the entity description.Then the representation of triple is computed in the corresponding semantic hyperplane.This paper could find the semantic correlations between entities.Experimental results show the effectiveness of SSP.This work has been published in AAAI 2017.(4)At the level of interpretability,this paper demonstrates a geometric principle of semantic representations.In the traditional models,geometric representation is difficult to explain,which are unfavorable both in theory and application.In this paper,we apply the multi-view clustering method to model the semantic organization in the knowledge graph.This method(KSR)has achieved a semantic break-through that can be better integrated into natural language comprehension.Experiments verify the semantics of our method,which achieves the best results in practical large-scale datasets.