Research On Hypernymy Recognition Based On Graph Contrastive Learning

Hypernymy is a fundamental semantic relationship in computer languages,which mainly expresses the ’is-a’ relation among generic concepts.In much of the downstream work in natural language processing,hypernymy is a key component.Therefore,recognition of hypernymy relation is a meaningful research task.Existing methods for hypernymy recognition include pattern based method and distributional method.The two methods are relatively mature,whereas expose some shortcomings.The pattern based method does not make good use of the Web-scale taxonomies;while the distributional method faces trouble when it tries to distinguish the hypernymy from other semantic relations.Moreover,it is difficult to model the mapping rule between a term and its hypernymy or non-hypernymy in the embedding space.In recent years,researches on word embedding-based methods has emerged.Although the existing word embedding-based hypernymy recognition methods have achieved good results,there are still some problems:(1)few methods can make full use of the hypernymy relations in the web-scale taxonomies,so that the rich contextual knowledge contained therein cannot be exploited;(2)the existing word embeddings are based on random initialization,which cannot effectively capture the asymmetry and metastability of hypernymy relations;(3)the existing models have the limitation that they do not adequately exploit the relationship between the predicted vector and the actual projection.To solve the above problems,this paper proposes two methods for hypernymy recognition:(1)The paper proposes a hypernymy recognition models based on word embedding projection.Two real hypernymy and non-hypernymy mapping networks are trained using existing taxonomies to learn how terms are mapped to its hypernymy and non-hypernymy in the embedding space;A new objective function is designed.It enables the model to map the vector of the predicted term to both its hypernymy and non-hypernymy,while taking into account the similarity of its hypernymy and nonhypernymy vectors in the mapping space and reducing the probability of overlapping hypernymy and non-hypernymy vectors,so as to better distinguish between the hypernymy and non-hypernymy.In addition,a cascading layer is designed to prevent model overfitting,make full use of the learned node features and speed up the convergence of the model and improve its generalization performance.Experimental results on two datasets demonstrate that the hypernymy recognition model based on word embedding projection can effectively exploit the hypernymy relations in webscale texonomies and improve the accuracy of relationship recognition.(2)Hypernymy recognition based on graph contrastive learning is proposed in this paper.The model uses graph contrastive learning to generate augmented subgraphs by data augmentation of the relationship graph so that the target node features in different augmented subgraphs have the highest similarity,and learn unperturbed feature representations based on maximizing mutual information between positive and negative samples,which contains global representations of the entire subgraph centred on the target node,i.e.feature invariance is achieved by specific expansion of the data.A comparison of the experimental results on two benchmark datasets demonstrates a significant improvement in accuracy over existing relatively advanced methods.