Domain Taxonomy Expansion Based On Graph Neural Networks From Text Corpora

Taxonomy is a hierarchically concept graph,which organizes hypernym-hyponym relationships between concepts,and it needs to mine the semantic relationships and induct a hierarchical structure.Taxonomy is of great significance to data analysis,and plays an important role in many applications,such as personalized recommendation,question answering,web search and so on.The primary problem is to extract some elements for taxonomy expansion,including entity attributes and their relationships.The second is to construct the domain taxonomy by unsupervised or supervised methods.Existing methods for constructing a taxonomy from text corpora rely on certain simple patterns and clustering algorithms.Traditional methods based on patterns largely ignore implicit features of context,complex structural semantics and many other latent features inside the concept terms;Clustering algorithms are employed to cluster similar terms into a taxonomy,where corpora are composed of different classifications and balanced types.In essence,clustering algorithms are slow computationally and needs too strong assumptions.Regarding of the issue above,the model based on graph neural networks is proposed to get concept feature,and the method based on joint similarity is put forward to construct a taxonomy.The main research work is as follows:(1)Existing methods for leveraging pattern-based could ignore cross-text and long-distance features,so the paper proposes a model based on graph neural networks,which can capture concept features.Specifically,the model uses point-wise mutual information(PMI)and node connection to construct a graph,which contains global concept similar information and local concept co-occurrence information;and it gets concept embedding by the text classification.On two real data sets,it can be seen that the concept embedding based on graph neural networks can shou good semantic features.(2)The methods based on clustering algorithms usually leverage simple features to classify,which could ignore concept co-occurrence similarity.The paper puts forward to a method based on joint similarity to construct a taxonomy.In particular,the paper uses similarity calculation to extract the relationship,and concept entropy to infer the hypernmy,and constructs the domain taxonomy.The experimental results conducted on two real-world datasets show that the method outperforms state-of-theart models,which confirms the effectiveness of the method.