Publication

Abstract : Digital technology and World Wide Web have re-sulted in a growth in the number of digital documents. The ma-jority of the data is unstructured, and extracting information into a structured machine-readable format remains a difficult under-taking. Clustering, which automatically categorizes information into meaningful groupings, is one of the most important activ-ities. Several information extraction and information gathering applications use document clustering. Document clustering is an unsupervised method for dividing a big corpus of documents into smaller, meaningful, identifiable, and verifiable sub-groups. But capturing the semantics of the documents is still an open problem. A knowledge graph can represent the relationships between the entities in the document collection. But a knowledge graph gets extremely dense and high-dimensional as the amount of data increases, requiring significant processing resources. We aim to explore this problem by using Knowledge Graph Embedding (KGE), which maps the high-dimensional representation into a compute-efficient low-dimensional embedded representation and then cluster these embeddings using the Gaussian Mixture Model (GMM)-based clustering technique. Dimensionality reduction of the embeddings has been done using t-SNE. We have found that the silhouette coefficient has improved considerably for t-SNE based GMM clustering as compared to Kmeans clustering alone.