Event-Driven News Stream Clustering using Entity-Aware Contextual Embeddings

We propose a method for online news stream clustering that is a variant of\nthe non-parametric streaming K-means algorithm. Our model uses a combination of\nsparse and dense document representations, aggregates document-cluster\nsimilarity along these multiple representations and makes the clustering\ndecision using a neural classifier. The weighted document-cluster similarity\nmodel is learned using a novel adaptation of the triplet loss into a linear\nclassification objective. We show that the use of a suitable fine-tuning\nobjective and external knowledge in pre-trained transformer models yields\nsignificant improvements in the effectiveness of contextual embeddings for\nclustering. Our model achieves a new state-of-the-art on a standard stream\nclustering dataset of English documents.\n