Word Embedding based Event Identification and Labeling of Connected Events from Tweets

Events conceived as facts which are fine grained entities that happen around us such as completing graduation, birthday celebration, and death of an individual. Events are regarded as happening in a certain place, during a particular interval of time which can be occurred planned or unplanned way. Social media is a platform where users share their attending events with others. In this paper, we present a novel machine learning based approach to identify events from social media, i.e., Twitter, by using Bidirectional Encoder Representations from Transformers (BERT) based word embedding technique. Events might be connected with each other which might have real life implications such as identifying causal-effect, investigating criminal activities, etc. We also demonstrate a mechanism which can organize relevant events into a cluster based on their spatiotemporal properties. Later, we develop an unsupervised connected event labeling technique by using BERT word embedding approach by exploiting its semantic strength from the content of tweets. Our model shows an outstanding performance which has an accuracy of 91%. We also compare our approach with two competitive baseline techniques (i.e., word2vec and tf-idf) to identify events and our model shows better performance (on an average 5% better accuracy) than that of those baseline models.