5W1H-Based Semantic Segmentation of Tweets for Event Detection Using BERT

Detection of events from Twitter has been one of the significant areas in the Text Mining domain due to the volume of content generated by online users. Twitter is considered as one of the top sources for disseminating information to the users. Due to the short length of texts on Twitter, the content generated is often noisy, which makes the detection of events very difficult. Though research on Twitter event detection has been in existence, most of them focused on implementing statistical measures rather than exploiting the semantics. The work presented in this paper presents an approach for the semantic segmentation of Twitter texts (tweets) by adopting the concept of 5W1H (Who, What, When, Where, Why and How). 5W1H represent the semantic constituents (subject, object and modifiers) of a sentence and the actions of verbs on them. The relationship between a verb and the semantic constituents of a sentence forms the basis for representation of an event. The basic approach of the proposed system is to segment the tweets based on the 5W1H contextual word embeddings generated with the help of recent state-of-the-art technology and then clustering the tweets for the representation of possible events. We compared our approach with a simple baseline system that does not segment the tweets. We evaluated the performance of both the approaches by measuring the cosine similarity of the tweets under a cluster. Our 5W1H segmentation approach produced a similarity score above 82% for the most similar tweets in a cluster against the baseline system that scored below 70%.