GraphEDM: A Graph-Based Approach to Disambiguate Entities in Microposts

The use of microblogging platforms such as Twitter has been growing rapidly. With about 500M tweets published per day, tweets are becoming a valuable source of information for several tasks such as event detection, sentiment analysis, or opinion mining, and are being leveraged by many prominent organizations. However, one must first be able to correctly capture the semantic content of a tweet prior to leveraging it for any automated analysis. Automatically understanding tweets is extremely challenging, as the information they contain is limited and insufficient for algorithms that need a larger context. In this work, we propose an approach that extends the context of a micropost by leveraging graph-based algorithms to further disambiguate the entities present in it. Our approach, GraphEDM, is divided into two phases. First, we use unsupervised clustering approaches to regroup tweets in semantic neighborhoods using embedding approaches. Next, each ambiguous entity in a cluster is iteratively disambiguated by leveraging a graph-based algorithm. Our experimental results reveal that GraphEDM outperforms the state of the art in tweet entity disambiguation by up to 15.13% on several gold standard datasets.