Relationship between detected events in online media

Abstract

Social media allow users read, post, and share information about real-world events. There are several techniques to detect or discover real-world events from online news or posts. However, a problem that is often overlooked in this context corresponds to discovering relationships between events. A particular kind of relationship is the similarity between two events, which can be used to organize and filter the flow of information provided to users. This thesis presents an approach to identify similarity relationships between events previously detected in short texts from online media. Thus, it proposes the Autoencoder Neural Event Model (AutoNEM ), an autoencoder-based unsupervised neural network model to discover similarity relations between events structured according to the 5W1H representation standard. This model meets the combination of a set of requirements that have not been satisfied in a single approach in the literature. AutoNEM can encode events in latent space, including each 5W1H -attribute separately, which allows the search for similarity relationships between events through their embeddings. The experiments use data collected from the news corpus EventRegistry to validate the proposed approach. The experimental evaluation indicates that proposed neural model for detecting similarity relationships is effective, and by comparing with some baselines is competitive too. The experiments also evidence some degree of similarity in other pairs of events that had not been evidenced by the manual curators of EventRegistry