Aggregating and visualizing news articles as relationships between entities

Abstract

In this project we present a system with two main sub-objectives. The first, relationship extraction, is performed by detecting named entities in news articles and then identifying the relationship predicate that relates these entities together. This is a process which is performed using deep linguistic analysis, for which we make use of the link grammar formalism. In this way, we are investigating the theory that grammatical patterns can be used to extract semantic relationships. Semantic relations are sought in the form of subject-predicate-object triples. An algorithm for Information Extraction (IE) based on a number of state-of-the-art IE techniques is implemented and used with the text of four Maltese news websites. The extracted semantic relationships are then classified as being valid or invalid using an RBF based SVM model that was trained on 2,000 triples and evaluated on roughly 3,000 triples. U pan further evaluation and analysis of the generated results, the system yielded a precision score of 88% and a recall of 21.45%. Since this value was unsatisfactorily low, we analyzed the error sources and types of errors that were generated in an effort to propose ideas for future work that we think would improve the performance of our system. The second objective explores a visualization technique that can be used to deliver the extracted data to the user. As a prototype, we developed a web application which allows for the browsing of the extracted data, however we also proposed a touch based application which would allow our system to be used on mobile devices. While the development of this application was out of the scope of our project, a number of mock-ups were created in order to describe its main features. For the proposed visualization technique, an evaluation framework that makes use of user acceptance testing is described. This can be used to identify whether or not delivering the news in a visual, relation based manner is as effective as more typical means, such as reading articles on news websites.