CATI: An Active Learning System for Event Detection on Mibroblogs’ Large Datasets

Gabriela Bosetti, Előd Egyed-Zsigmond, Lucas Ono


Today, there are plenty of tools and techniques to perform text- or image-based classification of large datasets, targeting different levels of user expertise and abstraction. Specialists usually collaborate in projects by creating ground truth datasets and do not always have deep knowledge in Information Retrieval. This article presents a full platform for assisted binary classification of very large textual and text and image composed documents. Our goal is to enable human users to classify collections of several hundred thousand documents in an assisted way, within a humanly acceptable number of clicks. We propose a graphical user interface, based on several classification assistants: text- and image-based event detection, Active Learning (AL), search engine and rich visual metaphors to visualize the results. We also propose a novel query strategy in the context of Active Learning, considering the top unlabeled bi-grams and duplicated (e.g. re-tweeted) content in the target corpus to classify. These contributions are supported not only by a tool whose code is freely accessible but also by an evaluation of the impact of using the aforementioned methods on the number of clicks needed to reach a stable level of accuracy.


Paper Citation