Extração automática de relações semânticas a partir de dados ruidosos

Abstract

Relationship extraction is a task performed in text-based continuous learning systems, aiming to find semantic relationships between categories or entities. NELL is such a system, which suffers from supervised labeling in its relationship extraction. One of the algorithms attempting to solve this task for NELL is OntExt, but it does not handle noisy input very well, and is computationally expensive. However this algorithm has interesting properties in the context of NELL’s application, not available in other methods. In this work, it is proposed a variant of the algorithm to reduce the impact of its flaws, using a graph-based representation, which is flexible in the handling of outliers. This new method has a comparable precision and higher recall, compared to the existing method. It is also shown an efficient way to represent the problem using sparse structures, reducing the computational cost from minutes to seconds.