Evolutionary Combinations of Text Classifiers: Investigating the Diversity of Contexts

Abstract

In this work the problem of text classification will be investigated, which becomes even more complex when the Internet context is considered: countless and ubiquitous computational devices allowing people to create a huge amount of text information stored in the cyberspace. Defining the latter as a virtual environment where information are produced and processed over time, it is also possible to consider them varying in geographical space as well as in the context, giving semantic to the text. The main contribution of this work is the search for text classification models in this dynamic environment. It is proposed an analogy with ecosystems, where texts are represented as habitats and classifiers as individuals that evolve in the dimensions of time, space and context. To demonstrate the initial results of the proposal, an algorithm that evolves combination of classifiers was developed. These classifiers have learned in different contexts and the results demonstrate how learning and evolution in diverse contexts can contribute with significant improvements in text classification tasks.