Evolving network structures for text classification using genetic algorithms

Hayden Andersen, Xiaoying Gao, Bing Xue, Mengjie Zhang

Abstract: Convolutional Neural Networks (CNNs) have shown themselves in recent years to be strong contenders for text classification and sentiment analysis, achieving high accuracies challenging other state of the art methods. However, their architectures need to be manually designed and the parameters need to be manually tuned. They suffer from requiring good domain knowledge to design the architecture for each individual problem, which often isn't possible and can increase operating costs for companies wishing to implement such methods. In this paper, two different methods are proposed that use Genetic Algorithms to automatically evolve CNNs without requiring any human intervention. The experimental results show that one of the proposed methods rivals results obtained from other CNN based methods based on classification accuracy, reaching high accuracy on the IMDB baseline dataset while requiring only a few hours training.