When You Should(n’t) Use Deep Learning for Text

• In Event: Machine Learning and Text Analysis

Abstract

Deep learning methods, a family of complex parameterized functions structured as a multi-layered neural network, have existed for over two decades and have been used primarily for image analysis. In recent years, these methods have been applied in a number of studies which use image, video and more recently, text data, to learn about a variety of political phenomena ranging from protests to political ideology. While deep learning methods are appealing because their ability to handle a wide variety of data with impressive accuracy rates, they have “black box” properties which include lack of transparency and lack of interpretability which are not offset by the marginal performance benefits they confer when compared with “traditional” machine learning methods. In this paper, I compare the performance, interpretability and transparency of deep learning methods such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) with that of traditional machine learning algorithms for classification of a number of canonical political texts. Through these analyses, I provide a series of general guidelines, promises and pitfalls for researchers thinking about using deep learning in their research and, where relevant, propose alternatives to deep learning.

Author

• Jason Anastasopoulos, University of Georgia