Deep Learning-Based Classification of Customer Communications of a German Utility Company

The Germany utility company, Zweckverband Ostholstein (ZVO), receives each year a large number of customer communications that are read and classified by its employees. To automate the processing of customer communications, University of Lübeck and ZVO collaborate to model the classification problem of customer communications over a challenging real-world dataset that is multi-label, imbalanced, extremely varied in length and full of noise. To find an optimal model, we first identify suitable classification algorithms and text encoding techniques and then develop 12 deep learning models that combine the different capabilities of neural networks with classical feature-extraction methods and modern word embedding techniques. These models are extensively tested and evaluated, and the results of evaluation are in-depth analyzed and discussed. Finally, we not only find an optimal model for the challenging dataset but also obtain interesting findings and insights that we believe will benefit other deep learning-based text classification projects.