Leveraging Text Classification by Co-training with Bidirectional Language Models – A Novel Hybrid Approach and Its Application for a German Bank

Labeling training data constitutes the largest bottleneck for machine learning projects. In particular, text classification via machine learning is widely applied and investigated. Hence, companies have to label a decent amount of texts manually in order to build appropriate text classifiers. Obviously, labeling texts manually is associated with time and expenses. Against this background, research started to develop approaches exploiting the knowledge contained in unlabeled texts by learning sophisticated text representations or labeling some of the texts in an automated manner. However, there is still a lack of integrated approaches, considering both types of approaches to further reduce time and expenses for labeling texts. To address this problem, we propose a new hybrid text classification approach combining recent text representations and automated labeling approaches in an integrated perspective. We demonstrate and evaluate our approach using the case of a German bank where the approach could be applied successfully.