Detecting Documents with Complaint Character

Recognizing complaint documents as early and as fast as possible is a worthwhile goal for companies. In this paper we present an analysis showing the complexity of this practically relevant problem. Therefore, we define the task and its challenges and investigate statistical methods for automated Complaint Detection in incoming text documents. Two different approaches for handling complaint documents are presented. First, we analyze various term weightings in a standard bag-of-words approach. Second, we show the effect of feature engineering techniques known from Natural Language Processing. The results on four German and one English corpora show that already a linear classifier achieves valuable results and is competitive to more sophisticated methods in most cases.