This paper describes a system for aspectbased sentiment analysis (ABSA) using a straight-forward supervised sequence labeling approach. Specifically, we apply a bidirectional, recurrent long short-term memory (biLSTM) architecture with a multilayer perceptron on top that predicts the labels token by token. We deal with the issue of rare words by dynamically switching between character-level and token-level representations depending on an occurrence threshold. A simple encoding of the aspects and their sentiments, a careful preprocessing of the data, and a generous ensemble of 24 single models beats the published state-of-the-art results for the GermEval 2017 ABSA data set for aspect-based sentiment analysis on the document level (joint prediction of aspect and sentiment in task C). For task D, the opinion target expression (OPE) detection task, our approach improves the current state-of-the-art even by 2.7-14.3 percentage points.