Recognizing emotions in textual conversations is highly context dependent, as chat messages are often very short and strongly linked to previously exchanged messages. However, the majority of current emotion labeled datasets are excerpts from forums, social media, or literature, most of which do not include dialogues and thus lack contextual information for the emotional states of the chat partners.

We therefore created a German chat dataset by conducting a chat simulation on highly emotional scenarios and evaluated the training on the generated data using different state-of-the-art machine learning models. Our findings indicate, that using context based datasets can improve emotion detection in chat messages. Furthermore, we propose a chat interface, which informs users intuitively about the emotional state of their chat partners by displaying the recognized emotion as an emoji.