"Natural language generation is one of the most active research areas within NLP and its barrier of entry has reduced dramatically. While applying supervised state of the art models to new data sets is becoming easier, the evaluation of models is becoming more challenging as models can produce completely fluent but meaningless or subtly flawed output. This leads to a disconnect between real-world needs of generation models and published research. Most of the disconnect can be bridged via in-depth evaluation and documentation of both data and models. To that end, the GEM workshop has three core goals: (1) Encourage the development of (semi-) automatic model audits and improved human evaluation strategies, (2) Popularize model evaluations in languages beyond English, (3) Provide a platform for discussions around evaluations to bridge the gap between industry and academia."