Input Conditional Language Models using Long Short Term Memory Networks

Several techniques have been used in the past to generate weather forecast texts using Natural Language Generation. This work presents a novel method using Long Short Term Memory Networks to produce weather forecasts using raw inputs. We also present 4 different variants of our method and show how they can be used as standalone systems to generate text without the need for classical NLG phases. We evaluate our method using several automated metrics and present visualizations for learned models.