End-to-end trainable system for enhancing diversity in natural language generation
Jan Deriu, Mark Cieliebak · Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences) · 2017
Natural Language Generation plays an important role in the domain of dialogue systems as it determines how the users perceive the system. Recently, deep-learning based systems have been proposed to tackle this task, as they generalize better and do not require large amounts of manual effort to implement them for new domains. However, deep learning systems usually produce monotonous sounding texts. In this work, we present our system for Natural Language Generation where we control the first word of the surface realization. We show that with this simple control mechanism it is possible to increase the lexical variability and the complexity of the generated texts. For this, we apply a character-based version of the Semantically Controlled Long Short-term Memory Network (SC-LSTM), and apply its specialized cell to control the first word generated by the system. To ensure that the surface manipulation does not produce semantically incoherent texts we apply a semantic control component, which we also use for reranking purposes. We show that our model is capable of generating texts that are more sophisticated while decreasing the number of semantic errors made during the generation.