- Felipe N. Ducau - [email protected]
- María Elena Villalobos Ponte - [email protected]
We explored the task of building a contextaware, open domain, non-goal driven conversational agent. This problem is non-trivial due to data sparsity and remains one important open question in the field. For this Natural Language Generation task in particular, there is a vast diversity of valid responses to a source utterance, so conditioning on larger context should help produce more on-topic and semantically relevant responses. In sup- port of this goal, we propose an Attention based Encoder-Decoder that depends not only on the previous utterance, but also in the conversational context. Our proposed model performs similar to non-context aware architectures in terms of automatic evaluation, but generates context-aware more human like responses when evaluated by a human.