End-to-end trainable goal-oriented conversational agent. Agent is trained to keep a distributed representation of the dialog state on multiple temporal scales. This gives the agent ability to have contextual conversation as well as optimize long term goals.
Agent trained on Frames dataset. Agent is pre-trained using maximum likelihood and then actor-critic algorithm is employed to maximize BLEU reward.
> Hey! I'm looking for a flight to toronto.
Agent (speak, 20.74): from which city would you like to depart from ? <EOS>
> I'm from calgary.
Agent (speak, 42.89): and when would you like to travel ? <EOS>
> August.
Query: {'dst_city': 'toronto', 'or_city': 'calgary', 'str_date': 'august'}
NO RESULT
Agent (speak, 41.67): there are no results from calgary to toronto . <EOS>
> What If I were to go from vancouver on august 26?
Query: {'dst_city': 'toronto', 'or_city': 'vancouver', 'str_date': 'august 26'}
NO RESULT
Agent (speak, 34.32): there 's nothing available from vancouver to toronto . <EOS>
> Okay. Thans.
Agent (end, 9.99): my apologies . have a nice day ! <EOS>