A Copy-Augmented Sequence-to-Sequence Architecture Gives Good Performance on Task-Oriented Dialogue

Abstract

Task-oriented dialogue focuses on conversational agents that participate in dialogues with user goals on domain-specific topics. In contrast to chatbots, which simply seek to sustain open-ended meaningful discourse, existing task-oriented agents usually explicitly model user intent and belief states. This paper examines bypassing such an explicit representation by depending on a latent neural embedding of state and learning selective attention to dialogue history together with copying to incorporate relevant prior context. We complement recent work by showing the effectiveness of simple sequence-to-sequence neural architectures with a copy mechanism. Our model outperforms more complex memory-augmented models by 7% in per-response generation and is on par with the current state-of-the-art on DSTC2, a real-world task-oriented dialogue dataset.

Self-Attentional Models Application in Task-Oriented Dialogue Generation Systems

“…Per-turn Accuracy: Per-turn accuracy measures the similarity of the system generated response versus the target response. Eric and Manning (2017a) used this metric to evaluate their systems in which they considered their response to be correct if all tokens in the system generated response matched the corresponding token in the target response. This metric is a little bit harsh, and the results may be low since all the tokens in the generated response have to be exactly in the same position as in the target response.…”

End-to-End Learning of Task-Oriented Dialogs

“…They used an RNN to encode the dialog state and applied end-to-end memory networks to learn it. In the same line of research, people explored using query-regression networks (Seo et al, 2016), gated memory networks (Liu and Perez, 2017), and copy-augmented networks (Eric and Manning, 2017) to learn the dialog state RNN. Similar to (Wen et al, 2017), these systems are trained on fixed sets of simulated and human-machine dialog corpora, and thus are not capable to learn interactively from users.…”

Generative Encoder-Decoder Models for Task-Oriented Spoken Dialog Systems with Chatting Capability

“…The third branch strives to solve both problems at the same time by building an end-to-end model that maps an observable dialog history directly to the word sequences of the system's response. By using an encoder-decoder model, it has been successfully applied to open-domain conversational models (Serban et al, 2015;Li et al, 2015Zhao et al, 2017), as well as to task oriented systems (Bordes and Weston, 2016;Yang et al, 2016;Eric and Manning, 2017). In order to better predict the next correct system action, this branch has focused on investigating various neural network architectures to improve the machine's ability to reason over user input and model longterm dialog context.…”