Key-Value Retrieval Network

Welcome! Below is a demo of the Seq2Seq network described in Key-Value Retrieval Networks for Task-Oriented Dialogue. This 2017 paper describes an end-to-end differentiable Seq2Seq network that is able to interact with an external knowledge base with no explicit representation of dialogue state. Additionally, the paper presents a new multi-turn, multi-domain, task-oriented dialogue dataset.

Check out my code for the API and model training here.

What is "Multi-Turn, Multi-Domain, Task-Oriented Dialogue"?

That's a lot of hyphens. Let's break it down:

• Multi-Turn:

  Multiple utterances take place between the user and the system. A single turn would be something like "How are you?", "I am good!".

• Multi-Domain:

  The dialogues occur in multiple domains, each with different knowledge base schemas. Here, we are concerned with the driving, weather, and calendar domains.

• Task-Oriented:

  There is a specific goal that the system is trying to achieve. This could be retrieving calendar information ("What time is my doctor's appointment?"), weather information retrieval ("What's the weather in Boston?") or navigation ("Take me to a coffee shop").

More info on the KVRET dataset of 3,031 multi-turn dialogues can be found here .

How do I use this demo?

To interact with the Seq2Seq chatbot, type a question into the text box below, and hit 'Enter'. Or, to sample dialogue from the KVRET test and dev data, hit 'Get Example Dialogue'. To clear the chat and the model's accumulated context, hit 'Clear Chat'. For best results, ask about specific values in the knowledge base table.