Building a Conversational Agent Overnight with Dialogue Self-Play

+ Perfect control of agent behavior + guaranteed scope of critical interactions - few recalls at unexpected interactions - agents do not learn from mistakes

+ Understand user / population interaction + Flexible agent that can be improved with more data - data collection and annotation are expensive - little control over agent behavior

Wang Yushi, Jonathan Berant, Percy Liang. "Construct a semantic parser overnight." The minutes of the 53rd Annual General Meeting of Computational Linguistics. 2015

Victor Zhong, Caiming Xiong, and Richard Socher. "Seq 2 SQL: Generating structured queries from natural language using reinforcement learning." ArXiv preprint arXiv: 1709.00103 (2017)

Pararth Shah, Dilek Hakkani-T × 1 / 4r, Gokhan Tur, Abhinav Rastogi, Ankur Bapna, Neha Nayak, Larry Heck. "Please build a dialog agent through dialogue and self game." ArXiv preprint arXiv: 1801.04871 (2018)

Pararth Shah, Dilek Hakkani-T × 1 / 4r, Gokhan Tur, Abhinav Rastogi, Ankur Bapna, Neha Nayak, Larry Heck. "Please build a dialog agent through dialogue and self game." ArXiv preprint arXiv: 1801.04871 (2018)

Liu Bin, Gokhan Tur, Dilek Hakkani-Tur, Pararth Shah, Larry Heck. "End-to-end optimization of task-oriented conversation model using deep reinforcement learning" Session AI Workshop, NIPS 2017

Liu Bin, Gokhan Tur, Dilek Hakkani-Tur, Pararth Shah, Larry Heck. "End-to-end optimization of task-oriented conversation model using deep reinforcement learning" Session AI Workshop, NIPS 2017

[1] Ankur Bapna, Gokhan Tur, Dilek Hakkani-Tur, Larry Heck. "Sequential Conversation Context Modeling for Spoken Language Understanding" The 103th Annual Scientific Meeting Record on Discourse and Dialogue 103-114. 2017

Pararth Shah, Dilek Hakkani-T × 1 / 4r, Larry Heck. "Interaction Reinforcement Learning for Task-Oriented Dialogue Management" Deep learning action and interactive seminar, NIPS 2016

To build a complete session proxy, Microsoft provides a bot framework. In the Bot framework, you can write loosely coupled scripts using Node.js or C #. In Node.js, business logic and session flow are described in callbacks listening for events, and when LUIS recognizes intents and entities, it issues related events and executes callbacks. Many things need to be managed by developers, but this is the best compromise between flexibility and not the need to build a framework. API training is organized around stories (domain specific use cases). Here, the engine learns the flow of conversation from examples of user input and robot response. Because the SaaS engine does not provide action support, external services must be called outside of the platform. Contexts are not explicitly managed by Bot Engine, but they can be passed as developers by JSON objects.

KITT.AI builds its own ChatFlow platform and enables users to create conversational or intelligent robots while visually describing conversations using a simple drag and drop interface while allowing processes that can be executed on the server Can be implemented. Dialogue design This platform features machine learning models for hotword detection (no internet required), semantic analysis, natural language understanding, session engine (multi-turn support), and neural network driven. ChatFlow supports Alexa, Facebook Messenger, Kik, Skype, Slack, Telegram, Twilio. This platform is currently free and you can sign up for the beta version. In the future, the team will provide free or cheap access to individual developers and small teams, and will provide higher cost to corporate customers.