It all started with a dream
Imagine that in the future, AI assistants will understand us and speak the same language with us. Imagine that they will learn and teach us. Imagine that they will become our true friends. They will do whatever we want ...
Today we are ready to share the first version of such an assistant. DeepPavlov DREAM is an AI assistant based on a social bot created by the laboratory team for the Alexa Prize Socialbot Grand Challenge 3 from Amazon . The bot is currently available on our demo site and in the Telegram messengerand is ready to communicate, but only in English. DREAM combines about 40 different communication skills to allow open dialogue on a variety of topics. The helper is based on modern NLP models and components, including 14 annotators, 4 post-annotators and knowledge graph integration from WikiData. And yes, do not forget to communicate with DREAM!
PS: Share your feedback with us on how to improve it (and yes, it's still early stage , so don't be too strict).
DREAM is built on top of DeepPavlov Agent , an open source dialog orchestrator. This orchestrator supports dialogue management across several different types of skills. In our Medium blog, we have prepared a simple example that can implement a simple DP Agent configuration using just one custom skill and built-in skill and response selectors. Read it to learn how to build AI assistants using DeepPavlov technology.
But if you want to dive deeper into the details of the DREAM bot architecture, check out the DREAM team's technical report for the Alexa Prize Socialbot Grand Challenge 3.
Why use DeepPavlov DREAM?
Developing a bot for a specific task requires setting up NLU (natural language understanding) and managing dialogs. At first, everything seems simple and straightforward, but soon many functions are added to NLU that can cover all the variety of both what the user can say and the scenarios needed to solve the problems of the users. More and more scripts / actions / models are being integrated to expand the functionality and value of the product. More scripts require more features and the whole thing turns into an advanced AI helper. This helper can reach a certain functional maturity, but it usually reaches a complexity ceiling when dependencies between components severely limit the helper's ability to refactor and further develop.This problem is especially acute in mission-critical applications with multiple domains, as even small changes to the system can lead to completely unpredictable behavior.
, ?
DeepPavlov addresses this problem by introducing a modular architecture for conversational agents. In the DP Agent framework, functionality for various tasks is packaged into separate conversational skills with clear interfaces. This additional level of abstraction allows new skills to be added without interfering with existing system components. Dialogue is controlled both at the level of individual skills and at the level of dialogue as a whole. The MVP for a new task can be tested in isolation before being integrated into the AI assistant.
However, many skills, such as chatting, being able to set an alarm or keeping a calendar, etc., are the same for most use cases and are already being actively pursued. Then why reinvent the wheel? But DP Agentallows you to create a distribution kit of an interactive agent that includes a set of basic basic skills. That is, by default it provides basic functionality out of the box, and then can be expanded by connecting special skills.
The open architecture of the DP Agent is especially effective for creating and maintaining complex conversational solutions. The integration of skills in the form of microservices makes the assistant highly scalable. And skill development and support can be effectively done by a group of enthusiasts, an entire product team, or subcontractors.
DREAMIs the first multifunctional and experimental AI assistant built using DeepPavlov's interactive product stack. In the coming months, we will begin uploading the source code for DREAM skills, services and tools. We will also continue to talk in our blog about how to create more complex AI assistants using DREAM.
The DeepPavlov DREAM roadmap includes the publication of examples of configs, tools for designing multiskill assistants and managing dependencies. We plan to support rich offerings, smooth form filling, context tracking, user learning, and many more scenarios in the next updates. You can read the roadmap in the blog .
How can I contribute to the development of AI?
AI is the frontier of technological progress, and open source is one of its main driving forces. Today we are pleased to tell you about our DeepPavlov Contributor Program. This is a fantastic opportunity to join DeepPavlov on our incredible adventure towards our big dream - creating AI assistants who can understand us, teach us, learn from us and help us become better.
We already have some inspiring stories from members of our DP library and we invite you to learn more about the program here .
We will continue to post updates on our DP blog, and if you want to receive news from us promptly, then subscribe to our newsletter .
And finally
- DeepPavlov DREAM Socialbot (web, Telegram) ( !).
- , - DeepPavlov.
- DeepPavlov.
- .
- , GitHub: DeepPavlov DeepPavlov Agent.