Building better language models to assist humans in various aspects of life continues to be an important domain for big tech firms. In fact, there have been so many advancements in this area that even Google engineers have started believing that the company's AI has become sentient. Meanwhile, Amazon is busy trying to make Alexa mimic the voice of your dead relative.
Now, Microsoft has announced that it is open-sourcing its Grounded Open Dialogue Language Model (GODEL) for use by a wider audience.
GODEL combines two important functionalities in a single language model. The first is being task-oriented and the second is to still make the conversation a realistic and social one. Typically, chatbots try to either be goal-driven - like support assistants - or they just attempt to be social without a purpose.
GODEL is a class of pre-trained models that basically allow a dialog agent such as a chatbot to provide answers based on the current conversation while also generating external and related information to make your interaction more realistic and grounded. An example provided by Microsoft is a conversation about restaurant recommendations branching out to discuss recipes, cooking techniques, and more.
Microsoft wants GODEL to be unrestricted in terms of the queries it can answer and expand upon, while still making the overall conversation useful.
GODEL is actually an evolution of DialoGPT, which Microsoft Research released in 2019. Microsoft's automated- and human-based testing has shown its latest conversational AI to be far superior to DialoGPT, which is why the firm believes that it is ready for broader use in diverse environments. Below, you can see an example of a conversation with GODEL in which it responds to a question about a current event that was not a part of its training set:
GODEL is now available as an open-source repository on GitHub here and you can also check out the associated research paper here. It is offered in three variants along with the code you would need to re-train GODEL on a specific set of information.