This is known as Open-domain ChatBot that is not specialized but is the place where the chat conversation can go anywhere, chat about anything a user wants. Although being a fascinating research problem, Google claims its Meena Open-domain ChatBot can handle a wide variety of conversational topics, It’s simply “a Human-like Open-Domain Chatbot. “
“ We present Meena, a multi-turn open-domain chatbot trained end-to-end on data mined and filtered from public domain social media conversations. ”
citation: Towards a Human-like Open-Domain Chatbot – Google
As it’s known for deep neural networks applied to natural language processing tasks, bigger neural networks is not always better. Thus Meena transformer isn’t the biggest in this area.
As it mentioned in “ Towards a Human-like Open-Domain Chatbot ”, Google presents Meena, conversational neural network with 2.6 billion parameters for a model trained with 40 billion words — 341 GB of text data mined from social media “multi-turn conversations” — using the seq2seq model.
It took 30 days of TPU training. That’s 2048 TPU cores comes out to over 1.4 million dollars in compute time!
Credit: Towards a Human-like Open-Domain Chatbot – Google
Google also proposes a human evaluation metric called Sensibleness and Specificity Average (SSA) to score each of Mina’s responses, is a new metric for open-domain chatbots that captures most important and basic attributes for “human-like multi-turn conversation”.
Credit: Towards a Human-like Open-Domain Chatbot – Google
There is two key elements of any human-conversations:
1) Sensibility (making sense).
2) Specificity (being specific).
The network was trained to improve that score of (SSA) as much as possible through gradient descent .
Sensibility alone was not enough because sensible statements can be very vague. that’s why they also optimize specificity given the context of the conversation to discourage vague replies.
The task was to learn the association between these dialogue pairs with an input sequence and an output sequence . The state-of-the-art in sequence to sequence modeling is transformer so they use it.
The most surprise at Meena is its ability to demonstrated the appearance of humor, because humor takes a lot of high-level abstract thinking including the ability to understand the context of a conversation.
Credit: Towards a Human-like Open-Domain Chatbot – Google
Unfortunately, Google says it won’t be releasing the bot to the public community as of now.
until it is sure of its safety and bias .That, however, might be a good thing.
You can read about Meena Bot in more detail in this paper.