Have you heard about the conversational agent, or chatbot, named Meena?

At the end of January, Google AI introduced it. It is a robust model with 2.6 billion parameters. The architecture of the model is based on 13 decoding blocks of the Evolved Transformer seq2seq. The model has been trained on 341 GB of text, primarily conversations from social networks (making it nearly twice the size of GPT-2 and trained on 8.5 times more data). In SSA (Sensibleness and Specificity Average) tests, it clearly outperforms competing chatbots like Mitsuku, Cleverbot, DialoGPT, and Xiaolce, even demonstrating a low perplexity.

The problem with older chatbots is that they excel only in narrowly defined areas, and if you steer the conversation elsewhere, they often struggle. For example, a bot with a wide range of conversational topics would be ideal for language learning or interactive games. Moreover, today's chatbots frequently spout nonsense. They produce texts that contradict what has already been said, lack basic knowledge about the world, and common sense. All too often, they tend to respond with phrases like "I don't know." I can confirm this from personal experience. When I created my chatbot about a year ago as a final project at the Moscow School of Artificial Intelligence (NRUHSE), it too often replied with "I don't know" :)

Future research on Meena will aim to enhance attributes such as personality and realism. The model has not yet been made publicly available, although discussions about this have been ongoing since the end of January.

Blog: https://ai.googleblog.com/…/towards-conversational-agent-th…
Paper: https://arxiv.org/abs/2001.09977¨
Examples: https://github.com/google-re…/google-research/…/master/meena