Jan 29, 2020 23:00:00

Google develops AI `` Meena '' that can talk normally like a human

AI that can talk to people like Apple's

Siri and Amazon's Alexa can answer people's questions properly and sometimes even joke. However, when conducting a 'conversation' rather than a question-and-answer exchange, it is also true that people often cannot communicate well unless they choose words in consideration of AI. Meanwhile, Google has developed AI that can conduct `` natural conversation '' using deep learning , and publishes comparison results and actual conversation examples with other chatbots

[2001.09977] Towards a Human-like Open-Domain Chatbot
https://arxiv.org/abs/2001.09977

Google AI Blog: Towards a Conversational Agent that Can Chat About… Anything
https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html

Meena is Google's attempt at making true conversational AI | VentureBeat
https://venturebeat.com/2020/01/28/meena-is-googles-attempt-at-making-true-conversational-ai/

The AI 'Meena' developed by Google uses ' Evolved Transformer seq2seq ', which is one of the methods to generate sentences by neural networks . The input conversation is encoded into a number, which is converted into a conversation by 13 decoders. By implementing many decoders, the quality of conversation has been raised. A 341GB text file containing conversations collected from social media was used for learning.

Along with the development of Meena this time, Google has designed a test 'Sensibleness and Specificity Average (SSA)' to know the performance of chatbots 'speaking normally', and Meena was tested by SSA. . In SSA, humans actually talk with chatbots and evaluate 'whether the story is reasonable' and 'whether the answer is limited'. 'Limited answer' here is not an answer that applies to any topic, such as saying 'I also like tennis' to 'I like tennis', but 'I am also.

Roger Federer can't be Roger Federer ! '

The following are the results of implementing SSA on people, Meena, Mitsuku , Cleverbot , DialoGPT , and Xiaoice . You can see that Meena performs better than other chatbots and is closer to human results.

In addition, the actual conversation content is published as a text file on GitHub, and you can read the conversation content of people, Meena, Mitsuku, Xiaoice.

google-research / meena at master · google-research / google-research · GitHub

https://github.com/google-research/google-research/tree/master/meena

This is part of the conversation between Mitsuku and people. In response to Mitsuku's answer to 'I just mentioned it,' people continue to say 'I don't know what that story is,' and I don't know the point.

On the other hand, this is what Meena and people talked about. Meena responded to the question 'What do you think is the title of the dissertation?', 'How to breathe underwater', and conversation without discomfort continues.

In this AI development, Google focused on the rationality of the story and the limitedness of the answer, but in the future development also emphasize individuality and factuality. In addition, there is a study that AI learns female discrimination and racism from human language , and it is said that it will also study the safety and prejudice of chatbots.