Chatbots have been around for a long time but more than a few minutes of interaction proves they’re not really up to scratch unless used for one narrow purpose (eg a talking FAQ for an insurance website). I’ve sat in on a Loebner Prize session where the most humanlike chatbot was awarded for “passing” the Turing test and they weren’t overly impressive.
Now, Google says it has created a chatbot, using a neural conversational model, that can achieve near-human conversational abilities: "Meena can conduct conversations that are more sensible and specific than existing state-of-the-art chatbots," say Google Research scientists Daniel Adiwardana and Thang Luong.
Very complex
Meena sounds very complex: it has 2.6 billion parameters (these are language-specific such as subject placement) and is trained on 341 GB of text taken from public domain social media conversations. Hopefully this doesn’t include too many politically polarised dust-ups on Twitter.
The chatbot was tested against others on two counts. First, for sensibleness, ie making sure it wasn't talking nonsense. Secondly for specificity, so asking the chatbot: "Do you like tennis?" might elicit a vague reply like "Yes, it's nice", or in the case of Meena, it would reply: "Me too, I can't get enough of Roger Federer! "
Google says Meena has applications including further humanising computer interactions, improving foreign language learning, and making more realistic and enjoyable interactive video game characters.
[ https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.htmlOpens in new window ]
