Alibaba – Domain game

AliMe is no Google Assistant beater.

• Alibaba is making strides to develop its AI, and its voice assistant in particular, but despite a very impressive demonstration, I think it remains very far away from what Google Assistant can achieve.
• At the 2018 Neural Information Processing Systems conference, Alibaba shared more details of its AliMe speech recognition and natural language processing engine which has been adapted for a number of tasks within Alibaba.
• One of this is within Cainiao logistics, a subsidiary of Alibaba which is responsible for delivering around 1bn packages per day.
• Other include a food ordering agent and a chatbot that aids price negotiations in Alibaba’s resale platform called Xianyu.
• Alibaba gave an impressive demonstration of a service call to Cainaio logistics that demonstrated three speech features that machines have historically struggled with.
• These are:
  • First, non-linear conversation: This is when the human participant changes the direction of the conversation with no warning.
  • Cainiao’s assistant was able to deal with this and then return to the business in hand.
  • This is a necessary feature but not something that other assistants are incapable of dealing with.
  • Second, implicit intent: In Alibaba’s demonstration, the human says that he is not home in the mornings rather than saying that a morning delivery won’t work.
  • The AI is able to deduce that the delivery won’t work and then offers other options.
  • This is one of the most difficult features for machines (see here) to deal with and I don’t think that this example, while functional. is a very good example of this critical feature.
  • Third, interruption: This is where the human interrupts the machine and the machine can deal with it smoothly and easily.
• Alibaba claims that this assistant is already live and dealing with millions of calls every day which will give it plenty of data from which to learn and improve.
• However, it does raise the question as to why Alibaba felt the need to use a demonstration call rather than a real one to make its point.
• If the assistant is indeed making millions of calls every day, then I am pretty certain that there will be plenty of instances where the three features that Alibaba wanted to demonstrate would be present.
• I very much doubt that this has been done for privacy reasons as Chinese users are far more relaxed about their privacy compared to users in developed markets.
• Hence, I question whether this assistant is really as good as Alibaba would have us believe.
• Google, on the other hand, has been much up front with Duplex and has shown numerous examples of the assistant making live calls to book tables at restaurants.
• Hence, while AliMe is clearly making progress, I think it still has a very long way to go before it can challenge Google’s leadership in AI.
• Furthermore, it is worth noting how these assistants are trained.
• They are not simply let loose but are programmed in functionality domains.
• This means that when one creates an assistant, one decides what it is most likely to be used for and ensures that it can handle as many possibilities within that function as possible.
• This is why AliMe and Duplex seem to be much cleverer than they actually are.
• Both of them are functioning within a very narrow domain (restaurant bookings and package delivery) which greatly simplifies the process of programming the bots.
• I suspect that if one asked AliMe what the delivery charge would be in US dollars or Duplex whether the user would like to sit outside based on the weather forecast, both would quickly fall over.
• Hence, I see these as steps forward in the ability of machines to understand humans and speech reproduction but not much else.
• The days of having a real conversation with a machine remain very far away ensuring that for now, humans still need to talk to each other.