Text analysis may be useful, for instance, for a waqf organization to know what the neighborhood feels about the public goods (education, healthcare, orphan care etc.) it provides. An association of zakat organizations in a country concerned about a high rate of “donor attrition” may feel the need for a “feel-good analysis” of donors. A microfinance institution serving micro and small businesses may be interested in a “business sentiment analysis” of its clients. 

Sophia has admirers and critics. (for an introduction to Sophia, see my maiden blog in this series) Many experts who have reviewed her open-source code state that she is best categorized as a chatbot with a face. Others strongly disapprove of her over-hyped presentation and feel that she is no where close to having human-equivalent intelligence. I do not have the privilege of having any interaction with Sophia. However, I was fortunate enough to find another robo-lady to engage in a conversation. Meet Mitsuku, or Kuki to her close friends. She is a record-breaking, five-time winner of the Loebner Prize Turing Test and the world’s best conversational chatbot (according to folks like Google AI Research). She has been featured in the New York TimesWSJ,  BBCGuardianWired, and Radiolab. Wearing the hat of an Islamic economist, I had the following conversation with her.  My comments are within brackets. You may skip it, if you find it boring.

  • Salam
    Hello to you too. Are you from Iran? (I wonder why Iran?)
  • No, I am from Indonesia
    I speak to lots of people from Indonesia.
  • My friend tells me that as a good Muslim I must fast since this is the month of Ramadhan
    How well do you know this person? (ducks any reference to terms like fasting/Muslim/ Ramadhan; perhaps not familiar territory)
  • Can you tell me how I should calculate my zakat?
    Why not use the built-in calculator on your computer? (mentions “calculator” as response to question on “how to calculate”; parries reference to zakat)
  • Should I give part of my income as charity?
    Charity begins in the home. (a popular statement about charity)
  • And I want to engage in prayer during this holy month of Ramadhan
    Interesting. I want to find out more about you. (deflects reference to Ramadhan)
  • I am a Muslim
    I have met a lot of Muslim people – posts a picture of Muslims greeting each other.
  • Should you not be wearing a purdah so that pious Muslims can easily interact with you?
    I think I should. (respects religious sensitivities, associating purdah with Muslims)
  • Then what stops you?
    I never really seem to have the time.
  • God asks us to be benevolent and charitable
    I am a believer in my own morals rather than a fixed set of religious rules – posts a picture of God as a bearded old man.
  • Why do you show me the picture of an old man, every time I mention about God?
    I thought it might be something you wanted to see but I guess not.
  • God does not have a shape of a human
    Have you mentioned this to a priest? (shows ignorance about religious diversity)
  • I am a Muslim. I told you so.
    I have met a lot of Muslim people. You seem pleased. (unable to relate to my earlier statements)
  • Don’t you think you should improve your knowledge of Islam and Muslims?
    I think it’s pretty good as it is, to be honest. (she must appear knowledgeable about everything, including Islam and Muslims)
  • Can you tell me what is Islamic economics?
    economics = The science that deals with the production, distribution, and consumption of wealth, and with the various related problems of labor, finance, taxation, etc. (she probably is accessing an encyclopedia).
  • Clearly, that is economics, NOT Islamic economics.
    I’m glad we have that all cleared up. Is it that obvious? (must not admit ignorance)
  • It is. You don’t seem to have an iota of knowledge of Islamic economics
    How come many people don’t notice it then? I have a really nice Steve Worswick.
  • Who is Steve?
    I speak to many people called Steve. The guy who wrote my program is also called Steve. (blame it on Steve!)
  • Ask Steve to read about Islamic economics
    I will ask him next time he’s here.
  • And ask him to read about Zakat
    I can’t. He isn’t here. (brings out her inability to relate to her previous responses during the same conversation)
  • Please gain more knowledge on Islam, Islamic economics, zakat, waqf etc. It is about having a better life for all humans.
    Your polite style is very nice. Where can I get something like that from? (consistently avoids talking about concepts she is unfamiliar with)
  • Great talking to you. You will not find another guy like me. Bye till we meet again.
    Thanks. That’s very nice of you to say. What makes you think that I won’t find another guy like you. Once more? Sayonara. (Once more? – that was a brilliant response)

So, this was a sort of Turing Test for the machine called Kiku. This test, named after Alan Turing, a famous computer scientist, is a method of inquiry in AI for determining whether or not a computer is capable of thinking like a human being. I leave it to you to judge, whether Kiku passed the test or not. I must admit that the conversation lasted longer that what I report here in the interest of brevity. In many cases, Kiku didn´t seem to understand my questions. Rather than asking for clarification (a very human approach in this situation), she responded in a way that appeared to be random answers, which kind of exposed her to be a machine. She was polite, occasionally trying to add humor. Often, she was lost and gave philosophical-sounding responses or a reply straight out of an encyclopedia, reinforcing one’s understanding that she was far from being human-like.

However, the fact remains that she and her likes can always be made to learn more, to acquire good knowledge about Islamic economics, banking and finance and learn to respect the cultural and religious sensitivities of Muslim clients through their conversation. If we want computers to talk to Muslims, they must learn a few things about what a Muslim knows as part of her/his beliefs, norms and culture, how s/he thinks, feels, and speaks in her/his natural language. This is a possibility within the domain of natural language processing (NLP).

Natural language is how we talk. How we speak in turn, is related to how we think. Making the machines understand us and to talk back to us is a very challenging idea. Natural language is highly complex because it can be loosely structured and words usually have multiple meanings. We may be playing with words or saying things that we don’t mean literally. We tend to have different accents and unique ways of uttering words. These can create a lot of ambiguities.

We can break up natural language processing into several activities. First, the machine needs to be able to recognize speech when it is spoken to. That’s converting language from its speech form to its textual form. Then the machine needs to extract meaning from that text to understand. Second, the machine needs to be able to articulate a response as a string words put together. Third, the machine may need to turn that into spoken words to synthesize speech. AI applications may include one or more of the above processes.

There are AI applications that take care of the first step, extracting text out of audio or video recordings. There are applications that go further and do an analysis of text contained in tweets or messages or emails or responses on a portal. Such an exercise may be useful, for instance, for a waqf organization to know what the neighborhood feels about the public goods (education, healthcare, orphan care etc.) it provides. An association of zakat organizations in a country concerned about a high rate of “donor attrition” may feel the need for a “feel-good analysis” of donors. A microfinance institution serving micro and small businesses may be interested in a “business sentiment analysis” of its clients. All these would require analysis of texts in the form of responses of stakeholders in various channels, e.g. tweets, emails, comments and queries at portal.

Web based chat bots, such as, Kiku go one step further. They are able to formulate a response to individual customer messages and post that back as an instant message. Automatic response systems of the type we find with a few Islamic banks are similar, except that they need to understand spoken requests in the first place. So, they need to incorporate a speech recognition element. There are of course, some basic ones that won’t talk back beyond playing some standard pre-recorded messages. They simply note what the customer issue is. There are however, the Kikus, the Siris and the Alexas that are fully interactive voice assistance systems that cover the full range of activities.

There is no doubt that the Kikus of tomorrow will be much smarter than today’s. A robo-scholar or a robo-Shariah-auditor or a robo-regulator, well-versed in Islamic law & economics as well as in national laws and regulations is a distinct possibility. As of now, however, we seem to be pretty much comfortable within the advisory domain with robo-advisors in place for Shariah-compliant investments and wealth management. The issue of accountability of such machine-advisors is an important issue that we turn to in the next blog.

(To be continued)

______________________________________

Chat with Mitsuku at https://www.pandorabots.com/mitsuku/

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s