Software Scientific: the competitive edge of intelligence.

Technologies

Auto Answer

Auto Answer's capabilities are well suited to automatically responding to SMS messages and emails.

Auto Answer - Present Generation

Capabilities

Auto Answer is an 'engine' designed for integration into a system, and has these capabilities:

All these comments apply to the capabilities of the present version.  It operates by reading an SMS message or email and then generating a reply, along with a confidence level, and an indication of whether or not the message should be reviewed by a human operator.
  1. In the present incarnation, the message is answered by having a set of model answers to the most frequently asked questions, and choosing the most appropriate.
  2. It is easy to 'teach' the system about new topics, since the specification language as to which answer to use is also natural language.
  3. If a query e-mail is about more than one question, then all the questions are answered, not just one.
  4. By taking into consideration the 'mood' of the subscriber (pleased, frustrated, angry, neutral), the message is then 'personalised' to some considerable degree, so that identical replies are rarely generated.
  5. The 'format' of the reply is specified in a template (as are many other parameters).
  6. The system 'knows' how well it has done - that is it also generates a confidence figure.  100% would mean totally sure, and 0% would mean totally unsure.  This can be used to send e-mails about which the system is unsure to a human operator.  This can also be used to balance work load - the human operators get to see the e-mails about which the system is least sure.
  7. The system generates a unique id which it puts in the reply text, and in the reply subject.  If a customer then replies to our reply, hopefully using the 'reply' button, it will be able to detect that this is so and not handle the message automatically, since it is possible that the client was not fully satisfied with the original answer.
  8. There are a number of reasons why the system will consider that an e-mail should be reviewed by a human operator, including:
  9. In addition to generating a reply, Auto Answer can also flag up tags or actions according to the content of the e-mail.  For example, it might be desired to forward e-mails asking about a product to marketing, or to accumulate statistics about the types of questions being asked.

Performance

  1. The quality of the replies generated can never exceed the quality of the model answers (or 'knowledge base').  For a high quality knowledge base, and assuming that many queries are simple, we anticipate about one third being handled fully automatically, one third being passed to an operator but then being immediately ratified, and one third being forwarded to an operator.  (In computer-science language, we put a greater premium on 'precision' than 'recall' - it is more important not to be wrong than it is to handle all e-mails.)
  2. On a typical platform, several e-mails a second can be answered.
  3. The system is multi-lingual.

Extensions

These extensions could be made available very rapidly:
  1. The system can automatically deduce the language of the query, and reply in the same language (provided we have given it the appropriate model answers in each language).
  2. We could reply in a different language to that of the query, but whilst technically clever, I don't think this ability is commercially useful!
  3. Deduce the language of the incoming e-mail for forwarding to an operator who understood that language.
  4. Answer SMS messages.
  5. Include diagrams in replies where they would help.
  6. Drive the system in reverse, and use it to check the accuracy of the human operators replies.  Whilst not totally accurate, this would be able to spot the worst performers.

Deep Technology

Using deeper techniques, we could also:
  1. Answer questions for which we do not have model answers, that is deduce new answers.
  2. Engage in questions with the user, in a conversation.  For example.  User Q: "I can't access System X with my browser."  Automatic Reply: "What browser are you using?"  User Answer: "Netscape"  Our answer: "System X does not work with Netscape.  Change to IE."  To do this, we must have incidents extending beyond the single question-answer pair, so that we can tie in previous e-mails in the conversation.
  3. Use other data clues, such as images.

Further Thoughts...

Paradoxically, a combined AI/human system works better then either on their own.  This is because the AI takes care of the boring routine queries, so the humans know in advance that they are only getting the 'interesting' queries, and therefore give them proper attention, rather than miss-skim-reading them.
 
There is a test for when a computer is genuinely intelligent proposed by the famous mathematician Allan Turing, called the Turing test.  A computer is deemed intelligent when, by conversing with it at arms length (through a terminal or equivalently by e-mail) you cannot distinguish it from a human operator replying with the same technology.  This system will pass that test and therefore may on the basis of this test be deemed the first genuinely intelligent industrial computer system.