Teaching machines to understand text

Teaching machines to understand – and summarize – text is an article from the The Conversation about the use of machine learning in text summarization. The example they give is how machines could summarize software licenses in ways that would make them more meaningful to us. While these seems a potentially useful application I can’t help wondering why we don’t expect the licensors to summarize their licenses in ways that we can read. Or, barring that, why not make cartoon versions of the agreements like Terms and Conditions.

The issues raised by the use of computers in summarizing texts are many:

  • What is proposed would only work in a constrained situation like licenses where the machine can be trained to classify text following some sort of training set. It is unlikely to surprise you with poetry (not that it is meant to.)
  • The idea is introduced with the ultimate goal of reducing all the exabytes of data that we have to deal with. This is the “too much informationtrope again. The proposed solution doesn’t really deal with the problems that have beguiled us since we started complaining since part of the problem is too much information of unknown types. That is not to say that machine learning doesn’t have a place, but it won’t solve the underlying problem (again.)
  • How would the licensors react if we had tools to digest the text we have to deal with? The licensors will have to think about the legal liability (or advantage) of presenting text we won’t read, but which will be summarized for us. They might chose to be opaque to analytics to force us to read for ourselves.
  • Which raises the question of just what is the problem with too much information? Is it the expectation that we will consume it in some useful way? Is it that we have no time left for just thinking? Is it that we are constantly afraid that someone will have said something important already and we missed it?
  • A wise colleague asked what it would take for something to change us? Are we open to change when we think of too-much-information as something to be handled? Could machine learning become another wall in the interpretative ghetto we build around us?