Notes from "Amplifying Human Cognition: Progress in Language Technologies"

Info

Notes

  1. What IBM Is Now
  • IBM now a cognitive business doing cognitive computing.
  • Using probabilistic models instead of deterministic models for building cognitive systems.
  1. History and Future of Cognitive Computing
  • Not replacing but amplifying human ability.
    • there will always be debates about whether computers will take over humans
  • Reason for going into cognitive era is due to amount of data that is growing so much.
  • In line of a continuous history of digital transformation:
    • Transaction -> Interaction -> Environment -> Reasoning/Learning/Knowledge
  • Need to be socially aware of connecting with people.
  • IBM business units are trying to become a platform (mostly starting from Jan 2014). Due to this, it is not just about runtime but also about tools.
  • Question answering requires much stronger mechanism for matching concepts, such as temporal, geo, etc.
  • word2vec is probably the biggest breakthrough among recent years.
  • For cognitive systems to have proper UIs, they have to be interactive in context.
  • NL context resolution is both important and difficult.
    • Belief tracking methodologies can be used for these.
    • There are structured rules in life, since it would be a waste of energy to do repeatable information processing otherwise. This is also an important concept in neural networks so that they do not compute unchanging beliefs again.
  • Conversational Technical Support is becoming a huge area business-wise.
    • "search of info" -> "interaction with knowledge"
  1. Questions
  • How to deal with ambiguity when solving semantics?
    • Ambiguity is among the nature of language. It can't go away. Meaning is resolved at the time of the conversation.
    • Machines probably won't need syntax at all (except for when communicating with people).
  • Issues of making general dialog system?
    • Conversational systems need domain of knowledge, just as people do. Without one having the knowledge that the other is talking about, both cannot simply perform conversations.
Open Comments