[Reading] EM-ONE: Chapter 1,2, 5

Chapter 1:

1. EM-ONE uses as its commonsense knowledge base a library of commonsense narratives, each a story describing a fragment of the physical, social, and mental activity that occurs during a particular interaction between two actors.
2. Mental critics are implemented as pattern matching procedures that solve problems by case-based reasoning using a library of narratives cases.
3. Critics recognize problems by matching patterns encoded in a frame-based knowledge representation language that supports the description of structured scenarios involving many connected actors, situations, events, objects, and properties, including mental relations such as "observes," "believes," and "desires."
   

Chapter 2:

1. Narratives connect knowledge to purpose
2. Narratives help us control inference => it tells you when to stop inferencing
   
3. Narratives are easy to acquire => just like OMCS easier to build than CYC
   
4. Narratives can contextualize knowledge => it gives you the surrounding context
   
5. ...Unlike natural language, there is no syntatic ambiguity [in EM-ONE Narratives], so this notation can be easily parsed.
   

So I think the topic for Society of Mind's term project could be as follows. I think I'm interested in understanding (and understanding the similarity/difference between two) stories, so I'd like to look more deeply into the knowledge representation in Push's EM-ONE and Schank's Conceptual Dependency, since they are used to represent the memory, the plans, the goals, and everything. Specifically:

• How easy is it to translate knowledge from natural language into their representaions?
• How well do they help to find the similarity/difference within different stories?
  
• How well do they facilitate the practice of applying the knowledge gathered from the Opend Mind website?
• How well do they suit the layered architecture in Emotion Machine?
  
• How well do they exhibit the benefits brought by narratives?

Chapter 5:

Then I found from this chapter that there actually isn't that much difference between the two knowledge representations, because Push's work was mainly focusing on how the mind works, in terms of making decisions at different levels for physical, social, and mental realms. Schank's stuff, on the other hand, focuses on the way of representing narration without using natural language. The ACTs that Push used include "is-touching", "is-holding", "at-location", "has-speed", "moves-to", "looks-at", "grasps", "releases", "attaches", "lifts", and so on, in which some belongs to the "sensor frame" category, and the others go to the "behavior frame" category. And, he used "observes", "does", "desires", and "believes", that utilize these predicates and make mental states for critics. The major difference between these two representations, I would say, is the relationships between event descriptions within a episode. Push used simply sequential orders to characterize causality, whereas Schank used 7 or more ways to differentiate links that exhibit different meanings or even functions.

If I can make a dicision about what we should exploit, I would say, we should make some network similar to ConceptNet, only that each of the nodes represent a sentence or narrative. I think it's nice way to represent all the things we know, because the innate characteristics of a network helps to make it easier to do analogy and spreading activation, and it's easier to build using OMCS stuffs that are so fragmented. If we can make this kind of semantic network that uses a static set of vocabulary without any ambiguity, I think EM-ONE might be able to use it, and maybe other systems can use it as well.