TCG refine

[victorbarres]

Deal with semantically empty constructions!

• Det_A cxn should NOT be semantically empty. It does signal that the object is not known or has not be referred to before. This could be encoded in different ways:
  • Added node to the graph (i.e. more symbolic information). The node could be treated through a modify relation ("KNOWN", or "UNKNOWN") [Which should refer to a discourse model or a model of what has already been uttered]
  • Find a way to use the activation values of the nodes to deal with that.
  • NOTE: In a way, in order to properly handle this issue, a discourse model is needed, on top of the semantic model. Like a discourse SemRep...
• There_is cxn (empty locative construction). Those are annoying as they pollute the utterance output since they can be added to any SemRep transduction (being semantically empty but forcing then relative clauses...). Think about what to do with it. Should I keep it? It is important to express a single entity node. But "there is a man who kicks a woman" should not be systematically preferred to "a man kicks a woman". This is where utterance length and other utterance style preference can come into play.

[victorbarres]

Det_A_CXN creates a problem when I used the SPECIFY relation as it matches on both the specified and the specifier. It wouldn't be the case if it could match only on the node that carry the "unknown" tag (ie not yet mentioned in discourse).

[victorbarres]

Could I use a feedback from the phonWM to set up a SemRep Node indicating that an entity has already been mentioned? The activity of this node would decay: High value -> "Pronoun", medium value "The X", low value "A X".

Could be a nice addition...