Unifying Models of Cognition - Rens Bod

What do different forms of cognition, such as language and music, have in
common? They involve perception, which groups parts in a hierarchical
structure. The result is compositional and recursive. 

It is relatively uncontroversial to suppose that one representation formalism
underlies all modalities: tree-like structures. A controversial hypothesis
however is that one model can predict all of perception (Newell). The problem
is that perception is highly ambiguous. Charniak claims that the average
sentence from the WSJ corpus has one million interpretations.  Yet Humans
usually perceive only one grouping structure. According to studies there is
more than 96% agreement among trained people about this structure.

Historically there have been two principles to resolve ambiguity: simplicity
and likelihood. Simplicity selects for economy, eg. the shortest derivation;
likelihood selects for the most likely derivation based on the joint
probability of earlier experience.

The Data-Oriented Parsing (DOP) framework combines these two principles: the
simplest derivation is sought, but at each step only the n most likely
candidates are considered. DOP is a tree grammar where tree units can be of
arbitrary size: from words to whole sentences. Simplicity is applied by
selecting the most similar fragment, likelihood by considering all possible
subtrees in the corpus and counting frequencies.



Word count: 204
Andreas van Cranenburgh 0440949