User:Oshanis/SCFG

Representational problem with training stochastic context-free grammars:

Unsupervised learning techniques such as the inside-outside algorithm produces grammars that structure text in ways contrary to our linguistic intuitions. For example, if we consider the phrase “walking on ice”, it is customary for a human to group the prepositional phrase “on ice” rather than to group the verb phrase “walking on”. However,

Computational problem with training stochastic context-free grammars

2. Simple linguistic example, similar to the one that deMarcken exhibits, illustrating the representational problem.

3. Summarize what deMarcken calls for as a solution to these two problems.

Head driven grammatical formalisms like link grammars. Framework for CFG induction that sidesteps many of the search problems the previous schemes have had.

With respect to the Collins paper:

(1) What parts of deMarcken’s representational proposals does it adopt? Be specific.

(2) Which proposals are novel to the paper itself? Please describe these via a short list. (Here we mean specific ideas and methods that are introduced here for the first time, as opposed to being algorithmic methods referenced in the paper itself – for example, the paper notes that the method of using ‘histories’ as a conditioning probabilistic context was first used by Black at IBM. How was this modified?)