Macro Grammars and Holistic Triggering for Efficient Semantic Parsing
Yuchen Zhang, Panupong Pasupat, Percy Liang
Consider a semantic parser that must search for programs (logical forms) with the correct execution results during training.
To speed up search, our method retrieves past training examples resembling the current example,
and then try out the search actions (represented as "program macros") from those examples.
This gives 11x speed up!
Errata: In Equation 9, the term in the logarithm should be
pi+ / pi-
pi+ / (pi+ + pi-).
Version 2 on arXiv has the corrected formula.