Bikel, D: Statistical Parsing Exposed

This book develops techniques and methodologies forthe examination of the complex systems that arelexicalized statistical parsing models. The primaryidea is treating the "model as data", which is not aparticular method, but a paradigm and a researchmethodology. I argue that lexicalized statisticalparsing models have become increasingly complex, andtherefore require thorough scrutiny, both to achievethe scientific aim of understanding what has beenbuilt thus far, and to achieve both the scientificand engineering goal of using that understanding forprogress. In this book, I take a particular, dominanttype of parsing model and perform a macro analysis,to reveal its core (and design a software engine thatmodularizes the periphery), and also cruciallyperform a detailed analysis, which provides for thefirst time a window onto the efficacy of specificparameters. These analyses have not only yieldedinsight into the core model, but they have alsoenabled the identification of "inefficiencies" in thebaseline model, such that those inefficiencies can bereduced to form a more compact model, or exploitedfor finding a better-estimated model with higheraccuracy, or both.