Externally indexed torrent

If you are the original uploader, contact staff to have it moved to your account

Textbook in PDF format

Grammars for natural languages show how sentences (and their meaning) are built up out of smaller pieces. Syntactic parsing is the task of applying a grammar to a string of words (a sentence) in order to reconstruct this structure. For example, “The dog thought there was day-old food in his dish” has a sub-structure “there was dayold food in his dish” which in turn contains structures like “day-old food.” Before we can build themeaning of the whole wemust at least identify the parts from which it is built. This is what parsing gives us.
As with most all areas of natural-language processing (NLP) parsing research has greatly benefited from the statistical revolution — the process of absorbing statistical learning techniques into NLP that began about twenty five years ago. Prior to that time we had no parser that could, say, assign a plausible structure for every sentence in your local newspaper. Now you can download several good ones on the web.
From the outside the result has looked sort of like a Moore’s law scenario. Every few years parsers got more accurate, or much more efficient, or both. From inside, however, things looked quite different. At more than one occasion we in the community had no idea where the next improvement would come from and some thought that we had, perhaps, reached the end of the road. The last time the improvement came from Slav Petrov and the ideas in this monograph. The embodiment of these ideas is the “Berkeley Parser.”
The best parsers models are all “supervised,” e.g., we have a corpus of sentences, in the case here the so-called “Penn tree-bank” where sentences have been analyzed by people so for each sentence has been broken down into a tree structure of components. A computer learns to parse new sentences by collecting statics from the training data that (we hope) reflect generalizations about a particular language, in this case English. We then recast the parsing problem as one of applied statistics and probability— find the most probable parse for the sentences according the the probabilities already obtained from the corpus