Natural Language Understanding

Last week, I was looking at the back flap of this book, and I saw something so shocking that I started choking on my pretzels! My wife, after slapping me on the back till I stopped choking, glanced down at the page--filled with bizzare symbols--and said, "How shocking could that page possibly be???" Well, I was looking at a list abbrieviations of the categories (parts of speech) which the book used, and I noticed, for the first time after owning this book for over 10 years, that there was no abbriviation for "conjunction" listed. And indeed, after consulting the index and looking through the book, it is plain to see that this book doesn't treat conjunction at all! I have many fond memories of this book--it is the book which my beloved professor at grad school taught me NLP from, and indeed, it contains far more information about NLP than most of its successors. For example, this book gives perhaps the best discussion of quantifier scope ambiguities of all the major NLP textbooks. (cf. with Jurafski and Martin's book, which devotes about 1/2 a sentence to quantifier scope ambiguities). But it has odd ommissions, one of which is the lack of treatment of conjunction/disjunction. After devoting so much time to quantifier scope, why does Allen leave me in the dark about whether "Every woman" can take scope over "a man" in the sentence "A man and every woman hug each other?" Does that scope differently from "Every woman and a man hug each other?" Or what about "Every woman and her mother fight?" Can that mean "Every woman fights with her mother" or are we to look for another antecedient for "her"? Or again, Allen's treatment of prolog-esque definite clause grammars. Allen deserves major kudus here for including them. Its obvious that he comes from the LISP side of the tracks, and most LISPy books on NLP ignore DCG's altogether (Norvig's "Paridigms of AI programming" being a notable exception). But it seems almost like Allen goes out of his way to present DCGs in the most unattractive light possible. Prolog has a nice syntactic sugar which makes a DCG look almost exactly like a context-free grammar specification, but you'd never know that if you only read this book--Allen chooses a wierd way to translate strings into clauses, which implies a bizzare-looking prolog grammar for them. The student naturally recoils in horror, but unless she reads a prolog-oriented book on NLP, she would never know how much easier DCGs are to program than ATNs or the bottom-up parsing methods which Allen goes on to expostulate. Since this book was published, the field of NLP has taken a bit of a side-track through statistical learning of grammars--the thought being that, well, we really don't know how to do knowledge representation or pronoun resolution very well, so lets all spend a decade or so on how to induce grammars from corpora. This book doesn't cover any of this research, but frankly, I really don't consider that a critique of the book. Bec

Allen's book is a standard introduction to NLP in industry and academia. His exposition permits straightforward implementation and provides a lucid motivation for the algorithms he describes. "Natural Language Understanding" was the first NLP text I read (for a summer job), and I've always referred to it first for its balance of formal and practical considerations.However, some developments in the past few years outpace his treatment. In particular, the stochastic viewpoint has become more common in natural-language processing, and Allen does not consider related innovations in great depth. Passing mention of the interaction between NLP and the speech processing and information retrievals would also benefit a revised version.

James Allen introduces the concepts required to build a NL system without losing you in the psycholinguistics, psychology and philosophy of language. The great part is that he gives you enough of a background in each of these fields before going into a topic in depth. Plenty of examples help define words like "morpheme" so you don't get lost in the psychobabble jargon of the required disciplines. (Wouldn't that be an ironic thing to have happen in a NLP book?!) :) I'd love to tell you more, but I only stopped by to catch the author's name and look for whitepapers on his (hopefully existant) website.

This books serves as a very useful introduction to computational linguistics. It is clearly laid out, and will serve the reader for many years as a continuing reference.The TRAINS project at Rochester (the author's institution) was based on many of the concepts outlined in this book... proof that they work and can be made to handle real-world situations.I continue to use it, now in its second edition as a reference for myself, and to train those who need to work with our project.I cannot recommend a book more highly. If you want to learn computational linguistics, or need to push the state of the art, this is the book you need.

An excellent book; the best introduction to the subject of Natural Language Processing, made even better by its second edition. If you wish to learn how computers can be made to process human language, go no further.