The authors have the challenge of covering a vast area, and they do a good job of highlighting the hard problems within individual sub-fields, such as machine translation. The availability of an accompanying Web site is a strong plus, as is the extensive bibliography, which also includes links to freely available software and resources.
Now for the negatives.
While I would still buy and recommend this book, you will need to supplement it with other material; in addition to the accurate "broad and shallow" comment made by another reviewer, I would add that much of the material, as presented, is aimed at the comprehension level of a computer-science PhD and doesn't really meet the definition of a textbook for either undergraduate or graduate students. It is not that the material is intrinsically difficult: one recurring problem in the book is the vast number of forward references, where a topic is introduced very briefly but not explained until 20-50 pages later. In most cases, if you don't understand a passage in the text, I would advise that you keep skimming ahead - you may be rewarded because in several cases, the book covers a particular approach for 2-3 pages before telling you that its underlying assumptions are flawed, and that modern methods for addressing the problem use alternative approaches.
In other cases, the authors try to explain topics that might deserve entire chapters in about ten lines - a poster child is the explanation on page 736 of how Support Vector Machines can be used for multiclass problems. To someone who is familiar with SVMs, this material is unnecessary, while those who are not will not be enlightened by knowing that SVMS are "binary approaches based on the discovery of separating hyperplanes". I understand that this is not a text on machine learning approaches, even though machine-learning approaches have revolutionized NLP, but if the authors are clearly in no position to do justice to a particular topic in limited space, I would have preferred that they do the reader the courtesy of acknowledging the same, and simply point to a useful source, preferably online. (While the Wikipedia entry on SVMs is, as of this writing,incomprehensible to non-Math PhDs, the 2nd Google link, at [...] provides a reasonable overview.)
On the other hand, in a book that has to cover a vast area in limited space, there is a surprising amount of repetition. The page-long explanation of F-measure, a statistic used to evaluate the accuracy of a method, is repeated in three places almost verbatim, on pg. 455, 479 and 733; the repetition 24 pages apart (in chapters 13 and 14) should be considered astonishing given that the same author in the two-author collaboration clearly wrote both passages.
Finally, given the way algorithms are described - some reviewers point to errors in some of the descriptions, but I can't verify this - you would be hard-pressed to complete many of the exercises that follow each chapter, in terms of being able to implement a working program.
A final word of advice to the authors: I really do want to see a Third Edition, but I would recommend that you beta-test your material on a sample of your target audience, and incorporate their feedback. When you write a textbook, you really need to make a serious effort to communicate: if smart undergraduates or grad students tell you certain material is hard to follow, the fault almost certainly lies with you and not them.