Browse Books

Reviewer: Jeffrey B. Putnam

Natural languages-for example, English or French-are messy. In fact, they are really messy. Programming languages usually have syntax that was designed to be easy for computers to understand, and the meaning of any bit of program text is usually well defined. In contrast, natural languages are hard to parse, and different languages are often very different when it comes to word order and sentence structure. To compound the difficulties, it may even be hard to figure out what a word is. And, naturally enough, natural languages are almost always changing-for example, the addition of new words, such as LOL (laughing out loud), and new punctuation, like the smiley emoticon. And just when you think things can't get any worse, there are all these odd idioms that don't mean what they look like they should mean. Basically, it's raining cats and dogs with difficulties. Still, we'd like to have programs that can do something reasonable with natural language. In particular, given the deluge of written texts that are now available electronically (and growing fast), and the number of languages involved, we'd like to be able to automatically translate one language to another. Statistical machine translation methods take parallel corpora-essentially the same content in two languages-and, using statistical models and some knowledge of how languages are generally structured, derive translation engines. Frequently, these parallel corpora are representative of a single knowledge domain. As just one example, the Europarl corpus is the proceedings of the European Parliament. These translation engines are built specifically to translate; they don't necessarily get at (or even need) the meaning of the texts in question. This effort has been met with some success, and this text provides an excellent overview into how it is done. The book is broken up into three parts. The first part introduces the topic, provides some basic notions-what is a sentence, for example-and provides a basic overview of the math required. The second part, "Core Methods," is the real meat of the book. One chapter is devoted to words and the problem of deciding which words in parallel corpora correspond to each other. One chapter covers phrases (which may require reordering of the constituent words), and another discusses decoding-the process by which the best translation is generated-a problem that is nondeterministic polynomial-time (NP) complete. The next chapter covers language models that mark sentences on how likely they are to occur in text generated by humans fluent in the target language. Finally, there is a chapter on how to evaluate constructed translation engines. Part 3, "Advanced Topics," covers discriminative training, adding linguistic information (transliteration, capitalization, and other factors), and tree-based models that construct parallel tree rewriting systems-that is, this kind of phrase as a tree translates to that kind of phrase as a tree with nodes labeled to correspond. Each chapter comes with an extensive bibliographic summary and exercises, which are usually based on software and corpora that are available online. There are some oddities. Entropy is defined twice-once in the preliminaries and once in a later chapter. A number of words and phrases are in boldface and noted in the margins; while this is a nice touch, it would have been even more useful to list the words in a comprehensive index. This is an excellent introduction for someone interested in statistical translation. It is quite readable, although a few dense sections could use a bit more explanation and a few more examples. As it is, it might be difficult to use for self-study. While the book might be appropriate for advanced undergraduates, it is certainly suitable for graduate-level courses. Online Computing Reviews Service