The Spoken Language Translator (review)

The creation of an automatic system of recognition and translation of human speech has been a dream of computer scientists and linguists since the appearance of computers. Still, the goal remains a long-term one, and the task requires plenty of research. This book tries to show what we can realistically attempt today in this intriguing field of human knowledge. Stating that spoken language translation is one of the most accessible speech and language processing tasks imaginable, the authors of the book present a detailed description of the Spoken Language Translator (SLT), one of the first major projects in the area of automatic speech translation.

The greater part of the book describes the language processing components, which are built on top of the SPI Core Language Engine and which use a combination of general grammars and techniques that allow them to be rapidly customized to specific domains. Though the SLT system can actually do some translation among English, French, and Swedish in the domain of air travel planning, we would say that the project is more of theoretical value: Though it has not yet resulted in a commercial automatic translation system, it represents a well-documented attempt to build a general grammar-based speech translation system. The most significant contribution of this book, as the editors state it, has been to demonstrate a serious possibility to use hand-coded grammars in such applications as rapid creating of a parsed corpora by graphical disambiguation, which could then be used to train efficient versions of a general grammar specialized to the particular characteristics of the training corpus. No doubt, the authors also have achieved solid progress in building large unification-based linguistic descriptions; in fact, some general recipes for converting grammars for English into corresponding grammars for French and Swedish are provided. The editors claim that similar methods would work for most other European languages. In the course of discussion, the editors show that it is practically feasible to write a single large grammar valid for two closely related languages—an idea which has only been vaguely speculated by now but which can provide a vast and very interesting field for further research. In general, they have developed a framework that, in their opinion, allows them to implement large-scale unification-based grammars by efficient reuse of existing resources.

As for translation itself, the book supports the idea of multi-engine translation, and the experiments suggested here can be considered the first formal description of the phenomenon of pipeline synergy. The editors present a hybrid transfer architecture based on quasi-logical form transfer and trainable preferences, which is believed to represent a good compromise between the opposing positions of pure grammar-based transfer on the one hand and pure statistical translation on the other.

Putting it all together, The spoken language translator is a most useful and interesting resource book. The results of the project seem to be solid to the best of modern knowledge. The book can answer some questions for those readers who would like to know what progress has already been achieved in the field of automatic translation and what problems still are waiting to be discussed in future.