SUBJECT: &NAME book review &NAME , I 've finally managed to finish the review of the &NAME book , which &NAME handed down to me . Sorry it took an extra &NUM weeks than expected , but a bunch of things came up right at the time I had planned to do it ... Anyway , I 'm attaching it to this message . Let me know if you need anything else , or if something is n't quite right . Also , since this is my first book review , feel free to let me know if I have n't adequately done the job ! Cheers , &NAME &NAME &NAME , editor . &NAME : Foundations of Speech-to-Speech Translation . &NAME . &NUM . &NAME &NUM . &NAME &NUM pounds ( hardback ) . xii+679 pages . &NAME was a project for speech-to-speech translation that lasted &NUM years and encompassed &NUM research groups , primarily in &NAME . The original goal was to develop a portable translation system for German and English for the domains of appointment scheduling and travel planning . Later , translation between German and Japanese was added and both a telephone translation system and a remote maintenance system for &NAME were developed . This book contains &NUM articles , grouped into &NUM sections , which address various aspects of the &NAME project . The title is somewhat misleading , as the phrase ' foundations of speech-to-speech translation ' seems to imply that the book will be instructive with respect to the task , or at least contain comprehensive high-level discussion about it . Instead , most articles of the book report on specific components and approaches developed and utilized to support speech-to-speech translation in &NAME . This is fine , but readers should be aware that the book is in no way a textbook ; background knowledge relevant to the individual topics is assumed by the authors of most articles . Having said this , if the term ' foundations ' is taken to mean consideration of all the various kinds of data , components , techniques , etc. that could take part in performing the task , then the articles of the book certainly cover the ground extensively . The articles themselves vary in their focus and emphasis . The stated goals of &NAME were essentially oriented toward producing an end application , and this course provided an excellent setting for a considerable amount of research to be done . Thus , some articles could be roughly characterized as ' here is how we made it work ' , while others could be captioned as ' here is this interesting thing we did within the scope of the project ' . As such , there is material of interest for both academically and commercially oriented readers . Throughout the book , there is a strong emphasis on producing modules that can perform their computations in real-time , and with such a large number of different technologies applied in &NAME , this should be of interest to those interested in developing various kinds of practical &NAME systems . With all the different topics covered , much more high level discussion of how the various bits and modules fit together would have been very welcome , including ( even brief ) overviews of the different sections explaining , basically , what the different articles are about and why they are there . Such overviews also could have provided a forum for more meta-level discussion on how well the different approaches and strategies worked out . It would also have been useful to have a map of the system with pointers to the articles where the relevant modules are discussed ; it is sometimes hard to tell what depended on what from just reading the articles , and it is often necessary to cross-reference several articles to properly situate them within the context of the overall system . The articles are too numerous and diverse in specialization to comment on each one individually in a short review by &NUM reviewer ; however , I will try to give an overall sense of the book 's contents in the remainder of this review . The first section contains &NUM articles that provide an overview of the completed &NAME system and facts and figures about the project itself . The overview is quite useful for getting a feeling for the system as a whole before launching into the articles on the specifics . It provides very high-level descriptions of some of the various components and how they fit together , as well as general discussion regarding the scope of the task which was undertaken , the data collection effort , and software architecture . It concludes by summarizing some of the lessons learned over the course of the project ; unfortunately , only &NUM pages are devoted to this , when in fact it seems that it would be worthy of its own chapter . Also , the overview clearly stays focused on the successes of &NAME , whereas it would be very beneficial for the reader , whether research or application oriented , to also hear about the problems that were encountered and the things that did not go so well . The next section contains &NUM articles that deal with topics in processing spoken input , with all its messy details , and transforming it into representations that can be processed by the syntax , semantics , and translation components . The first half of the section is concerned with the methodologies and techniques for speech recognition , covering topics such as language identification , providing robustness in the face of various noise conditions , optimizations for improving the efficiency of speech recognition components , and investigations into new techniques for capturing long range dependencies in language modeling . The latter articles report on the creation of pronunciation dictionaries from data to provide variants for canonical pronunciations , the &NAME and classification of prosodic information , detection of emotional behavior by users , and the detection and correction of speech repairs as a filter process between recognition and subsequent processing stages . The third section is the largest -- it includes &NUM articles addressing lexical and syntactic processing . &NUM different approaches to parsing were developed and integrated : a robust chunk parser built from a cascade of finite state transducers , a probabilistic &NAME parser , and an optimized &NAME &NAME &NAME Grammar ( &NAME ) parser . These approaches , which are each discussed in separate articles , provided the means to balance robustness with accuracy and depth of syntactic analysis . There is also a further , brief article describing how their results were combined . While the &NAME parser provides full semantic analyses , the chunk parser and &NAME parser produce only syntactic trees ; how interpretations are read from these trees is discussed in an article on semantic construction . The use of a &NAME syntactic framework such as &NAME required considerable justification in the face of the cost of developing grammars and efficient parsing techniques , the third section contains an article that lays out some of the arguments for using &NAME over other available frameworks and answers some of the concerns regarding the efficiency and complexity of having deep analysis as such at all . &NUM further articles describe aspects of the &NAME grammars which were developed for German , English , and Japanese . Since the results of the &NUM parsers need to be integrated into a single interpretation for further processing , it was necessary to develop a standard interface representation that abstracted from the particularities of each of the approaches . For this , &NAME &NAME Terms ( &NAME ) were developed and employed , and are motivated and described in detail in the fourth article in this section . &NAME provide an underspecified , flat representation that permits many different kinds of information about an utterance to be stored and accessed , including basic predications , scope constraints , sorts , discourse relations , syntactic features , tense and aspect , and prosody . The final article of this section describes the multilingual lexical databases that were developed and highlights some of the issues in dealing with the morphology of the different languages and ensuring that it was possible to retrieve translation equivalents from &NUM language to the other for the semantic transfer translation component . There are &NUM articles on aspects of semantic processing in the fourth section . The first discusses how the analyses produced by the &NUM different parsers were combined to produce a single semantic output , and the second deals with processing the intermediate &NAME hypotheses of the parsers in robust manner that corrects certain recognition errors and disfluencies in the semantic output of each parser . The next paper focuses on how some of the non-local ambiguities in &NAME were resolved by incorporating discourse and prosodic information . The rule-based module employed for this handled aspects such as verb and adverb disambiguation , anaphora and ellipsis resolution , and focus projection . The final article in this section relates how the semantic databases were designed to support multiple languages and ensure that they map material in a given source language to comparable semantic relations in &NAME to assist in semantic transfer translation . In the fifth section , we come to descriptions of the actual translation engines employed in &NAME . &NUM main approaches were explored : semantic transfer , statistical translation , substring-based translation , case-based translation , and dialog-act based translation . The different engines tackled the problem in significantly different ways that gave them , in some cases , complementary strengths and weaknesses . Echoing the multi-engine approach to parsing discussed earlier , the translations ultimately produced by the &NAME system were chosen from those created by these different modules , based on the confidence values they assigned to their outputs . The sixth section contains &NUM articles which address the recognition , representation and use of discourse information within the system . The first article explicates the representation of dialog , including illocutionary force , propositional content , and temporal expressions , and the following article discusses how dialog structures are produced for the purposes of both context-sensitive translation and producing multilingual summaries of the dialogs mediated by &NAME . The final article describes the module which performs contextual disambiguation using the variety of sources of information built up during the ' deep processing ' path in &NAME and which plays a role in semantic transfer translation . The seventh section deals with sentence realization and speech synthesis . First to be discussed is the &NAME generation component , which takes the &NAME output by the semantic transfer translation module and determines how they should be expressed . After a &NAME is fully specified , it is handed to &NUM syntactic realization components , one which uses &NAME Grammar ( &NAME ) and another which uses &NAME . The grammars for the former realizer are compiled from the English and Japanese &NAME grammars discussed earlier ; the HPSG-to-TAG compilation process is the subject of the second article in this section . The next article deals with the creation of dialog minutes and summaries , a useful task which crucially relies on the dialog processing component described in the previous section and the generation component covered in this one . The final article discusses the concept-to-speech synthesis component used for producing the final spoken output of the system . The articles in the eighth section describe various aspects of the data collection and corpus creation efforts and evaluation and analysis of different components of &NAME . These include information about recording , transcription , treebanking , and analysis of the data utilized to develop the &NAME system . &NUM articles provide evaluations of the speech recognition and synthesis components and the quality of the different translation paths . These evaluations were carried out at different sites from those which produced the modules , which was certainly of particular importance for the translation evaluation which involved &NUM competing approaches . There is a further article on a module which attempts to select the best output from one of the translation modules based on the confidence values they produce ; this article might have been better situated in the fourth section of the book , especially since it provides some higher level discussion of the different translation approaches . Finally , the &NUM articles in the ninth section deal with concerns of software architecture and integration . Since &NAME involved &NUM research groups with scientists and programmers with varying competencies with respect to programming skills and knowledge of software engineering , it was no small task to coordinate the components that were produced at the individual sites and enable them to work together as a whole . A special group was devoted to this task , and the article on their effort provides an informative look at the technical ( and sociological ) issues which were overcome in this integration . The other article describes how &NAME was extended to provide telephone translation services . &NAME tackled all sorts of problems on all sorts of fronts , resulting in an impressive and extensive body of work that led to &NUM scientific publications and &NUM spin-off products . The project 's funding and activities provided an immense boost for computational linguistics research and development in &NAME . This book provides significant insight into the project , and it will be useful to researchers and commercial developers who are interested in specific aspects of speech-to-speech translation and how they fit within the context of an overall system for translation . The articles are so diverse that it is hard to imagine any single person reading all of the articles contained in the collection since only a subset of them are likely to be of particular interest to any given reader ; nonetheless , it will certainly be an excellent resource for anyone who is specifically interested in &NAME itself and can use it to quickly find information on all the modules in &NUM big handy volume .