Reviewer: Tania Reynolds

Computing in Linguistics and Phonetics constitutes a comprehensive review of the current computing applications in linguistics. It is a general introductory text aimed at the linguistics undergraduate.

The book is set out in seven sections, each covering a specific area of application. The first chapter begins with an introduction to computing and computing terms. It describes a computer and a set of basic terms.

The second section introduces the application of data banks in linguistics. Their usages are briefly described, with examples of the types of searches that are carried out. The author goes on to describe the structure in which the data is stored and how the structure is related to the retrieval mechanism. The stored text may be spoken or written. Text is tagged with different types of information such as grammatical, phonetic or typographical. Examples of each sort are given.

The third chapter describes the ways in which computers are used to analyse, produce and recognise speech. Speech can be analysed acoustically. Waveforms are represented graphically and are analysed by frequency, cycle and filtering. Applications of this type of work include work improving intelligibility of speech such as that involved in teaching languages or teaching deaf people to speak.

The analysis of speech production is concerned with picking up airflow information which can be converted into an electronic signal that then is processed by a computer. Another area tries to study the changes in dimensions of the vocal tract when a sound is produced by the movements of the jaw, tongue, lips and soft palate. Speech recognition is another major area of research.

The fourth chapter deals with Natural Language Processing. The methods of phrase structure grammar, Transition Networks, Sentence Networks, and Augmented Transition and Sentence Networks are described in some detail.

Chapter five deals with speech production by machine. It describes the process involved in producing a speech signal. Two basic methods are described. These are the Terminal Analog System, which simulates human speech, and the Live Analog System which copies the way in which sound is produced.

Chapter six introduces the subject of Machine Translation. A potted history of the development of MT is briefly discussed, with examples of MT systems to illustrate the points. Translation is described as being a three stage process. These stages are analysis, transfer of the logical representations of language from the source language to the target language. MT and its limitations are also covered.

Computer Assisted Language Learning (CALL) is the subject of the last chapter, which presents a comprehensive survey of currently available software and a concise history of developments in the field. There are also some suggestions for future developments.

Each chapter deals with a field or branch of computational linguistics and as such represents an extensive body of learning in its own right. Since the text is an introductory one, care has been taken to present a broad overview of each area. In some cases this has led to a rather simplistic approach. For example, there are many limitations to the ways in which computers are able to represent or process natural language: computers cannot, for example, translate perfectly between languages, or spontaneously generate sentences in the way in which human beings do. The creative element in language cannot be mirrored by a computer and forms a constraint on all types of applications.

Hardware limitations are briefly discussed in the book, but I suspect that greater emphasis could have been placed on the fact that a computer solution to a language processing problem generally requires a large amount of storage area and power.

To sum up, the book admirably fulfils its main objective, which is to provide an informative guide for the linguistics undergraduate. The articles are all precisely written in straightforward language, and do not require any previous computing experience. The scope of subjects includes applications in most aspects of linguistics, ranging from syntax (phrase-structure grammars), to phonetics and language acquisition. A bibliography of further reading is provided with each chapter. Areas such as historical linguistics and sociolinguistics pose particular problems which are harder to solve by computer, and these are left untouched by the book. Overall, its simplistic approach and breadth of topics make it a highly suitable text for non-computational undergraduates.

Reviewer: Derek Lewis

The authors are, or have been, members of the Computational Linguistics and Machine Translation Group at the University of Essex. They have aimed in this volume to provide a genuine introduction into MT theory and practice by affording a broad range of insights into the field: how MT systems work; their capabilities and limitations; methods for evaluating systems; and likely developments in the short to medium future. Targeted at students, translators and managers, the book presupposes no specialist knowledge of computational linguistics. It is clearly organised, with each chapter providing a summary of points covered and recommendations for further reading. A short glossary of technical terms is included, together with a list of useful addresses.

The practical emphasis is evident from the start. The first chapter tackles head-on popular (and occasionally amusing) conceptions and misconceptions about MT. The second follows the passage of a document within a fictional multinational organisation in order to illustrate how MT could actually contribute to the process of translation within a commercial environment. Such perspectives are particularly useful for language students who are contemplating translation as a career but have little conception of either the type of language which is most often translated or of how documents are prepared and handled within large organisations. The status and nature of electronic documents, the role of controlled languages, and the complexities of evaluating MT systems are further issues considered later in the volume which retain the focus on MT as a practical activity.

Perhaps even more interesting for a computationally non-expert academic readership are the chapters describing the architectures and linguistic models underlying MT systems. For those unschooled in phrase structure theory, the outline of the principles of constituent structure and tree diagrams are especially useful. There are clear expositions of basic parsing techniques and of how linguistic information may be represented in either surface-based forms or 'deeper', more abstract and semantics-orientated ways for subsequent processing by the transfer and generation components of an MT program. Such representations are at the heart of current research and development within the wider field of NLP. Individual chapters are devoted to word formation and morphological processing, the structure of dictionaries and lexical information, and to the identification and handling of terminology within an MT context.

Current and future developments are addressed throughout the book. They include the long-standing preoccupation of builders of MT-systems with such problems as ambiguity, anaphoric references, lexical and structural mismatches between languages, and the extent to which real-world knowledge may be represented within the machine. A separate chapter is devoted specifically to new directions in MT. One such direction is the extension of rule-based approaches to incorporate various levels and types of linguistic information. Further directions include the interest in knowledge-based MT, with its focus on subject domains, and the use of existing language resources for MT purposes, in particular corpora and parallel multilingual texts. There is also a useful exposition of how statistical methods are being employed to perform certain MT tasks, such as the disambiguation of word senses.

A most welcome feature of the book is the way in which the authors have taken care to illustrate translation processes with examples from actual languages (my only quibble in this respect is that the German zurück is curiously identified at one point as a preposition). This explanatory approach lends the book transparency and gives the reader a genuine feeling as to what goes on under the bonnet of an MT system. While many of the main research and operational systems are mentioned, they are not described in any detail. In view of the extent of the MT field, however, this is wholly in keeping with the character of an introductory guide. In summary, this volume admirably achieves its aims and succeeds in combining breadth of coverage with both theoretical and practical depth. It avoids the mystification so often found in computational linguistics and should indeed stimulate its readers to take a long-term interest in MT applications.

Reviewer: Derek Lewis

PLNLP (Programming Language for Natural Language Processing) evolved during the 1970s and 1980s as a general purpose software tool for writing natural language applications. Typically, the developer of the application, a linguist, builds a grammar and lexicon from PLNP rules and data structures in order to parse a sentence. Possible applications range from checking grammar and analysing style to machine translation. Although most applications operate at sentence level, tentative progress has been made in building a model of discourse by linking sentences within paragraphs. The system has various components organised as a continuum and is claimed to provide a highly flexible and integrated set of tools for processing and analysing languages (which include Japanese and Arabic, although most applications are based on English). This volume presents the first published collection of papers (twenty two in all) about the PLNLP system and its applications. Many of the authors were originally members of the group at IBM which pioneered this work at the Thomas J. Watson Research Centre.

NLP systems vary as to their relative reliance on dictionary information or processing rules. Although PLNLP is not committed either way, it is clear that dictionaries are indispensable for many of its components. Of particular interest is the emphasis on employing on-line machine-readable versions of standard lexica from which the computer program extracts the information it needs in order to produce a correct parse or otherwise represent the structure of an input sentence. Most of the papers describe a stage in the processing and its relevance to a particular application.

The component which supplies the initial syntactic sketch of an input sentence, the transductive grammar, aims to produce a parse for any input; that is, it responds to real-world textual data more robustly than parsers in the tradition of generative grammar. The transductive grammar written for English (PEG), operates with a large but unsophisticated lexicon based on Webster's Seventh New Collegiate Dictionary. In the next stage of processing, semantic information is applied to correct, or re-assign, the syntactic sketch produced by PEG. Re-assignment relies on large amounts of information per word held in complete on-line dictionaries, so the lexicon plays an unusually important role for an NLP system. Dictionaries mentioned in the applications described here are Webster's Seventh and the Longman Dictionary of Contemporary English (LODCE). Lexica play no role in the third stage (PEGASUS), where underlying logical forms of sentences are constructed solely from the parse structures derived so far. During the next stage, sense disambiguation , on-line dictionaries are used extensively to narrow down senses of verbs and nouns in the parse. The disambiguation program, called MAST (MApping Senses to Text), works for English by comparing the structure and context of a word in the input text with information in a dictionary entry. Lexical entries in the LODCE follow a strict argument structure which enables MAST to extract the semantic features on subjects and objects of verbs. In this way MAST is able to assign the input word its most probable sense. The final two stages, in which rules and dictionaries interact closely, are still under development. The first of these aims to produce common semantic representations for sentences which mean the same thing (a process started by PEGASUS). This is achieved by a concept grammar which establishes underlying semantic graphs for different surface structures (for example: A is under B is equivalent to B is over A ). Lastly, the system identifies links between words and concepts to build a conceptual model of the topic of a paragraph (this is perhaps the most theoretical of the papers).

The flagship application of PEG is the Critique system, which identifies grammar and style problems in English texts. Relying primarily on syntactic rules to parse input and flag possible errors, it looks for the best structural 'fit' by progressively relaxing grammatical constraints, and, if no complete parse is found, combining the available analysed sub-structures. The user may tune the system to suit his own profile and be presented with message, help and tutorial information. Critique worked best with standardised non-literary texts containing short sentences and has office, publishing and educational applications.

PEG has been used as a front end for a number of prototype machine translation systems, SHALT (English to Japanese), C-SHALT (English to Chinese) and PORTUGA (English to Portuguese, Norwegian and other languages). A transfer system relying on contrastive knowledge of a language pair, PORTUGA aims for broad-based coverage, i.e. it will handle ill-formed and incorrectly parsed input and is not restricted to a limited domain. Its bilingual dictionary is notable for its relative simplicity (possible translations of a source language word are separately annotated with the conditions for its translation) and multi-word expressions can be imported directly from published dictionaries.

The PLNLP group provides convincing evidence of the potential of machine readable dictionaries as sources of extractable information for NLP applications. The implication is that natural language may be a convenient and rich medium of knowledge representation, for instance, in the form of semantic networks, if we can develop the computational techniques to retrieve the information. Illustrations of this approach, which basically uses the dictionary as a source for setting up a knowledge base, include analysing the semantic structure of head nouns as obtained from dictionary definitions (chapter 10), finding the most likely attachment for English prepositional phrases by comparing them with similar constructions in a lexicon (chapter 11), uncovering the semantic relations between members of compound noun sequences, such as 'vegetable market' or 'lemon peel' (chapter 13)), and identifying patterns of semantic relations between verbs and their noun phrases as typically found in dictionary entries (chapter 15).

Although the PLNLP group makes no strong claim for a particular model of linguistic representation, it is evident that they eschew a strict rule-based approach to natural language processing and see the most exciting prospects in the extraction and application of lexical information. Although expensive, the volume represents a unique and clearly presented collection of valuable papers.