Communication and Cybernetics – serie

Since its publication in 1976, the original volume has been warmly received. We have decided to put out this updated paperback edition so that the book can be more accessible to students. This paperback edition is essentially the same as the original hardcover volume except for the addition of a new chapter (Chapter 7) which reviews the recent advances in pattern recognition and image processing. Because of the limitations of length, we can only report the highlights and point the readers to the literature. A few typographical errors in the original edition were corrected. We are grateful to the National Science Foundation and the Office of Naval Research for supporting the editing of this book as well as the work described in Chapter 4 and a part of Chapter 7. West Lafayette, Indiana March 1980 K. S. Fu Preface to the First Edition During the past fifteen years there has been a considerable growth of interest in problems of pattern recognition. Contributions to the blossom of this area have come from many disciplines, including statistics, psychology, linguistics, computer science, biology, taxonomy, switching theory, communication theory, control theory, and operations research. Many different approaches have been proposed and a number of books have been published. Most books published so far deal with the decision-theoretic (or statistical) approach or the syntactic (or linguistic) is still far from its maturity, many approach.

This monograph deals with curious and little-known acoustic systems of commun­ ication based on whistling that are in use at the present time in many parts of the world and have had, on the evidence available, even greater currency in the past. In spite of their wide distribution, they have not so far received the attention they might have been expected to attract: linguistically and in other ways they are most interesting. For it should be realized that they are much more than mere codes or conventional systems of signals; they function in exactly the same way as speech (in the sense we normally use the term), being in fact rather extraordinary realizations of the languages spoken in the re­ gions where they occur. They utilize the vocabulary, grammar and, in many cases to a large extent, the phonology of the local speech. However their phonetic system is profoundly modified acoustically and (in some cases less so) from the point of view of articulation, because the glottal tone of every­ day communication, the "voice", is repl aced by a whistl e which carries the in­ formation. The advantage of this procedure, from the point of view of the user, is a vastly increased range as well as, under certain circumstances, a degree of secrecy. It will be shown that whistled languages appear in two main, very different forms.

Information theory and cybernetics have developed along somewhat different lines in Europe and in the U. S. A. This book is to be seen as a contribution towards bridging the gap. Anyone who seeks to apply information theory in the fields of education and psychology very soon comes up against a central diffi­ culty: in the form in which it was developed by Shannon information theory excludes the semantic aspect. This problem is fundamental for in education, as in psychology, the semantic aspect is the very heart of the matter. Thus, while Attneave, Miller and Quastler, among others, successfully employed the concepts and units of measurement of in­ formation theory in the interpretation of the findings of experimental psychology, they were obliged to restrict their work to its syntactic and statistical aspects. Before we can make use of the methods and results of information we have to solve the central problem: How can theory in actual teaching, we measure the semantic information of a verbal message? The only way to do this is by extending the theory. A special concept has been deve­ loped for this purpose: subjective information. In place of an objectively measurable quantity (frequency of sign sequences) we set an empirically determined one: the subjective probability with which the recipient expects a certain sign sequence.

During recent years auditory research has advanced quite rapidly in the area of experimental psychology as well as in that of physiology. Scientists working in both areas have in cornrnon the study of the process in HEARING, yet different scientific areas always tend to diverge. A SYMPOSIUM ON PSY­ CHOPHYSICAL MODELS AND PHYSIOLOGICAL FACTS IN HEARING was or­ ganized for the exchange of information and to stimulate dis­ cussion between research workers in psychoacoustics, neurophy­ siology, anatomy, morphology and hydromechanics. The basic aim of holding this syrnposium was to halt the divergence and to initiate the kind of multi-disciplinary research that will be need ed to elucidate the hearing process as a whole. The present proceedings comprise the papers, which were circulated to the participants two months before the syrnposium and discussed during the syrnposium, together with some cornrnents and additional re­ marks. These cornrnents and rernarks do not, however, represent the full discussions but only the parts available in written form. We have arranged the material in five sections: I. Structure and Neurobiology of the Inner Ear II. Cochlear Mechanisms III. Auditory Frequency Analysis IV. Auditory Time Analysis V. Nonlinear Effects Within the limits of a syrnposium, none of these topics could be treated comprehensively; moreover, most of the papers concerned problems having several aspects.

This book contains among other things, developments of ideas suggested in my papers "Pseudo-units in phonetics" and "On linguistic terminology" read to the Congresses in Prague and Bucharest (both 1967) and in my book Linguistische Ein­ heiten im Rahmen der modernen Sprachwissenschaft (1966), which, in its turn, includes ideas contained in previous papers. It is my belief that units and relations between units are the basic linguistic facts to be described. Units must be defined in relation to other units, i.e. as elements of systems. Relations or systems contain as elements units having some "positive" definition (and not merely a "negative" one), i.e. units have describable features which can be compared with features of other units. It is not sufficient to establish that different units are functionally different but one must also describe in detail the character of the difference between them. A language has a paradigmatic system, i.e. a network of paradigmatic relations, and a syntagmatic system, i.e. a network of syntagmatic relations. Following J. Baudouin de Courtenay, F. de Saussure, N. S. Trubetzkoy, L. Hjelms­ lev and so many others I think that the difference between form and substance, be­ tween units and their realizations, henceforth called items, should be emphasized. A clear recognition of language facts leads to a drastically increased number of units and items. This will not amaze anybody having at least an intuitive under­ standing of the immense complexity of a natural language.

During the past ten years a new area in speech processing, generally referred to as linear prediction, has evolved. As with all scientific research, results did not always get published in a logical order and terminology was not always con­ sistent. In mid-1974, we decided to begin an extra hours and weekends project of organizing the literature in linear prediction of speech and developing it into a unified presentation in terms of content and terminology. This effort was completed in November, 1975, with the contents presented herein. If there are two words which describe our goals in this book, they are unifica­ tion and depth. Considerable effort has been spent on showing the interrelation­ ships among various linear prediction formulations and solutions, and in develop­ ing extensions such as acoustic tube models and synthesis filter structures in a unified manner with consistent terminology. Topics are presented in such a manner that derivations and theoretical details are covered, along with Fortran sub­ routines and practical considerations. Using this approach we hope to have made the material useful for a wide range of backgrounds and interests.

The many different mathematical techniques used to solve pattem recognition problems may be grouped into two general approaches: the decision-theoretic (or discriminant) approach and the syntactic (or structural) approach. In the decision-theoretic approach, aset of characteristic measurements, called features, are extracted from the pattems. Each pattem is represented by a feature vector, and the recognition of each pattem is usually made by partitioning the feature space. Applications of decision-theoretic approach indude character recognition, medical diagnosis, remote sensing, reliability and socio-economics. A relatively new approach is the syntactic approach. In the syntactic approach, ea ch pattem is expressed in terms of a composition of its components. The recognition of a pattem is usually made by analyzing the pattem structure according to a given set of rules. Earlier applications of the syntactic approach indude chromosome dassification, English character recognition and identification of bubble and spark chamber events. The purpose of this monograph is to provide a summary of the major reeent applications of syntactic pattem recognition. After a brief introduction of syntactic pattem recognition in Chapter 1, the nin e mai n chapters (Chapters 2-10) can be divided into three parts. The first three chapters concem with the analysis of waveforms using syntactic methods. Specific application examples indude peak detection and interpretation of electro­ cardiograms and the recognition of speech pattems. The next five chapters deal with the syntactic recognition of two-dimensional pictorial pattems.

The attempt to simulate dialogues in Natural Language by a machine requires extensive analyses of Natural Language's expression and content phenomena. Carefully deducted natural laws must be extracted. A division of all existing Natural Languages into carrier systems of a) agglutinated and b) isolated mor­ phological structures appears to be of principal significance. Thus morphology can be clearly separated from syntax. While morphology concerns structural phenomena, syntax refers to functional customs and rules of language expressions. Expression morphologies of usual language systems like English, French, German or, Russian exhibit tightly agglutinated characteristics. A smaller portion of Natural Language carrier systems provides morphologies of considerably less degrees of agglutination. Among them are ideographic-, pictographic-and, gesture systems as well as air-controller and children languages within a certain phase of development. Sometimes fully self-explanatory or content-related expression units within carrier systems of isolated morphologies guarantee significant insights into phenomena of Natural Language's content. Therefore evaluations on Natural Language expression structures can never be limited exclusively to auditive and phonographic morphologies. They also incorporate transport means of ideo- and pictogenetic characteristics, in order to evaluate morphology and syntax distinctively. The process of formalizing Natural Languages is highly interdisciplinary. It consists of 1) analyzing, 2) possible enumerating, 3) optimizing, and 4) synthesizing procedures. Irrelevant domains of formalization excesses are avoided by keeping strictly to definitions demarcating natural from artificial languages. Comparative evaluations of agglutinated as well asisolated morphologies are necessary.

The purpose of the Symposium was to provide a meeting place for those working in the field of s·peech perception, whose main in­ terest is in the study of the perceptual processes in the deco­ ding of connected speech, hence the title Dynamic Aspects of Speech Perception. It was felt, after the meeting of the 8th ICA in London and the 2nd Speech Communication Seminar in Stockholm, 1974, that there should be an opportunity for an exchange of ideas on this topic with the emphasis on discussion and interpretation, rather than on the presentation of experimental results. The initiators set themselves up as a planning committee and asked the present editors to organize a symposium in Eindhoven at the Institute for Perception Research, which has a well established tradition in perceptual work. The present proceedings contain papers by contributors invited by the planning committee, as well as discussions. Most of the papers were circulated well in advance of the Symposium. We arranged the material in five sections: I. Theoretical Issues in Speech Perception Research II. Prosody in Speech Perception III. Some Storage Properties of Speech Perception IV. Short-term context effects, Dichotic listening, Speech Perception in the very young V. General discussion and tentative conclusions The order of papers is not the same as at the Symposium. Two papers are printed that were not introduced orally by their authors, viz.

The point of departure of this general survey of modern structural linguistics is the place of language in human relations. Linguistics will consequently be understood as a science of communication. My book is not intended as an elementary handbook. The readers are supposed to be in the first place advanced students of linguistics and phonetics and of neighbouring fields where a real awareness of linguistic methods and problems is essential (such as psychology, phoniatrics, speech therapy, language teaching, communication engineering). The book may, however, be of some value also for the general reader who is interested in language, in language learning, or in communication processes. It might finally serve as an introduction to structural theories and practice for those linguists of traditional orientation who would like to make contact with the new trends in the study of language. It is self-evident that, under such circumstances, any reader will find certain chapters in this book rather complicated, others irritatingly elementary. This is, however, unavoidable in a work whose aim is to cover a vast field of knowledge and to offer the reader a synthesis of what appears at first sight to be widely disparate facts and phenomena. Many of the facts brought together here may, regarded superficially, seem to have few or no mutual connections. They can, nevertheless, be combined into a wide humanistic and scientific unity within which numerous lines of relationship bind together physical and psychic, individual and social phenomena.

In trying to give an account of the statistical properties of language, one is faced with the problem of having to find the common thread which would show the many and multifarious forms of language statistic- embodied in scattered papers written by linguists, philosophers, mathe­ maticians, engineers, each using his own professional idiom - as belong­ ing to one great whole: quantitative linguistics. This means that the investigator has to find the system of this branch of science which would enable him to arrange the vast material in an orderly fashion, and present it as an organic whole. Such a system is conceived in this book, as comprising the following disciplines as the four main branches of literary statistics: Statistical Linguistics, Stylostatistics, Optimal Systems of Language Structure, and Linguistic Duality (Parts I-IV). The Introduction is meant to define the position of the book with regard to both, linguistics and statistics.

About the Scope of This Text This book contains two types of material ~ first, the many divergent and often diffuse meanings given to the concepts of association, associative memory, and associative recaZZ are expounded. A review of this kind was felt necessary because there apparently does not exist any single monograph which could serve as a reference to these topics. But the presentation of the main body of this text is motivated by quite other reasons: in recent years, plenty of interesting mathematical and system-theoretical material has been published which makes it possible to gain a view of associative memory which is different from the conventional abstract and computationally oriented approaches. It seems that the basic operation of associative memory, the storage of information together with the relations or links between the data items, and the selective recall of stored information relative to a piece of key or cue information presented, is not restricted to certain computer-technological implementations but can also be reflected in more general mathematically describable processes in certain physical or other systems, especially in their adaptive state changes. It further seems that some generally known forms of associative memory, namely, certain computer technological artifacts, or abstract systems of concepts or data, are in fact special representations of a class of processes characterized as associative memory.