The following online article has been derived mechanically from an MS produced on the way towards conventional print publication. Many details are likely to deviate from the print version; figures and footnotes may even be missing altogether, and where negotiation with journal editors has led to improvements in the published wording, these will not be reflected in this online version. Shortage of time makes it impossible for me to offer a more careful rendering. I hope that placing this imperfect version online may be useful to some readers, but they should note that the print version is definitive. I shall not let myself be held to the precise wording of an online version, where this differs from the print version.

Published in Philosophical Transactions of the Royal Society (Mathematical, Physical and Engineering Sciences) 358.1339–54, 2000.

The role of taxonomy in language engineering



Geoffrey Sampson


University of Sussex





Software engineering versus programming


The aim of this paper is to suggest that natural-language computing at present needs to take on board, more fully than it has done up to now, lessons which the wider IT profession learned some twenty to thirty years ago.  The lessons I have in mind were those that led to the creation of the discipline of software engineering.  Let me quote historical remarks from two standard textbooks:


The term “software engineering” was first introduced in the late 1960s at a conference held to discuss what was then called the “software crisis”. …  Early experience in building large software systems showed that existing methods of software development were not good enough.  Techniques applicable to small systems could not be scaled up.  Major projects were sometimes years late, cost much more than originally predicted, were unreliable, difficult to maintain and performed poorly.  Software development was in crisis.  (Sommerville 1992: 3)


In the middle to late 1960s, truly large software systems were attempted commercially. ... The large projects were the source of the realization that building large software systems was materially different from building small systems. ...  It was discovered that the problems in building large software systems were not a matter of putting computer instructions together.  Rather, the problems being solved were not well understood, at least not by everyone involved in the project or by any single individual.  People on the project had to spend a lot of time communicating with each other rather than writing code.  People sometimes even left the project, and this affected not only the work they had been doing but the work of the others who were depending on them.  Replacing an individual required an extensive amount of training about the “folklore” of the project requirements and the system design. ... These kinds of problems just did not exist in the early “programming” days and seemed to call for a new approach.  (Ghezzi et al. 1991: 4)


The new approach was what came to be called “software engineering”, which is nowadays a fundamental component of the training of computing professionals.


There are different ways of glossing the term “software engineering”, and I hope my definition will not seem objectionable to readers who are involved with it more centrally than I am, but one way of explaining the concept in a nutshell might be to call it a systematic training of computing professionals in resisting their natural instincts.


For most individuals who are attracted to working with computers, the enjoyable aspect of the work is programming, and running one’s programs.  Writing code, and seeing the code one has written make things happen, is fun.  (It is fun for some people, at any rate; it leaves others cold, but those others will look elsewhere for a career.)  Even inserting comments in one’s code feels by comparison like a diversion from the real business; programmers do it because they know they should, not out of natural inclination.  As for documenting a finished software system on paper, that is real punishment, to be done grudgingly and seeming to require only a fraction of the care and mental effort needed in coding, where every dot and comma counts.  What is more, these instincts were reinforced in the early years by the instincts of IT managers, who wanted objective ways of monitoring the productivity of the people under them, and quite inevitably saw lines of code per week as a natural measure.


These instincts seem to be widely shared, and they were often harmless in the early years, when software development was a small-scale, craft-like rather than industrial process where all the considerations relevant to a particular system might reside in a single head.  They led to crisis, once the scale of software projects enlarged, and required teamwork and integrity of software operation under different conditions over long periods of time.


Software engineering addresses that crisis by inverting computing professionals’ instinctive scale of values and sequence of activities.  Documentation, the dull part, becomes the central and primary activity.  Developing a software system becomes a process of successively developing and refining statements on paper of the task and intended solution at increasing levels of detail – requirements definitions, requirements specifications, software specifications; so that the programming itself becomes the routine bit done at the end, when code is written to implement specifications of such precision that, ideally, the translation should be more or less mechanical – conceptual unclarities that could lead to faulty program logic should be detected and eliminated long before a line of code is written.  Gerald Weinberg (1971) argued for a culture of “egoless programming”, which systematically deprives computing professionals of the pleasures of individual creativity and control over the programs for which they are responsible, as a necessary price to be paid for getting large systems which work as wholes.


Nobody suggests that now we have software engineering, all the problems described thirty years ago as “software crisis” have melted away and everything in the software development garden is rosy.  But I think few people in the IT industry would disagree that the counter-instinctive disciplines of software engineering are a necessary condition for successful software development, though those disciplines are often difficult to apply, and clearly they are not sufficient to ensure success.



How far we have come


Natural-language computing is not a new application of computer technology.  When Alan Turing drew up a list of potential uses for the stored-program electronic computer, a few weeks after the world’s first computer run at Manchester in June 1948, the second and third items on his five-item list were “learning of languages” and “translation of languages” (Hodges 1983: 382).  Some of the early machine translation projects must have been among the larger software development projects in any domain in the 1950s and early 1960s.  But on the whole natural-language computing has been late in making the transition from individualistic, craft activity to industrial process; and, where work was being done in a more realistic style, for instance on Peter Toma’s “Systran” machine-translation system (Hutchins & Somers 1992: chapter 10), for many years it was given the cold shoulder by computational linguists within the academic world (Sampson 1991: 127-8).


Since the 1980s, in some respects the subject has made great strides in the relevant direction.  It is hard, nowadays, to remember the cloistered, unrealistic ethos of natural-language computing as it was less than twenty years ago.  To give an impression of how things were then, let me quote (as I have done elsewhere) a typical handful of the language examples used by various speakers at the inaugural meeting of the European Chapter of the Association for Computational Linguistics, held at Pisa in 1983, in order to illustrate the workings of the various software systems which the speakers were describing:


Whatever is linguistic is interesting.


A ticket was bought by every man.


The man with the telescope and the umbrella kicked the ball.


Hans bekommt von dieser Frau ein Buch.


John and Bill went to Pisa.  They delivered a paper.


Maria é andata a Roma con Anna.


Are you going to travel this summer?  Yes, to Sicily.


Some critics of the field were unwilling to recognize such material as representing human language at all.  Michael Lesk (now Director, Information and Intelligent Systems, at the U.S. National Science Foundation) once characterized it acidly as “the imaginary language, sharing a few word forms with English, that is studied at MIT and some other research institutes” (Lesk 1988).  To me, there was nothing wrong with these dapper little example sentences as far as they went; but they were manifestly invented rather than drawn from real life, and they were invented in such a way as to exclude all but a small fraction of the problematic issues which confront software that attempts to deal with real-life usage.  Focusing on such artificial examples gave a severely distorted picture of the issues facing natural-language engineering.  Contrast the above examples with, at the other extreme, a few typical utterances taken from the structurally-annotated CHRISTINE Corpus[1] which my research group released this summer, based on real-life material extracted from the British National Corpus:[2]


well you want to nip over there and see what they come on on the roll


can we put erm New Kids # no not New Kids Wall Of # you know


well it was Gillian and # and # erm {pause} and Ronald’s sister erm {pause} and then er {pause} a week ago last night erm {pause} Jean and I went to the Lyceum together to see Arsenic and Old Lace


lathered up, started to shave {unclear} {pause} when I come to clean it there weren’t a bloody blade in, the bastards had pinched it


but er {pause} I don’t know how we got onto it {pause} er sh- # and I think she said something about oh she knew her tables and erm {pause} you know she’d come from Hampshire apparently and she # {pause} an- # an- yo- # you know er we got talking about ma- and she’s taken her child away from {pause} the local school {pause} and sen- # is now going to a little private school up {pause} the Teign valley near Teigngrace apparently fra-


Whatever IT application we have in mind, whether automatic information extraction, machine translation, generation of orthographically-conventional typescript from spoken input, or something else, I think the degree of complexity and difficulty presented by the second set of examples, compared with the first set, is quite manifest. 


Of course, I have made the point vivid by using examples drawn from spontaneous, informal speech (but then, notice that the last, at least, of the examples quoted from the Pisa meeting was clearly intended to represent speech rather than writing).  Some natural-language computing applications are always going to relate to written language rather than speech, and writing does tend to be more neatly regimented than the spoken word.  But even published writing, after authors and editors have finished redrafting and tidying it, contains a higher incidence of structural unpredictability and perhaps anarchy than the examples from the Pisa conference.  Here are a few sentences drawn at random from the LOB Corpus[3] of published British English:


Sing slightly flat.


Mr. Baring, who whispered and wore pince-nez, was seventy if he was a day.


Advice – Concentrate on the present.


Say the power-drill makers, 75 per cent of major breakdowns can be traced to neglect of the carbon-brush gear.


But he remained a stranger in a strange land.


In the first example we find a word in the form of an adjective, flat, functioning as an adverb.  In the next example, the phrase Mr. Baring contains a word ending in full stop followed by a word beginning with a capital which, exceptionally, do not mark a sentence boundary.  The third “sentence” links an isolated noun with an imperative construction in a logic that is difficult to pin down.  In Say the power-drill makers ... , verb precedes subject for no very clear reason.  The last example is as straightforward as the examples from the Pisa meeting; but, even in traditional published English, straightforward examples are not the norm.  (Currently, technologies such as e-mail are tending to make written language more like speech.)


There were no technical obstacles to real-life material of this sort being used much earlier than it was.  The Brown Corpus[4] of American English, which is proving to be a very valuable research resource even now at the century’s end, was published as early as 1964; for decades it was all but ignored.


For computational linguists to develop software systems based entirely on well-behaved invented data, which was the norm throughout the 1980s, is rather analogous to the home computer buff who writes a program to execute some intellectually-interesting function, but has little enthusiasm for organizing a testing régime which would check the viability of the program by exposing it to a realistically varied range of input conditions.  And this approach to natural-language computing militates against any application of statistical processing techniques.  Speakers of a natural language may be able to make up example sentences of the language out of their heads, but they certainly cannot get detailed statistical data from their intuitions.


One must learn to walk before one runs, and the 1980s reliance on artificial linguistic data might be excused on the ground that it is sensible to begin with simple examples before moving on to harder material.  In fact I think the preference of the discipline for artificial data went much deeper than that.  In the first place, as we have seen, computational linguistics was not “beginning” in the 1980s.  More important, almost everyone involved with linguistics was to a greater or lesser extent under the spell of the immensely influential American intellectual Noam Chomsky of MIT, who saw linguistics as more an aprioristic than an empirical discipline. 


One of Chomsky’s fundamental doctrines was his distinction between linguistic “performance” – people’s observable, imperfect linguistic behaviour – and linguistic “competence”, the ideal, intuitively-accessible mental mechanisms which were supposed to underlie that performance (Chomsky 1965: 4).  Chomsky taught that the subject worthy of serious academic study was linguistic competence, not performance.  The route to an understanding of linguistic performance could lie only through prior analysis of competence (op. cit.: 9, 15), and the tone of Chomsky’s discussion did not encourage his readers to want to move on from the latter to the former. 


For Chomsky, this concept of an ideal linguistic competence residing in each speaker’s mind was linked to his (thoroughly misguided) idea that the detailed grammatical structure of natural languages is part of the genetic inheritance of our species, like the detailed structure of our anatomy.[5]  But Chomsky was successful in setting much of the agenda of linguistics even for researchers who had no particular interest in these psychological or philosophical questions.  In consequence, if computational linguists of the 1980s noticed the disparity between the neatly-regimented examples used to develop natural-language processing software and the messy anarchy of real-life usage, rather than seeing that as a criticism of the examples and the software, they tended obscurely to see it as a criticism of real-life usage.  Aarts & van den Heuvel (1985) give a telling portrayal of the attitudes that were current in those years.  Not merely did most natural-language computing not use real-life data, but for a while there seemed to be an air of mild hostility or scorn towards the minority of researchers who did.


Happily, from about 1990 onwards the picture has completely changed.  Over the last ten years it has become routine for natural-language computing research to draw on “corpora”, machine-readable samples of real-life linguistic usage; and the validity of statistics-based approaches to natural-language analysis and processing is now generally accepted.  I am not sure that one can count this as a case of the profession being convinced by the weight of reasoned argument; my impression of what happened was that American research funding agencies decided that they had had enough of natural-language computing in the aprioristic style and used the power of the purse to impose a change of culture, which then spread across the Atlantic, as things do.  But, however it came about, the profession has now accepted the crucial need to be responsive to empirical data.



The lesson not yet learned


In another respect, though, it seems to me that natural-language computing has yet to take on board the software-engineering lesson of the primacy of problem analysis and documentation over coding.


I shall illustrate the point from the field of parsing – automatic grammatical analysis of natural language.  I believe similar things could be said about other areas of natural-language processing; but I am a grammarian myself, and automatic parsing is a key technology in natural-language computing.  Many would have agreed with K. K. Obermeier’s assessment ten years ago that parsing was “The central problem” in virtually all natural-language processing applications (Obermeier 1989: 69); more recently, I notice that “parsing” takes up more space than any other technology name in the index of an NSF/European Commission-sponsored survey of natural language and speech computing (Cole et al. 1997).[6]  As these pointers suggest, a large number of research groups worldwide have been putting a lot of effort into solving the parsing problem for years and indeed for decades.  Many parsing systems have been developed, using different analytic techniques and achieving different degrees of success.


Any automatic parser is a system which receives as input a representation of a spoken or written text, as a linear sequence of words (together possibly with subsidiary items, such as punctuation marks in the case of written language), and outputs a structural analysis, which is almost always in a form notationally equivalent to a tree structure, having the words of the input string attached to its successive leaf nodes, and with nonterminal nodes labelled with grammatical categories drawn from some agreed vocabulary of grammatical classification.  (A minority of research groups working on the parsing problem use output formalisms which deviate to a certain extent from this description, for instance the “dependency” notation due to Lucien Tesnière (1959), but I do not think these differences are significant enough to affect the substance of the point I am developing.)  The structural analysis is something like a representation of the logic of a text, which is physically realized as a linear string of words because the nature of speech forces a one-dimensional linear structure onto spoken communication (and writing mimics the structure of spoken utterances).  So it is easy to see why any automatic processing which relates to the content of spoken or written language, rather than exclusively to its outward form, is likely to need to recover the tree-shaped structures of grammar underlying the string-shaped physical signals.


Obviously, to judge the success of any particular parser system, one must not only see what outputs it yields for a range of inputs, but must know what outputs it should produce for those inputs: one must have some explicit understanding of the target analyses, against which the actual analyses can be assessed.  Yet it was a noticeable feature of the literature on automatic natural-language parsing for many years that – while the software systems were described in detail – there was hardly any public discussion of the schemes of analysis which different research groups were treating as the targets for their parsing systems to aim at.  Issues about what counted as the right analyses for particular input examples were part of what Ghezzi et al. (see above) called “the ‘folklore’ of the project requirements”.  Members of particular parsing projects must have discussed such matters among themselves, but one almost never saw them spelled out in print.


Of course, unlike some of the topics which software is written to deal with, natural-language parsing is a subject with a long tradition behind it.  A number of aspects of modern grammatical analysis go back two thousand years to the Greeks; and the idea of mapping out the logic of English sentences as tree structures was a staple of British schooling at least a hundred years ago.  So computational linguists may have felt that it was unnecessary to be very explicit about the targets for automatic parsing systems, because our shared cultural inheritance settled that long since.


If people did think that, they were wrong.  The wrongness of this idea was established experimentally, at a workshop held in conjunction with the Association of Computational Linguistics annual conference at Berkeley, California, in 1991.  Natural-language processing researchers from nine institutions were each given the same set of English sentences and asked to indicate what their respective research groups would regard as the target analyses of the sentences, and the nine sets of analyses were compared.  These were not particularly complicated or messy sentences – they were drawn from real-life corpus data, but as real-life sentences go they were rather well-behaved examples.  And the comparisons were not made in terms of the labels of the constituents:  the only question that was asked was how far the researchers agreed on the shapes of the trees assigned to the sentences – that is, to what extent they identified the same subsequences of words as grammatical constituents, irrespective of how they categorized the constituents they identified.


The level of agreement was strikingly low.  For instance, only the two subsequences marked by square brackets were identified as constituents by all nine participants in the following example (and results for other cases were similar):


One of those capital-gains ventures, in fact, has saddled him [ with [ Gore Court ] ].


If specialists agree as little as this on the details of what parsing systems are aiming to do, that surely establishes the need for a significant fraction of all the effort and resources that are put into automatic parsing to be devoted to discussing and making more publicly explicit the targets which the software is aiming at, rather than putting them all into improving the software. 



The scale of the task


I do not mean to imply that every natural-language computing group working on English ought to agree on a single common parsing scheme.  In the context of applications executing commercially or socially valuable natural-language processing functions of various kinds, automatic parsing is only a means to an end.  It may well be that the kind of structural analysis which is most appropriate with respect to one function differs in some details from the analysis that is appropriate for an application executing a different function.  But the lack of agreement revealed at the 1991 workshop did not arise because various research groups had made explicit decisions to modify the details of a recognized public scheme of English-language parsing to suit their particular purposes.  No such public scheme existed.  Separate groups were forced to use different parsing schemes, because each research group had to develop its own standards, as a matter of internal project “folklore”.  The analytic concepts which we inherit from traditional school grammar teaching may be fine as far as they go, but they are far too limited to yield unambiguous, predictable structural annotations for the myriad linguistic constructions that occur in real life.


And, because research groups developed their parsing standards independently and in an informal fashion, not perceiving this as truly part of the work they were engaged on, they were in no position to develop schemes that were adequate to the massive structural complexity of any natural language.  The results of the 1991 ACL workshop experiment came as little surprise to me, in view of earlier experiences of my own.  From 1983 onwards, as a member of the University of Lancaster natural-language computing group, I had taken responsibility for creating a written-English “treebank” – a sample of structurally-annotated real-life sentences[7] – which was needed as a resource for a statistics-based parsing project led by my senior colleague Geoffrey Leech.  I remember that when I took the task on and we needed to agree an annotation scheme for the purpose, Leech (who knows more about English grammar than I ever shall) produced a 25-page typescript listing a set of symbols he proposed that we use, with guidelines for applying them in debatable cases; and I thought this represented such a thorough job of anticipating problematic issues that it left little more to be said.  All I needed to do was to use my understanding of English in order to apply the scheme to a series of examples. 


I soon learned.  As I applied the scheme to a sample of corpus data, the second or third sentence I looked at turned out to involve some turn of phrase that the typescript did not provide for; as I proceeded, something on the order of every other sentence required a new annotation precedent to be set.  Written items like names, addresses, money sums, weights and measures have linguistic structure of their own; the grammatical tradition says little about them, so one has to make new decisions about how to represent that structure.  But plenty of decisions are needed also in the more linguistically “central” areas of clause and phrase analysis.  Often, alternative structural annotations of a given construction each seemed perfectly defensible in terms of the school grammatical tradition – but if we were going to use our treebank to produce meaningful statistics, we had to pick one alternative and stick to it. 


Consider, to give just one example, the construction exemplified in the more, the merrier – the construction that translates into German with je and desto.  Here are three ways of grouping a sentence using that construction into constituents:


[ [ the wider the wheelbase is ], [ the more satisfactory is the performance ] ]


[ [ the wider the wheelbase is ], the more satisfactory is the performance ]


[ [ [ the wider the wheelbase is ], the more satisfactory ] is the performance ]


The two clauses might be seen as co-ordinated, as in the first line, since both have the form of main clauses and neither of them contains an explicit subordinating element.  Or the second clause might be seen as the main clause, with the first as an adverbial clause adjunct.  Or the first clause might be seen as a modifier of the adjectival predicate within the second clause.  There seemed to be no strong reason to choose one of these analyses rather than another.


Linguists influenced by the concept of innate psychological “competence” tend to react to alternatives like this by asking which analysis is “true” or “psychologically real” – which structure corresponds to the way the way the utterance is processed by speaker’s or hearer’s mental machinery.  But, even if questions like that could ultimately be answered, they are not very relevant to the tasks confronting natural-language computing here and now.  We have to impose analytic decisions in order to be able to register our data in a consistent fashion; we cannot wait for the outcome of abstruse future psychological investigations. 


Indeed, I should have thought it was necessary to settle on an analytic framework in order to assemble adequately comprehensive data for the theoretical psycholinguists to use in their own investigations.  In biology, the Linnaean binomial classification system (which Linnaeus and everyone else knew to be unnatural, but was practical to apply) was a prior requirement for the development of modern theories of cladistics.  A science is not likely to be in a position to devise deep theories to explain its data, before it has an agreed scheme for identifying and registering those data.  To use the terms “true” and “false” in connexion with a scheme of grammatical annotation would be as inappropriate as asking whether the alphabetical order from A to Z which we use for arranging names in a telephone directory or books on shelves is the “true” order.


At any rate, within the Lancaster group it became clear that our approach to automatic parsing, in terms of seeking structures over input word-strings which conformed to the statistics of parse configurations in a sample of analysed material, required us to evolve far more detailed analytic guidelines than anything that then existed; without them, the statistics would be meaningless, because separate instances of the same construction would be classified now one way, now another.  We evolved a routine in which each new batch of sentences manually parsed would lead to a set of tentative new analytic precedents which were logged on paper and circulated among the research team; weekly or fortnightly meetings were held where the new precedents were discussed and either accepted or modified, for instance because a team member noticed a hidden inconsistency with an earlier decision.  The work was rather analogous to the development of the Common Law.  A set of principles attempts to cover all the issues on which the legal system needs to provide a decision, but human behaviour continually throws up unanticipated cases for which the existing legal framework fails to yield an unambiguous answer; so new precedents are set, which cumulatively make the framework increasingly precise and comprehensive.  We want our nation’s legal system to be consistent and fair, but perhaps above all we want it to be fully-explicit; and if that is possibly not the dominant requirement for a legal system, it surely is for a scientific system of data classification.  To quote Jane Edwards of the University of California at Berkeley: “The single most important property of any data base for purposes of computer-assisted research is that similar instances be encoded in predictably similar ways” (Edwards 1992: 139).


Ten years of our accumulated precedents on structural annotation of English turned a 25-page typescript into a book of 500 large-format pages (Sampson 1995).  Beginning from a later start, the Pennsylvania treebank group published their own independent but very comparable system of structural annotation guidelines on the Web in the same year (Bies et al. 1995). I am sure that the Pennsylvania group feel as we do, that neither of these annotation schemes can be taken as a final statement; the analogy with the growth of the law through cumulation of precedents suggests that there never could be a last word in this domain.  My own group has been elaborating our scheme in the last few years by applying it to spontaneous speech; but although the main focus here is on aspects of the annotation scheme that were irrelevant to the structure of written prose, for instance mechanisms for marking what is going on when speakers edit their utterances “on the fly”, we continue to find ourselves setting new precedents for constructions that are common to writing as well as speech. (In due course we plan to cumulate them into a supplement to the 1995 book.)



Differential reception of data and specifications


The only way that one can produce an adequate scheme of structural annotation is to apply an initial scheme to real data and refine the scheme in response to problem cases, as we have been doing; so in developing an annotation scheme one inevitably generates a treebank, an annotated language sample, as a by-product.  The Lancaster treebank which started me on this enterprise in the mid-1980s was for internal project use and was never published, but I did publish electronically the annotated samples on which later stages of annotation-scheme development were based.  This “SUSANNE Corpus”, as it is called, was released in successively more accurate versions between 1992 and 1994.  Part of the point I am seeking to make in the present paper can be illustrated by the different receptions accorded by the research community to the SUSANNE Corpus, and to the published definition of the SUSANNE annotation scheme. 


Because it emerged from a manual annotation process which aimed to identify and carefully weigh up every debatable analytic issue arising in its texts, the SUSANNE Corpus is necessarily a small treebank; there is a limit to how reliable any statistics derived from it can hope to be.  Yet it has succeeded far beyond my expectations in establishing a role for itself internationally as a natural-language computing research resource.  Accesses to the ftp site originally distributing it at the Oxford Text Archive quickly rose to a high level (and subsequently other “mirror” sites began distributing it, so that I no longer have any way of monitoring overall accesses).  I quite often encounter in the professional literature references to research based on the SUSANNE Corpus, commonly by researchers of whom I had no prior knowledge. 


Conversely, the book defining the annotation scheme has found no role that I have detected.  Reviewers have made comments which were pleasing to read, but almost no-one has spontaneously found reasons to get into correspondence about the contents of the annotation scheme, in the way that many researchers have about the SUSANNE treebank – indeed, it often becomes apparent, when people who have been working intensively with the SUSANNE Corpus get in touch, that they have never looked at the published definition of the SUSANNE annotation scheme on which the corpus is based.  My guess is that the only place where that reference book is in routine day-to-day use is in my own research group at Sussex.[8]


Now, like every academic, I am naturally quite delighted to find that any research output for which I was responsible seems to be meeting a need among the international research community.  The welcome that the Corpus alone has received is certainly more than a sufficient professional reward for the effort which created Corpus and annotation scheme.  Nevertheless, I find the imbalance in the reception of the two resources rather regrettable in what it seems to say about the values of the discipline.  In my own mind, the treebank is an appendix to the annotation scheme, rather than the other way round; the treebank serves a function similar to what I believe biologists call a type collection attached to a biological taxonomy – a set of specimens intended to clarify the definitions of the taxonomic classes.  The SUSANNE treebank is really too small to count as a significant database of English grammatical usage; whereas the published annotation scheme, although it unquestionably has many serious limitations and imperfections, can (I believe) claim to be a more serious attempt to do its own job than anything that existed in print before.  If the research community is not taking up the SUSANNE annotation scheme as a basis from which to push forward the enterprise of taxonomizing English structure, that could just mean that they prefer the Pennsylvania scheme as a starting point for that work; but in fact I do not get the impression that this sort of activity has been getting under way in connexion with the Pennsylvania scheme either.  (The fact that the Pennsylvania group limited themselves to publishing their scheme via the Web rather than as a printed book perhaps suggests that they did not expect it to.)


When Geoffrey Leech began to look for support to create the first corpus of British English, about thiry years ago, I understand that funding agencies were initially unreceptive, because at that time a simple collection of language samples did not strike reviewers as a valid research output.  People expected concrete findings, not just a collection of data from which findings could subsequently be generated – although Leech’s LOB Corpus, when it was eventually published in 1978, served as the raw material for a huge variety of research findings by many different researchers, which collectively must far exceed the new knowledge generated by almost any research project which seeks to answer a specific scientific question. 


We have won that battle now, and it is accepted that the compilation of natural-language corpora is a valuable use of research resources – though now that massive quantities of written language are freely available via the Internet, the need at the end of the century is for other sorts of language sample, representing speech rather than writing and/or embodying various categories of annotation.  But there is still a prejudice in favour of the concrete.  When I put together a new research proposal, I couch it in terms of compiling a new annotated corpus, rather than extending and testing a scheme of structural annotation.  If I wrote the proposals in the latter way, I am convinced they would fail, whereas research agencies are happy to sponsor new corpora even though (given our method of working) the ones I can offer to create are very small.  Before software engineering brought about a change of vision, IT managers measured their colleagues’ output in terms of lines of code, and overlooked the processes of planning, definition, and co-ordination which were needed before worthwhile code could be written.  At present, most computational linguists see the point of an annotated corpus, but few see the point of putting effort into refining schemes of annotation.


Some encouragement to give more priority to the annotation-scheme development task has come from the (perhaps unexpected) direction of the European Commission, whose Directorate-General XIII induced the predominantly US-sponsored Text Encoding Initiative[9] to include a small amount work on this area about ten years ago, and more recently established the EAGLES group[10] to stimulate the development of standards and guidelines for various aspects of natural-language computing resources, including structural annotation of corpora. 


The EAGLES initiative has produced valuable work, notably in the area of speech systems, where the relevant working group has assembled between hard covers what looks to me like a very complete survey of problems and best practices in various aspects of speech research (Gibbon et al. 1997). But in the area of grammatical annotation the EAGLES enterprise was hobbled by the obvious political necessity for EU-funded work to deal jointly with a large number of European languages, each of which has its own structure, and which are very unequal in the extent to which they have been worked over by either traditional or computer-oriented scholarly techniques (many of them lagging far behind English in that respect).  Consequently, in this domain the EAGLES initiative focused on identifying categories which are common to all or most EU national languages, and I think it is fair to say that its specific recommendations go into even less detail than the inherited school grammar tradition provides for English.  The nature of the EAGLES enterprise meant that it could hardly have been otherwise. 


What is needed is more effort devoted to identifying and systematically logging the fine details of spoken and written language structure, so that all aspects of our data can be described and defined in terms which are meaningful from one site to another, and this has to be done separately for any one language in its own terms (just as the taxonomy of one family of plants is a separate undertaking from the taxonomy of any other family).  European languages obviously do share some common structural features because of their common historical origins and subsequent contacts; but a language adapts its inherited stock of materials to new grammatical purposes on a time-scale of decades – think for instance of the replacement of might by may in the most respectable written contexts just within the last ten or twenty years, in constructions like if he had been in Cornwall he may have seen the eclipse – whereas the EU languages have developed as largely independent systems for millennia.  We do not want our grammatical classification systems to be excessively dominated by ancient history. 


In developing predictable guidelines for annotating the structure of spontaneous spoken utterances, my group faced large problems stemming from the fact that, within English, there are different groups of speakers who, for instance, use the verb system in different ways.  If a speaker of a nonstandard version of English says she done it, rather than she did it or she’s done it (which speakers very often do), to a schoolteacher this may represent heresy to be eradicated, but for us it is data to be logged.  We have to make a decision about whether such cases should be counted as:  simple past forms with nonstandard use of done rather than did as past tense of do; perfective forms with nonstandard omission of the auxiliary; or a third verbal category, alongside the perfective and simple past categories of the standard language.  The idea of developing guidelines at this level of detail which simultaneously take into account what happens in German or Modern Greek is really a non-starter.


In any case, encouragement from national or supranational government level will not achieve very much, unless enthusiasm is waiting to be kindled at grass-roots level among working researchers.  Natural-language computing researchers need to see it as just as fascinating and worthwhile a task to contribute to the identification and systematic classification of distinctive turns of phrase as to contribute to the development of language-processing software systems – so that taxonomizing language structure becomes an enterprise for which the discipline as a whole takes responsibility, in the same way as biologists recognize systematics as an important subfield of their discipline.  The fact that natural-language computing is increasingly drawing on statistical techniques, which by their nature require large quantities of material to be registered and counted in a thoroughly consistent fashion, makes the task of defining and  classifying our data even more crucial that it was before.  It is surely too important to leave in the hands of isolated groups in Sussex or Pennsylvania.



The fractal analogy


If people are attracted to the task, there is plenty of work for them to do.  Richard Sharman, of SRI International, Cambridge, has likened natural languages to fractal objects, in the sense that there is always more structural detail to be revealed as one looks at them more closely.  I carried out a statistical analysis on the first treebank that we developed at Lancaster in the mid-1980s which suggested that this analogy may be rather exact.[11]  For one high-frequency category of phrase (the noun phrase), I looked at the frequencies of each alternative realization in the treebank.  That is, I listed the various sequences of daughter labels found on nodes immediately dominated by a mother node labelled “noun phrase”, using a fairly coarse alphabet of grammatical category labels, and I counted the number of times each distinct daughter-label sequence recurred in the data.  The total size of the treebank was not large, for the reasons I have already discussed.  But, within the limits of that data-set, there proved to be a rather precise relationship between numbers and frequencies of different constructions (that is, different ways of realizing the given mother category as a sequence of daughter categories).  As one looked at constructions with lower and lower frequencies, the number of different constructions each occurring at those frequencies grew, in a regular way, so that constructions which were individually very rare were collectively quite common.  Putting it formally:


if m is the frequency of the commonest single construction (in my data, m was about 28 per thousand words)


and f is the relative frequency of some construction (fm is its absolute frequency)


then the proportion of all construction-tokens which represent construction-types of relative frequency less than or equal to f is about


The significance of this finding is that it contradicts the widely-shared picture of a natural language as containing a limited number of “competent” grammatical structures, which in practice are surrounded by a penumbra of more or less random, one-off “performance errors” – and it is this picture, I believe, that has done much to discourage linguists (even many linguists who would claim to disagree with Noam Chomsky’s theories) from seeing structural taxonomy as a worthwhile activity. 


If the competence v. performance picture were correct, then identifying the “competent” constructions would be a task more akin to defining the syntax of Pascal or Java than to classifying the genera and species of the family Compositae; and investigating the “performance errors” would be a fairly unattractive task to most researchers: mistakes are mistakes.  But that picture seems to imply that construction frequencies ought to be distributed bimodally, with competent constructions occurring at reasonably high frequencies, individual performance errors each occurring at a low frequency, and not much in between.  My data were not like that; constructions were distributed smoothly along the scale of frequencies, with no striking gaps.  (It would in any case be surprising to find that “performance error” was an important factor in this treebank, which was based on published writing.)


Consider what the mathematical relationship I have quoted would mean, if it continued to hold true in much larger data sets that I was in a position to investigate.  (I cannot know that it does, but as far as they went my data contained no suggestion that the relationship broke down at the lower frequencies.)  If the relationship held true in larger bodies of data, then:


      one in every ten construction-tokens would represent a type which occurs at most once per 10,000 words


      one in every hundred construction-tokens would represent a type which occurs at most once per 3.5 million words


      one in every thousand construction-tokens would represent a type which occurs at most once per billion words


One per cent of all construction-tokens is surely far too high a proportion of language material for natural-language computing to wash its hands of as too unusual to deal with.  One might feel that one in a thousand is still too high a proportion for that treatment.  Yet, if we have to search millions, or even hundreds of millions, of words of text to find individual examples, we seem to be a long way away from the picture of natural-language grammars as limited sets of rules, like programming languages with a number of irregularities and eccentric features added.  A natural fractal object such as a coastline can never be fully described – one has to be willing to ignore detail below some cut-off.  But the degree of detail which natural-language computing ought to be taking into account extends well beyond the range of structures described in standard grammars of English.





At the end of the twentieth century, mankind’s honeymoon with the computer has not yet quite faded, and software development still has a glamour which is lacking in research that revolves round ink and paper.  But a computational linguist who helps to develop a natural-language software system is devoting himself to a task which, realistically, is unlikely to achieve more than a tiny advance on the current state of the art, and will quickly be forgotten when another, better system is produced.  To improve our system for registering and classifying the constructions of English, on the other hand, is to make a potentially lasting contribution to our knowledge of the leading medium of information storage and exchange on the planet.  Of course I do not suggest that computational linguists should migrate en masse from the former activity to the latter, but it would be good to see more of a balance. 


Some researchers perhaps feel that the kind of detailed study of natural-language structure which I have been advocating belongs to the domain of the humanities rather than science, and consequently is not for them.  But arts-based researchers are not inclined towards the work of imposing rigid and sometimes artificial classificatory divisions on inherently continuous clines which is needed as a prerequisite for gathering large amounts of quantitative data on a consistent basis.  The flavour of the work makes it more akin to other IT activities than to the humanities, and it is computer-oriented researchers who have a motive for engaging in it.  I hope that more of them will begin to do so.



Aarts, J. & T. van den Heuvel (1985)  “Computational tools for the syntactic analysis of corpora”.  Linguistics 23.303-35.

Bies, Ann, et al. (1995)  Bracketing Guidelines for Treebank II Style,

Chomsky, A.N. (1965)  Aspects of the Theory of Syntax.  MIT Press (Cambridge, Mass.).

Chomsky, A.N. (1976)  Reflections on Language.  Temple Smith.

Cole, R., et al., eds. (1997)  Survey of the State of the Art in Human Language Technology.  C.U.P.

Edwards, Jane (1992)  “Design principles in the transcription of spoken discourse”.  In Svartvik (1992), pp. 129-44.

Ghezzi, C., et al. (1991)  Fundamentals of Software Engineering.  Prentice-Hall.

Gibbon, D. et al., eds. (1997)  Handbook of Standards and Resources for Spoken Language Systems.  Mouton de Gruyter (Berlin).

Hodges, A. (1983)  Alan Turing: The Enigma of Intelligence.  Hutchinson; my page reference is to the Unwin Paperbacks ed., 1985.

Hutchins, W.J. & H.L. Somers (1992)  An Introduction to Machine Translation.  Academic Press.

Lesk, M. (1988)  Review of R.G. Garside et al., eds., The Computational Analysis of English.  Computational Linguistics 14.90-1.

Obermeier, K.K. (1989)  Natural Language Processing Technologies in Artificial Intelligence: The Science and Industry Perspective.  Ellis Horwood.

Pinker, S. (1994)  The Language Instinct: The New Science of Language and Mind.  William Morrow (New York); Penguin ed., 1995.

Sampson, G.R. (1987)  “Evidence against the ‘grammatical’/‘ungrammatical’ distinction”.  In W. Meijs, ed., Corpus Linguistics and Beyond, Rodopi (Amsterdam), pp. 219-26. [Reprinted as ch. 10 of Sampson, Empirical Linguistics.]

Sampson, G.R. (1991)  “Natural language processing”.  Chapter 8 of C. Turk, ed., Humanities Research Using Computers.  Chapman & Hall.

Sampson, G.R. (1992)  Probabilistic parsing”.  In Svartvik (1992), pp. 425-47.

Sampson, G.R. (1995)  English for the Computer: The SUSANNE Corpus and Analytic Scheme.  Clarendon Press (Oxford).

Sampson, G.R. (1997)  Educating Eve: the “Language Instinct” Debate.  Cassell; revised ed., 1999.

Sommerville, I. (1992)  Software Engineering (4th ed.).  Addison-Wesley (Wokingham, Berks.).

Svartvik, J., ed. (1992)  Directions in Corpus Linguistics: Proceedings of Nobel Symposium 82.  Mouton de Gruyter (Berlin).

Tesnière, L. (1959)  Eléments de syntaxe structurale.  Klincksieck (Paris).

Weinberg, G. (1971)  The Psychology of Computer Programming.  Van Nostrand Reinhold (New York).








[5] Chomsky expounded this doctrine of linguistic nativism in books such as Reflections on Language (Chomsky 1976); it has been popularized recently by Steven Pinker’s The Language Instinct (Pinker 1994).  I have pointed out the vacuousness of Chomsky’s and Pinker’s various arguments for linguistic nativism in my Educating Eve (Sampson 1997).

[6]Only general concepts such as corpora, dialogue, speech, word occupy larger sections of Cole et al.’s index.

[7] I believe that the term “treebank” was first coined in this sense by Geoffrey Leech in connexion with our Lancaster project.  It has subsequently become current internationally.

[8] One of the organizers of the Royal Society Discussion Meeting informs me that this guess is not wholly correct.  Nevertheless, the difference between reception of the scheme and reception of the data is striking.


[10] The “Expert Advisory Group on Language Engineering Standards”,

[11] The analytic findings are presented in Sampson (1987); see Sampson (1992: 440-5) for discussion of a critique of my conclusions by Briscoe and others.