Hans Karlgren
But pausing in our scientific pursuit, as we do during a panel discussion, we might be justified in standing back to see what effects, if any, our studies have on the object studied. Intended or not intended, scientific studies and results do have an effect, and those involved would be the first to know. And the first to provide early warnings.
Danger occurs only when the restrictions are not deliberately made, e.g., if it is overseen that human communicants do not only have a larger vocabulary than that of the present experimental set-up but also always one which is and remains incomplete. And if it is not recognized that well-formed input is a special case of communication, the idea of the fundamental separation of our linguistic "competence" and our ability to handle incomplete and distorted messages will be continuously and unduly corroborated.
In engineering, of course, we want machines to perform well, but successful software execution does not necessarily mean success for an experimental system. After all, computational linguistics is a way of studying language, or ourselves as creatures who interact and think using language. Whenever we can reduce human behaviour to mechanical procedures, we gather urgent, not seldom painful knowledge about ourselves; but when we fail, and recognize why, we may learn even more. It is not only in thermodynamics that the great failures mark the fundamental advances. Unspecified input manipulation or selection will rob us of the most instructive failures.
To begin with, today's technology does not enable machines to understand "natural language" as used by innocent people. We have to restrict it. Our choice is whether we admit the restrictions or hide them. Many systems claim outright that they accept natural language, thus avoiding the crucial question how to tell the users what are the restrictions and in fact often offering a seductively English-like - or Japanese-like etc as the case may be - unspecified superset of an unspecified subset of what native English-speakers would have used in comparable situations. Is it really easier for a human to learn restrictions on something confusingly similar to normal language than to learn an explicitly arbitrary problem-oriented code? The answer depends on the applications and is rarely trivial.
No doubt machine analysis can be made much simpler and safer if we grant ourselves the right to redefine the input language to make it manipulatable. Minor restraints on the input make processing much easier.
Do they? If the restraints are really "minor"?
Restraints on form are easily enforced and seem easy to live with for a user of an interactive system.
Thus, if queries directed to a retrieval system as well as the specifications of the objects of the universe to be retrieved from must be stated with a controlled vocabulary of indexing terms, one can reduce the matching of query and object specification to finding identities instead of evaluating similarities. Any unauthorized term utilized by indexer or user will be immediately rejected. The simplification is successful - at a premium of course: nicety in description is lost, making selections less appropriate, and the system will not grow and adapt itself smoothly to new ideas and manners of expression but will age so that it will have to be revised or replaced now and then.
Similarly, restraints on sentence length or syntactic types are not hard to uphold and do not cause very great inconvenience. Nor, unfortunately, do they eliminate the most crucial computational problems. To have a significant effect on computational feasibility, restraints must eliminate the need for disambiguation via long-range context.
Now, if we put semantic restrictions on what looks like natural language, are these still minor restraints? If the interpretation of permitted expressions is changed, how do we humans learn the new semantics? How do we learn that familiar expressions have acquired another value?
Is it easy to remember that "very big" has a different meaning in an almost-natural language from which the word "rather" has been purged out? Do we readily unlearn to recognize coordination in English-like sentences except where there is an explicit "and" or "or"? Can we get accustomed to reading "or" in the non-exclusive sense in every context so that "Wednesday or Thursday" necessarily covers "Wednesday and Thursday"? I doubt it: the very familiarity of the words will lead us astray. And can we inhibit our habits to expect a dialogue partner to react cooperatively on the presuppositions, omissions and ambiguities of which we are guilty?
When are such quasi-natural languages, with their attractive mnemotechnical features, easier to learn than overtly artificial languages? To learn well?
When we address the machines of today they do not often let us forget that we are talking to heartless fault-finding things, but when their surface has been polished a few degrees more, what happens then? Paradoxically, the more intelligently and cooperatively a machine is programmed to behave, the more deceptive it becomes.
One of the most ambitious modern attempts at an artificially restricted language is the AECMA Simplified English adopted by a large group of European aircraft manufacturers for use in maintenance manuals. This "English" is claimed to be easier to handle in machine translation but its primary purpose is to make the technical documents clearly understandable to a wider circle of non-native English readers. - A similar Rationalized French has been developed for the same industry.
The AECMA language was discussed at recent workshops arranged by the Linguistics in Documentation Committee within the International Federation for Information and Documentation ("FID"), in meetings hosted by Kval Instititute for Information Science, in Mariehamn in the Åland Isles in 1987 and in Malta in 1988. The properties of the language were presented by Erik Lebreton of the Itep translation organization, Toulouse et al., France. Cf. the committee's report Industrial Product Documentation, in International Forum on Information and Documentation 1988:1, Moscow 1988.
This language is essentially a subset of natural English as used in aircraft maintenance manuals. The restraints are so strong that a significant amount of training is necessary for a person proficient in English to learn to write, but the claim is that any reader of normal English can read AECMA texts.
The restrictions on vocabulary are heavy. In addition to an open class of well-defined technical terms of the trade, only members of a listed vocabulary are legitimate, and these are to be used only in prescribed word classes and meanings.
Thus, "mass" can be used only as a noun and only in the sense "A value that shows how resistant an object is to acceleration". That is not how I spontaneously perceive this frequent English word, but the specialization seems fair enough in technical manuals. Other rulings are harder to swallow.
Some lexical restraints affect the semantics. The words "below" and "behind" must refer to position: "Gauge 3 should not read more than 1.5 psi below gauge 2" should be replaced by "Gauge 3 must not read 1.5 psi less than gauge 2".
Other words which could cause ambiguity are banned altogether, to be replaced by words with a unique interpretation. For "beyond" use "more", the instructions say, suggesting by way of illustration that "Components that are worn beyond their maximum limits" should be replaced by "Components that have more wear than the maximum limits".
Structural restraints are coarse but come close to common sense rules for easy texts: sentences must be no longer than 24 words, paragraphs no longer than 6 sentences, etc.
Many of the recommendations seem to have a general healthy effect on style. Many phrases, though cumbersome, become unambiguous. And non-native readers are freed of many obstacles. It is clearly easier for us to read "The movement forward of the control lever must be ..." than "The advance of the control lever must be ...".
The AECMA language has been found to be harder to learn to write and less acceptable to native English-speakers than to those for whom English is a second language, whether they have a good or poor command of it. Some phrases land a little bit outside the border-line of grammatical English, and it is in the nature of the whole scheme that one will deviate from idiomatic English usage. It is a moot question whether AECMA is, strictly speaking, a subset of English.
Thus, since "test" is an AECMA noun but not an AECMA verb, one cannot say "to test" but has to say "to do a test", which sounds complicated to the English. The word "action" is not in the vocabulary so for " This action must not be carried out if the aircraft ..." the instruction book recommends "You must not do this procedure if the aircraft ...".
These and similar features make AECMA texts sound awkward, ugly and complicated (!) to natives, and they certainly cannot boast elegance or conciseness.
However, native English speakers seem to underestimate the degree to which we non-native English-users - who, incidentally, constitute the majority of all English-users - are inconvenienced by lexical ambiguity, above all by the multiple word classes of English words.
Substitutions for, e.g., "once", "inside" and "next" surely do not yield texts which are easier to digest for uneducated natives: "The pump is adjacent to the..." for "The pump is next to the ...". But structures where "once" is a conjunction and "inside" a preposition are really confusing to us non-natives, who have to run the sentence a few more iterations through our brains before it occurs to us that "once" which we trusted to be an adverb is here functioning as a conjunction and that "inside" which has all appearances of being a noun has suddenly turned itself into a preposition. And few natives recognize how gratifying are simple words with a unique function, like "adjacent"; its Latin roots and low frequency in everyday English are no drawback for us.
Even "to do a test", though admittedly longer than "to test" does feel helpful; most of us are not accustomed from our home languages to handle more than an occasional blur of the noun/verb dichotomy. The insertion of a dummy do-verb does not disturb us overly, once we have reconciled ourselves to the odd fact that English, whether simplified or not, inserts a "do" for which we do not feel any necessity in every negative and interrogative clause. The circumlocution "to do a test" does not strike us as more absurd than does "I do not see" for "I see not".
But there are more fundamental shortcomings. During the discussions of the AECMA language the experience was noted that texts in this language tended to look easy and offer little resistance for cursory reading by persons with a weak command of English but that they often do not pay up when pressed for a precise answer to a specific question. The smoothness is often bought at the price of vagueness. This feeling of evasiveness is disconcerting to technical readers, whether good or poor in English, and is particularly obstructive for translators when trying to translate out of the AECMA language.
If this is a general and inherent feature of the language, it is of course a fundamental objection to the whole scheme. It means that the obstacles have been moved from mere vocabulary and syntax, which readers can, after all, learn to master after a while, to text interpretation, which is a permanent source of nuisance and failure. It may be, however, that writers can learn to use the language more effectively and that slight modifications of its rules can mitigate these effects. The area should be given serious attention by linguists, applying what we know of grammar, text linguistics and translation theory.
The next attempt at making a strategically simplified English for some domain is likely to be cleverer, built as it will be on longer experience and, hopefully, exploiting more of linguistic insight into what constitute the fundamental difficulties. The fact that a constructed language will take time to learn - learn to write if not to read -, that it will feel awkward at first to everybody and possibly remain revolting to stylistically sensitive natives, that it will convey attitudes and enthusiasm less effectively and even be less clarifying in matter than good English texts are to educated natives - all these I find are sometimes acceptable. They are not an unreasonable price to pay if simplification does make reading easier for some humans and all machines and thus permits more mechanical processing, including none-too-misleading fully automated translation into other languages.
We are confronted with some genuine questions: what simplifications are necessary to make a language essentially more processable? What kinds of simplifications are we prepared to tolerate in order to smooth the way for machines and non-native readers? How far human behaviour and technology should each go to meet the other is, of course, not a scientific issue: it is a matter of human engineering and social tolerance as well as of ingenuity of technical construction. It is not certain that the problem becomes easier if linguistic computation is improved. Almost-human language is perhaps in the long run a worse intruder than square artefacts.
We can conclude that it is possible to introduce strong restraints which make mechanical processing easier, that many restraints are tolerable for human users and quite a few of them justifiable as simplifications for any reader, whether man or machine.
Systematic control of human language is more accepted in some countries than in others. For centuries, royal or otherwise highly placed academies have produced dictionaries, grammars and advice with normative intentions in many countries. These attempts at creating or preserving "The Good Language" have had effect primarily on small groups with literary ambitions, and have given rise to counter-reactions among many others. In the U.S. most people seem to reject institutionalized cultivation of language as interference into a very private sphere. In later years, however, systematic terminological efforts in international cooperation on the principles of industrial standardization has had a significant impact on every language in the industrialized world.
In Europe, we have a long tradition of systematic control of language, particularly technical language. We should mention in the first hand the large-scale and well-informed early efforts made by Eugene Wuester and his followers in many countries.
Today, language maintenance is an established and effective activity in most European countries for technical language in the widest possible sense, including the special-purpose languages - "Fachsprachen" - for medicine, law, banking etc. The idea is well anchored in wide circles that, although every language is spontaneously developing in the sense that is incessantly changing, there is no necessarily fortunate language evolution. Much effort is brought into what is called "Sprachpflege" in German. This term is, for want of a better term, traditionally rendered as "language planning" in English, thereby evoking exactly all the most fierce opposition towards planned societies that exists in the Anglo-Saxon my-home-is-my-castle tradition. Sprachpflege and its counterparts in a number of European languages does not, however, mean planning and operates with more subtle means than commands from above. The second component means "care" or "(technical) maintenance" and is the same as is used for, say, hospital care of humans or overhaul of cars or airplanes to keep them fit and up-to-date.
The field is particularly developed in Scandinavia, where highly respected advisory boards give recommendations, on terminology and on usage otherwise, both in special-purpose and general-purpose languages. These boards are typically non-governmental and have no legal power; the closest they have come to officialdom is that in some cases certain recommendations have had to be observed for school books to be adopted by the public school system. But the impact is none the less - or therefore - great, and recommendations are often adopted as in-house language policy by newspapers, industries or other large organizations who, internally, do not hesitate to use a centralistic approach.
The normative power vested in these advisory bodies depends on a consensus that consensus is a value in itself and that deliberate action is better than random change; the criticism launched is in the majority of cases that the language maintenance offices have not yet given sufficient directions; as a typical answer from a user organization when asked to give their reaction to a draft terminology proposal on an important field, I personally remember the car manufacturer Volvo, an industry not without self-esteem, criticizing some proposed decision on central computer processing terms but concluding that it was more important for the terminology committee to publish their guide-lines than to optimize the result. The idea that the country's largest industrial corporation would not obey the guide-lines did not occur to Volvo or to anybody else, particularly not since their view had been properly polled before the final decision.
In such a climate of popular support for joint efforts, and presumably only in such a climate, a nation-wide - or, for internationally used languages such as English or Swahili, language-wide - mechanical surveillance would accelerate the speed of implementation of decisions.
As linguistics develops, deliberate actions need no longer remain on low levels of language, such as terminology, but can climb towards the summits of grammar and style. At the same time, the language-maintenance bodies will at small marginal cost automatically be provided with a better and more recent basis for decisions, in that virtually all sentences written, whether transmitted and read or just written and rejected as faulty, will be recorded and available for individual or statistical scrutiny.
Just one question: Do we want this?
We are not thinking of conventional censorship where messages are suppressed if they have illegitimate contents; no one doubts that we can make machines recognize documents en claire having certain contents - not one hundred per cent of course but a significant proportion. We are considering the effects machine-supported language control may have on the documents which are not suppressed.
As we all know, George Orwell did not hesitate about the answer. With Newspeak people in his society had been given a tool to express authorized thoughts and these only. The words and their meanings were so selected that opposition and criticism literally could not be uttered and new perspectives and dimensions on our lives were not expressible and therefore unthinkable. This was the very opposite of censorship, since there was no longer need to prevent anyone from saying anything that could be said to anybody else.
Was George Orwell right at all? If yes, computational techniques can assist, in manners he could not possibly have foreseen in 1948, in creating a stable society. It is now possible to enforce large-scale and in minute detail the nation's adherence to language norms and automatically and exhaustively record and report any abortive attempt to deviate. If that is so, computational linguistics becomes more socially important than usually realized.
Like most other language reformers, Orwell focussed on vocabulary. If there is no word for a controversial entity, the speaker does not have such concept. With the idea the word is borne on man's lips, says the poet, but Orwell and his likes mean that the idea also dies with the word.
Empirically we know that intellectual development and vocabulary go hand in hand. Is it true then that we can inhibit thoughts in a particular direction by cancelling the words required for the purpose?
In this immediate form, this hypothesis reflects a naively static conception of language. Natural language - and that is one of the key features of its naturalness - is extendable. We can almost imperceptibly and without explicit defining operations or declarations, introduce new terms or modify the meaning of existing ones. That is how we once learnt most words and that is how we daily acquire new ones. Explicit definitions are exceptional, even in scientific prose; in everyday life they are rare, being the last resort to save cooperative discourse from degenerating into quarrel. Therefore, if at some point of time a language lacks words for some concept or if an existing word is tabooed, an interactive user group which feels the need of expression will in a short time have built its own vocabulary which caters for its perceived needs. Restrictions from above will therefore have little effect unless supported by a population already homogeneous and unanimously hostile towards change. Prohibiting the use of a subset of an existing vocabulary therefore seems only to delay unwanted thinking - or even accelerate thinking along unconventional paths.
A good old trick to check how realistic is our evaluation of potentialities is to swap risks and hopes. How insignificant is, to our mind, one-in-a-million if it is not the risk of being killed in a car accident but the chance of winning at a lottery? If we are inclined to disregard the potential adverse effects of machine-supported vocabulary pruning, why are we so ready to believe the positive statement that by providing people with a vocabulary for some purpose, we help them develop their thinking in that direction? By offering terms loaded with experience and associations, we supply ready-made language macros; that is the credo of much educational effort. Like all macros, they can in principle do nothing beyond what the underlying primitive language could achieve, but the difference in speed and convenience may be a matter of orders of magnitude.
To counter unwanted development, vocabulary control would have to be complete: no new terms would be permitted unless specifically authorized by the System's Manager. Such restraints are no doubt easy to enforce in a computerized organization or wired society with one competitor-free telecommunication admninistration (as we have in, say, Sweden). Such an environment would, it seems, damp creativity to such an extent that it would be unacceptable on the ground of immediate productivity and is therefore unlikely to be implemented large-scale. Besides, it would still not introduce an unsurmountable obstacle to illegitimate thinking and communication: if there were not a word for "stupid", one would still be able to say about someone that his mental development asymptote is parallel and close to the x-axis. To say the least, we are far from any effective procedure to detect automatically such unintended uses of matter-of-fact words like "parallel" or "x-axis".
But there is more to it than mere words. No doubt there is some relationship between grammar and style on one hand and contents and attitudes on the other, though the relationship is surely not so simple that certain thoughts require certain sentence or text patterns. One can encourage a type of attitudes by requiring all messages to have some particular linguistic shape.
Thus, if one is obliged to write "I request you to ..." when addressing someone lower in rank and "I hope you might consider it attractive to ..." when turning to a superior, these phrases are likely to direct attention towards social hierarchies - approvingly or resentfully as the case might be.
Another silly example. We know that the words for "and" and for "that" appear at or near the very top of any degressive word frequency list for any language and that "and" gains over "that" in straightforward narrative prose whereas "that" wins in argumentative presentations. These quantitative observations, of course, reflect the tendency towards paratax and hypotax, respectively. I cannot imagine any usable medium of expression which does not include some "and" as well as some "that" but suppose we set a price on the choice! By letting the system tax the user every time he uses "and" or full stop - say, by a 5 seconds delay whenever either occurs - could we bend writers towards more structured thinking? I find it obvious that we could teach them a manner of presentation where logical and textual coherence is made more explicit, but can we more than marginally influence their thoughts?
(The example is not too silly to be almost realistic. For many years, journalists in some evening newspapers had to have their product's readability numerically evaluated by means of one or other statistical "lix" measure, essentially some weighted mean of word length, sentence length and similar quantifiable features. If his lix value did not show his style to be popular enough, the journalist was seriously reprimanded. It goes without saying that he evaded the pressure by replacing some conjunctions by full stop, thus improving his score but leaving the reader with as complex a text as before but with fewer cues to its interpretation.)
These are very crude illustrations to make the conclusion plausible that computable restraints on communication can promote a particular manner of thinking among people who want to belong, although it will remain ineffective as an obstacle to whose who want to circumvent it. It will not make heretical ideas unthinkable as Orwell expected but it will make them conspicuous and therefore easier to avoid by those who are anxious to fall in. There is no indication that social control via language assimilation in a group will work in essentially new ways than they have since it all began, but computation seems to offer more rational means of enforcing homogenization on a fast and large scale. - Again, I am thinking not of the primitive beginnings of computation we have today but of what we will have when we have learnt to handle linguistic structures more delicately.
In the discussion above I have focussed on control in the sense of centripetal force towards a standardized minimum, limiting vocabulary and reducing variation in grammar, text structure and contents. Control may work in the opposite manner and offer additional support for one manner of thinking, giving it a competitive advantage. Guidance can have a more constructive form than lists of shalts and shalt nots.
An example, insignicant in itself, to illustrate non-punitive control. In several European countries we have lately seen the spread of unidiomatic manners of paging, paragraphing and dating documents. The reason is obviously that the writers have inherited foreign patterns from the home countries of their word processors which the software importers have failed to "localize" and the users have not been enterprising enough to "customize". One is, of course, free to override the inbuilt macros. No word-processor, after all, is so rigid that one must adopt a given month-day-year format in one's own texts. But it is so much more convenient not to override.
This is true also in things that matter: it is so much more convenient not to override. We have illustrated by salient detail but control effects from biased helpfulness should be expected to be major.
Such as those on global text and argument structure. There already exist writer's tools, from synonym finders to authoring assistants, which help to sort out all links and detect stale repetition. It is reasonable to asssume that these will develop into really powerful tools for structured writing but that equally effective instruments will not come forth for other genres. These tools will place no extra burden on an impressionistic poet but they will make it relatively easier to construct logically coherent intelligible prose. It would therefore be surprising and indeed very sad, if the writer's tools will not within a generation or so have taught authors, ceteris paribus, a more transparent and structured manner of presentation and, for that very reason, better organized thinking. I expect the average technical report or conference contribution to become clearer in matter when written by those who will have grown up with the new text-manipulation tools. And I would be surprised if the love-letters of that generation would not also have "profited" from technology-guided thinking and contain fewer untenable statements and fewer sloppy glides between post and propter.
Could we even combine constructive and destructive control by designing machinery which would filter out all texts which do conform to conventions on all levels? In a time of explosive increase of computer-assisted mass production of well-meant and well-formed next-to-nonsense texts for computer-mediated international dissemination, a mechanical Dullness Detector is perhaps what writers and readers need best.