This paper looks at three characteristics of traditional formal approaches to language learning: discrete treatment of isolated steps in a linear syllabus; concern with scoring and numerical evaluation; and the use of convergent language and discourse which lacks uncertainty of outcome. It then describes the use of a group of dialogue simulator programs which, while focussing on a limited set of language forms, seek to reverse these three attributes and to create a friendly practice environment. The programs, under the collective name of GRAMMARLAND, generate questions, find answers to questions, obey commands or assimilate new knowledge, all within a highly restricted framework of discourse relating to a graphic display.
We hear a great deal nowadays of the distinction between Learning and Acquisition. Much of this began with Pit Corder’s key paper on student errors (1967), although Corder himself refers to work done by Harold Palmer in 1922, so the concept is not new. Stephen Krashen (1982: passim) has made the distinction the basis of his Monitor Theory, in which he claims that learned language and acquired language are stored and used separately by the brain, and that what is consciously learned cannot originate spontaneous speech. To acquire language one must participate in the exchange of meaningful messages. My GRAMMARLAND project began as an attempt to see if it was possible to exchange meaningful messages in natural English with a machine, and whether this could aid acquisition.
For many learners the classroom is the main, perhaps the only, available environment for learning or acquisition to take place. Traditional classroom exercises concentrate on formal manipulations, which will tend to favour learning at the expense of acquisition. Acquisition, however, may be at work in ways which the teacher and learner do not perceive directly, eg through rubrics or the teacher’s explanations, all of which are or should be meaningful messages. While an exercise may concern replacement of one tense form with another in a fairly automatic process, the learner may be absorbing and trying subconsciously to account for data about the article system. Say the exercise asks learners to change HE WRITES BOOKS. into HE IS WRITING A BOOK. Some of the learners may well be puzzling over the reason for that plural suddenly becoming a singular, possibly perceiving even without the teacher’s help that it relates to the way in which the second sentence is more specific in what it refers to than the first. However, and more importantly, they may not yet be acquiring any feel for the distinction between the simple and continuous verb form, perhaps because that particular insight does not fit into the mental jigsaw puzzle of language which each learner is building. This could be true even for learners who are successfully giving right answers to the exercise items. Corder finds evidence for what he calls a “built-in syllabus” (1967:24). Dakin also shows how insights into language do not occur predictably when the teacher wants them to but rather when the learner is ready (Dakin 1973:14-15 and 145-146). Krashen talks about a Natural Order hypothesis, though perhaps drawing conclusions which are not warranted by the rather sketchy evidence he cites (1982:12-15). All of these writers, however, do give us an explanation for the difference between the smooth linearity of learning predicted by the syllabus maker or textbook writer and the lurching diversity of achievement encountered in real life.
It is characteristic of a linear approach to learning that it makes us think of the pace of learning as the major variable. The matter to be learned is organised into steps, itself a metaphor of distance rather than quantity or space. The teacher has so much time available to present what is to be learned, and the learners are thought to vary only in the pace at which they can absorb it. Mass teaching imposes a single pace on a class. In a drill session some of the students need more time than is available, while others already command the skill and resent spending more time on it. One of the advantages most often claimed for the computer (eg by Last 1982) is that it will solve this problem, allowing each student the optimum time for his or her need. Underlying this kind of claim is an image of the language class as a race track, with the teacher as handicapper, trying to bring the whole class to the finishing line abreast.
To support any claims made for the computer (or any other panacea), there has to be evaluation, and this tends to come in the form of so many items mastered for so many hours of tuition. Mastery, in this context, means correct responses to exercise prompts rather than spontaneous usage of taught forms of language. The conditions under which tests are taken bring the learning mechanism to the forefront, and to infer how much acquisition has taken place from standard test scores is a dangerous extrapolation. In fact, such evaluations of computer-assisted language learning programs as have been carried out are often embarrassingly inconclusive, showing modest gains, or perhaps no overall gains at all, and often showing that it is not the weaker students who benefit but the more gifted ones. Clearly the computer, as a delivery medium for drill and practice, is not solving the problem of mass education nor is it levelling standards. If it achieves so little in the service of learning, what can we expect it to do for acquisition? In particular, how can we expect it to measure gains in acquisition? The answer to that question is, I believe, that we should not be asking it to carry out such measures until we know better just what we are measuring. There is no scoring routine anywhere in the GRAMMARLAND programs.
Natural language use is highly divergent; one cannot normally predict the
form of an utterance from the preceding one except in highly ritualised
uses such as catechisms, jokes and word play. It is characteristic of
learning (as opposed to acquisition) that it uses a good deal of
convergent language in which a prompt determines arbitrarily the form and
content of the response. Acquisition requires divergent language, always
provided that there is enough help in the situation for the meaning to be
clear. Of course acquisition is not achieved by indiscriminate exposure to
the foreign language: there has to be ‘caretaker speech’,
language which is simpler, more careful, more repetitive, and more
concerned with what is visibly present or known about than would be found
in unsimplified natural language. However, if caretaker speech is to
contain meaningful messages, there must be some uncertainty of outcome in
it; otherwise there would be no point in listening to it. A good deal of
classroom discourse contains questions of the type I call quizzes, those
to which only one answer is legal by the rules of the discourse so far. In
real life no question ever forecloses all possible answers but one; if it
did it would be a non-question, a piece of word-play such as “Who was the
George Washington Bridge named after?” (Of course there may be questions
to which there is only one true answer, but that is different from
questions to which there is only one possible answer.) Teachers’
quizzes, sometimes explicitly and often implicitly, do foreclose variant
answers. I remember one episode I saw in a class where the teacher was
drilling “Yes, I am” and chose to do it by asking questions round the
class, “Are you a girl?”, “Are you a student?” and so on. She asked one
small boy, “Are you sick?” He looked puzzled and said, “No, I not sick.”
“Wrong!” the teacher screamed, “You must say ‘Yes I
am.’”
The GRAMMARLAND programs use question and answer pairs in which either the
learner or the machine may be the questioner. The answers to questions are
derived from a truth table, incorporating what the machine currently
‘knows’ about its micro-world. If the machine asks a question,
it may judge an answer true or false in accordance with its knowledge, but
not right or wrong in accordance with its discourse pattern. One
interesting consequence of this is that the answer “I don’t know”,
whether given by the learner answering the machine’s question or by
the machine answering a learner’s question, may often be correct. In
a c onventional classroom one hardly ever hears “I don’t know”
except as a confession of failure.
In its first fifteen years (1965-1980) the development of computer-assisted language learning directed itself towards learning rather than acquisition, and in the process found itself dealing with isolated features of language or simple contrasts rather than with total systems; concerning itself with numerology and generating quantities of numerical data which it could not quite interpret; and suppressing uncertainty of outcome to the point where meaningfulness and interest almost vanished to be replaced by a buttonholing intimacy of presentation called ‘user-friendliness’. A number of commentators assumed that this was an inevitable constraint of the nature of computers. The current view is that this was a failure of imagination. Of course drill and practice has value, if for no other reason than that learners sometimes demand it and feel comfortable with it. But to assume that a computer is capable of nothing else is to misunderstand grossly the nature of the machine.
One can make a computer serve as an acquisition environment in many ways. One of the simplest would be to have it deliver programmed learning on a subject other than language. A programmed tutorial on how to play chess or on the history of the theatre would display many of the characteristics of caretaker speech: a simplified and careful style of exposition, and much reference to information displayed graphically. In the same way many games, simulations, and logic problems in computerised forms will turn the student’s attention away from language manipulation towards some external task, but will still require some use of language. They may, too, generate a great deal of animated spoken language around the screen if tackled as groupwork (Higgins and Johns 1984:37). However, one may take a more direct approach and turn the computer into a conversational partner, which is what GRAMMARLAND attempts.
It is not possible to make conversation, in whatever medium, with a
computer in natural language as one would with a human being, and is
unlikely to be so for many years. The friendly and chatty robots of Star
Wars are creatures of fantasy. This is not an inherent limitation of the
technology, but rather of our own ignorance of what is involved in
understanding language, of how the human mind brilliantly draws on its
diverse store of knowledge and uses it to make sense of what it hears or
reads. However, two key pieces of research in artificial intelligence
demonstrated the possibilities of simulating some aspects of conversation.
The first was Weizenbaum’s 1966 program ELIZA, which showed how easy
it was, using some trickery with the rules of discourse, to persuade
people that they had been understood when in fact they had not. By using a
keyword recognition technique in conjunction with echo questions and some
bland conversation fillers (such as “Please go on”), the program was able
to sustain a plausible imitation of a consultation with a psychiatrist.
A more significant piece of work was that done by Terry Winograd in 1972,
in which he showed that if one limited the subject matter of discourse to
a ‘micro-world’ within which the machine could be given all
the relevant knowledge, and if one further stipulated that all the
human’s inputs would be grammatical and relevant, then one could
have man/machine conversation which was fully meaningful. Winograd’s
original micro-world was a table-top assembly of different coloured blocks
which could be manipulated by a robot arm. With his program, which he
called SHRDLU, one could interrogate the computer about the current
positions of the blocks and how they came to be in those positions, give
orders to manipulate the blocks, and teach the computer new terms for
categorising the blocks and the actions carried out on them. In exploring
a program such as SHRDLU, the major interest for a human being is to find
out what the machine can do, what forms it does and does not understand,
and how much complexity it can handle. One’s instinct is to test it to
destruction, to force it to make a mistake.
The GRAMMARLAND programs are an application of the SHRDLU principle to language practice. The original program, called TENSELAND, is still unfinished, but a description of its structure will show what I am attempting to do. The program contains a database of facts about the inhabitants of TENSELAND, whose names are displayed at the beginning:
| Jim Brown | the taxi-driver |
| Maria Moreno | the tennis champion |
| Susan Dandridge | the novelist |
| Bob Matthews | the maths teacher |
The program ‘knows’ each inhabitant’s name and profession. In addition it stores a simple graphic showing each person doing something other than their work. Thus, if we ask to look at Jim, we see a picture of him watching television, and Maria is shown playing a guitar. The program is equipped with some rules of inference. The evidence of the picture is that Maria is playing the guitar. From this the program infers that she plays the guitar, that she can play the guitar, and that she owns a guitar. The contrary inference, ie from She plays tennis to She is playing tennis, is blocked. The program proceeds with a turn-taking sequence of user’s input, typed at the keyboard, and screen message. It is essentially a neutral and responsive environment; the machine takes no initiatives and never tells the user what to do next. It does not want to converse. It will, however, respond to the prompt ASK ME SOMETHING by generating a random question, and then attempting to understand the answer. The machine will respond to the command SHOW ME BOB or SHOW ME MR MATTHEWS by displaying the relevant graphic, but cannot otherwise manipulate its micro-world. In this respect it is rather more limited than some of the other GRAMMARLANDs I have worked on. However, it does incorporate a learning routine. A simple statement by the user of a form such as BOB IS SINGING is added to a special part of the database reserved for acquired knowledge. When this knowledge is called up, it is always prefaced by YOU THINK ... A sample interaction with the program might run as follows:
What is Susan?
SHE’S A NOVELIST.
Is she writing a novel?
NOT AT THIS MOMENT
Who is watching TV?
JIM IS.
What is Jim’s surname?
BROWN.
Who is singing?
I’M SORRY. I DON’T UNDERSTAND ‘SINGING’.
Show me Bob.
OK.
Bob is singing.
YOU THINK BOB IS SINGING.
Who is teaching?
NOBODY.
Does Bob play tennis?
I DON’T KNOW.
Is Jim singing?
I DON’T KNOW. YOU THINK BOB IS SINGING.
The structure of the parser limits inputs to simple questions and statements. The range of possible inputs may seem large, but it is of course minute compared with what could be interpreted by a human interlocutor. Nevertheless, it is unlikely to be covered by one learner in one session. What one hopes is that the learner, by playing around with questions suggested by the pictures, may gain a feeling for the range of meanings associated with present simple and present continuous forms. In particular one hopes that he or she may pick up the same rule of inference, from the specific HE IS WATCHING TV to the general HE WATCHES TV but not vice versa, that the machine has been given.
The main purpose of including the ASK command, which temporarily reverses the human/computer roles, is to suggest forms of question which the user may not yet have thought of trying. Nor is the user obliged to answer. The machine will answer its own question. A null input, ie pressing the ENTER key without typing any words, is always interpreted as ASK or ANSWER according to context, so that the user, by pressing this key repeatedly, can watch the machine have a rather disjointed conversation with itself. A learner can enjoy unlimited examples, joining in when he or she is ready. This perhaps highlights an important and underrated role that the machine can have, that of demonstrator. Even when asking questions the machine is not carrying out any form of testing. It keeps no record of the interaction, and, of course, does no scoring. How can it when it has no means of knowing whether an input was ‘wrong’ or merely too difficult for its limited computer intelligence to understand? Its role is to respond in a slave-like fashion, not to attempt anything which would conventionally be called teaching.
TENSELAND, since it is still not fully coded, has not been subject to any trials or evaluation. The only program in the GRAMMARLAND group which has been used with learners is a very simple one concerned with space and movement called JOHN AND MARY, in which the micro-world consists of two figures, two rooms and a door. The figures can be moved from room to room, and the door can be opened and closed. The program is described in Higgins and Johns (1984:74-80) and the BASIC listing is given in the same book (177-187).
I have always used the program with small groups of learners and have tried to direct them as little as possible beyond telling them about the demonstration mode. Some groups launch quickly into questioning, while others, after watching the demonstration, will stick to answering the machine’s questions, perhaps staying in this mode for ten minutes or more. If they do not spontaneously try out commands, I may eventually invite them to do so, setting them the task of reversing the positions of JOHN and MARY. Sometimes they solve this easily, but it may be a long and slightly frustrating process, since they have to discover the words BRING and SEND, the only movement commands which the present form of the program recognises. Future elaborations of the program will widen its vocabulary to include MAKE JOHN GO OUT and MAKE JOHN COME IN. A little frustration can be stimulating, but too much will be counter-productive. The most valuable part of the experience has turned out to be the de-briefing, when, after twenty minutes or so, I invite the group to go over what they did, and then to decide what language they would like the program to handle. They may then return to the program for further exploration, although the limited parsing ability of this program, which can handle only about twenty possible inputs, does not permit much exploration. The parser for TENSELAND, by contrast, will be able to deal with some eighty thousand inputs.
Each GRAMMARLAND project is in essence an enquiry into a system or sub-system of English, so it is difficult to automate their production, or to say just what subject matter is feasible or rewarding. Two projects which I would like to work on are a SHOPPINGLAND in which the discourse concerns quantities and prices of goods displayed in a shop, and DIARYLAND in which the subject is an appointments diary, with both past and future reference. It would be very easy to create a GRAMMARLAND which focused on description and comparison; my PHOTOFIT program, though set up as a task, is effectively that already. I am not sure if one could reasonably expect to cover comprehensively the grammatical systems which an intermediate learner needs to command, though I know we shall not find out except by making the attempt.
Obviously a library of GRAMMARLANDs is no substitute for a real human being at one’s elbow, ready to talk, explain, answer questions, or make jokes in the language you are learning. It might, however, be a useful adjunct to an overworked human teacher, trying to teach a language to classes of forty children, unable to respond to each one on demand. The name GRAMMARLAND derives from some observations of Seymour Papert:
... learning to communicate with a computer may change the way other learning takes place. The computer can be a mathematics-speaking and an alphabetics-speaking entity. We are learning how to make computers with which children love to communicate. When this communication occurs, children learn mathematics as a living language. Moreover, mathematical communication and alphabetic communication are thereby both transformed from the alien and therefore difficult things they are for most children into natural and therefore easy ones. The idea of “talking mathematics” to a computer can be generalised to a view of learning mathematics in “Mathland”; that is to say, in a context which is to learning mathematics what living in France is to learning French. (1980:6)
The problem, particularly for an adult, about going to France to learn French is that the natives talk too fast and use too much difficult language. Perhaps GRAMMARLAND can be regarded as a port of call, a protected environment where acquisition is nourished and unlimited stopovers are permitted.
This is a condensed and revised version of a paper written in April 1982 for the English Speaking Union’s annual competition, and fairly widely circulated in mimeo at the time. Since its original appearance I have changed my theoretical position slightly. I am indebted to Peter Roe for this use of the terms convergent and divergent. An example of convergent language used in an acquisition environment would be cyclic repetition in a children’s story. (See Dakin 1973:34 seq).
Corder, S P. 1967. “The significance of learners’ errors.” IRAL, Vol 5, No 4.
Reprinted in Richards, Jack C (ed) 1974. Error Analysis. London: Longman. p. 19 - 27.
Dakin, Julian. 1973. The language laboratory and language learning. London: Longman.
Higgins, John. 1983. “A question of questions”. ELT Journal, Vol XXXVII/2:269-270.
Higgins, John and Johns, Tim. 1984. Computers in language learning. London: William Collins.
Krashen, Stephen. 1982. Principles and practice in second language acquisition. Oxford: Pergamon Press.
Last, Rex. 1982. Address to the CILT conference on computers in language learning, Lancaster, April 1982. Unpublished.
Papert, Seymour. 1980. Mindstorms: children, computers and powerful ideas. Brighton: the Harvester Press.