|
Paper given by John Higgins at the University of Rostock, 15 November 1989, immediately after “Die Wende”, the opening of the Berlin Wall. |
Abstract: This paper looks at several different senses of the word error in relation to trial-and-error learning, and suggests that there is a danger in labelling as ‘errors’ what are linguistic experiments. It deals with a research project which studied peer-correction in a problem-solving activity, and describes two computer applications, a reading pacer and a voice digitiser card, which have potential applications in the training aspects of language learning.
“The computer puts the trial back into trial-and-error.” For several years I have been using that sentence as a kind of slogan. What I have been trying to say with it is that the computer permits and to a great extent encourages an experimental approach to learning which is hardly possible with a mass class served by a single teacher. This is the result of two factors; firstly the computer is a responsive rather than an initiating device, and secondly the computer normally interacts with only one user or a small group simultaneously. Experimenting, and I mean experimenting by learners rather than teachers, is an unpredictable process; a human teacher with thirty pupils cannot realistically respond to thirty or more unpredictable events. A computer, while it can respond to only one event at a time, is so fast that it can normally cope with the inquiries or experiments of more than one person. Therefore a group can use it to try out inputs and will get feedback as fast as they can think of things to try.
The type of activity I had in mind was very much the problem-solving or puzzle-based activity exemplified in programs such as my own STORYBOARD (1982), now re-published as ECLIPSE (1989), though the spirit can be and has been extended to such traditional exercises as multiple-choice grammar questions. One of the most interesting CALL novelties to have emerged in the last three years was TESTMASTER by Chris Jones(1987), which looks like a standard multiple-choice test generator apart from one tiny difference: when you have found the right answer to a question, you are not forced to move on to the next item. The program does not assume that getting a question right means that you have no further interest and allows you to find out what you would have been told if you had got it wrong. Most of the time, of course, you will have no interest in that, but where the interest exists, you are encouraged to pursue it. In the paradigm of teachers' questions and pupils' answers, there is an implicit assumption that a right answer indicates command of a skill while a wrong answer shows the opposite:
| Answer right | Answer wrong | |
| Learner can | ✓ | |
| Learner can’t | x |
The real situation is much more complicated:
| Answer right | Answer wrong | |
| Learner can | ✓ | Carelessness or experiment |
| Learner can’t | Lucky guess or double error | x |
The case of a learner who gets the right answer as a result of a double or self-cancelling error is particularly interesting. Dakin (1973) discusses several such cases. What the author of TESTMASTER has realised is that the learners themselves are often more aware than teachers that they are in one of the 'wrong boxes' in the diagram, and it is the learners who can best help themselves by finding out what feedback is given for other answers.
What does trial-and-error actually mean? The word error in the phrase is potentially misleading. Let's take a situation where a teacher asks a question. The teacher, of course, knows in advance what the ‘right’ answer is, and labels all different answers as ‘errors’. But there is an enormous difference in the way that wrong answers can be given. A student may give a wrong answer and then realise it is wrong. The answer was just a careless slip and hardly needs any correction. Some writers use the label mistake rather than error for this behaviour. A student may give a wrong answer believing it to be right. In that case one can well label the answer an error, something which sends a warning signal to the teacher that something must be done to correct whichever faulty rule has led to the generation of the response. A student may also give a wrong answer in order to find out if it is right. You can hardly call that an error; it is an experiment, an attempt to collect the data which will lead to the internalisation of an appropriate rule to deal with future tasks of the same kind. “I know I can say a quarter past eight and half past eight; can I also say three-quarters past eight?” No you can’t, I’m afraid, but if you never try in circumstances where feedback is available, then you may never know for sure.
Perhaps the computer’s greatest weakness is that it is very poor at distinguishing mistakes from errors and errors from experiments. The efforts that are being made to create intelligent tutoring systems sometimes include an attempt by the machine to build a representation of the learners and their learning behaviour precisely to improve the computer’s performance and let it give relevant feedback. Although they may build a machine which recognises carelessness, I will be surprised if anybody succeeds in writing a program which can reliably distinguish error from experiment. Human teachers are not always perfect at doing this either, though they are, I believe, a good deal more perceptive than any machine and can usually pick up the signals sent out by learners to show when feedback is overabundant or irrelevant. The real experts in distinguishing the different types of error are the learners themselves, and feedback does not have to come from a teacher; it can come from any conversational partner including fellow learners. A recent study carried out in Bristol (Naanda, 1989) used transcriptions of 90 minutes of group-work with a computer-generated puzzle, and showed a surprising amount of peer-correction of pronunciation and syntax going on, including some which echoed the kinds of rules that one would expect to hear only from a teacher (“Magic E”).
The type of error classified here as experiment is probably vital in the second-language learning process, simply because explicit rules leave so much out. In another paper (Higgins 1988b) I commented on the curious affinity which links the present continuous tense with the indefinite article:
(1) She is writing a book.
contrasted with the use of the simple present with generic plural:
(2) She writes books.
The reverse of these cases are grammatical sentences but require much more contextual justification. Out of context they sound distinctly odd:
(3) She writes a book.
(4) She is writing books.
This kind of knowledge, which most competent speakers of English have, is not learned from explicit statements of rules. The rule which makes (1) more natural than (3) is not in any grammar book that I know. I noticed a similar case from a television interview the other day. The speaker, a West German, said
(5) If this is going on, there will be difficulties.
From the context he presumably meant to say:
(6) If this goes on, there will be difficulties.
To a native speaker, both sentences are quite acceptable but they mean different things. (5) could be paraphrased as:
(7) I don't know if this is happening now, but if it is the case, then there will be difficulties.
while (6) could be paraphrased as:
(8) This is happening now. It may continue, and I believe in that case it will cause difficulties.
This rather subtle difference in the time reference of the verbs is not likely to be learned from a grammar book nor, for that matter from a computer. The most likely way of learning it is from reading, talking and having one's misunderstandings put right.
One of the greatest benefits of the computer is in supplying subject matter to be talked about. Naanda's study, which I have already mentioned, used a logic puzzle generator called TRACK which some of you may have seen, and there are a host of similar activities available. Text reconstruction, too, has as its main benefit the provision of reasons for collaboration and thinking about a topic. Some language input and a task to carry out with it are both supplied by the machine, but the computer itself does no teaching. Other speakers will have more to say about such work. Meanwhile I want to mention a different application of trial-and-error, its use in what we could call training.
There are essentially two strategies in skills training. One is gradual withdrawal of support. In learning a new script you start by tracing letters, continue by copying them to a different place on the page, and finally you write them. A driving instructor might make use of dual controls at the beginning, but then uses them less and less. Behaviourist theories of language learning, with their expressed aim of eliminating error as far as possible, clearly embraced this approach.
The second strategy is imitation and approximation with feedback. Like the artillery gunner who fires a shell and waits for the observer to tell him where it has landed, the trainee serves the tennis ball and waits for the instructor to tell him or her what was wrong with the stroke. This is what trial-and-error conventionally means, and it is something which the computer can do very well provided it is only reporting measurable external facts.
One application of the principle which has recently concerned us at Bristol is reading speed. Clearly faster reading is of practical benefit, and it is probably though not provenly the case that some learners with enough general control of the language to read well are held back by attitudes and behaviours which reflect their anxiety about comprehension. They have been implicitly trained to believe that you must understand word 1 before you can understand word 2, that only total comprehension will do. In fact, such an approach leads to rather poor comprehension; the learners lose sight of the forest while identifying the trees. A reading speed program in that case needs only to persuade them that they can read faster and still emerge with undiminished comprehension of the totality of the text.
We have written a program called HOPALONG, still not quite polished enough to publish but good enough for research use, which tries to make learners much more conscious of what they do when they read. It does this by running a highlight through a text, the speed of which can be controlled by the keyboard or a joystick. It is a diagnostic tool rather than a training program; the chief instruction given to the user is to find the speed of the highlight which feels most comfortable. The program allows you to pause the highlight whenever you feel you are losing the thread, or to re-read pages. At the end of the session the program delivers two printouts: a graph showing the reading speed through the text, and a transcript of the text with inserted marks at every decision point. Here is a case where trial supplies a particular kind of feedback which would hardly be available from any non-mechanical source. What we are not yet sure of is the long term benefits that may or may not arise from repeated use of the program. Some studies have been carried out (Wallace, 1989), but we have not had the resources to establish the effects of the program over a period of use with the same population.
One other device that we have begun to experiment with is a speech digitiser. This is an expansion card which fits inside a computer, to which a headset with built-in microphone is attached. The card is not all that expensive; the one I use cost £200. What it does is to allow speech to be edited very much as one edits text in a word-processor, inserting blocks of sound or silence, cutting out particular fragments, or even reversing a stretch of speech, which is a wonderful way of showing learners what English diphthongs actually consist of. The most exciting application seems to be in playing around with unstressed syllables and weak forms, to see what effect there is when those sounds are masked. A recent experiment (Graham 1989) has used a gating procedure to reveal sentences in fragments of increasing size, starting from a point in the middle where there is a cluster of unstressed syllables. Subjects were asked to write down what they heard in any way they could, and the point was to see at what stage they would write intelligible English rather than nonsense. Some interesting findings have come out of this, and the work gained a prize in this year's English-Speaking Union competition. The potential for the device seems vast; I use it in almost every phonetics class I give, simply to let the learners have some form of visual feedback from whatever noises they make, so that there is an independent measure of sameness or difference. It also underlies an interesting project in listening (Al-Hinai, in progress) using the ‘hidden word’ technique in measuring comprehension, which has already thrown up some interesting data on the role of weak-forms in listening.
In a memorable passage Bright and McGregor (1970, p. 130) once said, “The pupil does not learn from his mistakes; if he did, the more mistakes he made, the more he would learn. Common experience, however, shows that the pupil who makes the most mistakes is the one who has learnt and will learn least.” Advocating the kind of trial-and-error learning that I have described is not opening the door to unlimited error and communicative pidgin. Indeed that result is more likely to arise from over-reliance on techniques such as role play and improvisation. What I am advocating is more trial and more feedback, so that the errors of experiment achieve what the learners want them to achieve, and that the crystallised errors due to faulty internalised rules diminish. Perhaps I should have called this talk More Trial — More Experiment.
Al-Hinai, Ahmed (1989 in progress). The use of hidden word tests in measuring the listening skills of Omani secondary school pupils. M Ed dissertation, School of Education, University of Bristol.
Bright, John and McGregor, Gordon (1970). Teaching English as a second language. London, Longman.
Dakin, Julian (1973). The language laboratory and language learning. London, Longman.
Graham, Anne (1989). “Speech-gating and the intelligibility of utterances”. Unpublished article drawing on DASE research project (winner of the ESU prize in 1989).
Higgins, John and Davies, Graham (1982). STORYBOARD, software for BBC microcomputer. Wida Software, re-issued as COPYWRITE by ESM Living and Learning.
Higgins, John (1988a). Language, learners and computers. London, Longman.
Higgins, John (1988b). “Power to the pupils”. In Das, Bikram K (ed), Materials for Language Learning and Teaching. Singapore, SEAMED Anthology Series 22, pages 89 - 99. And elsewhere on this site.
Higgins, John (1989). ECLIPSE, software for IBM computers. Distributed as shareware. Published in the USA as RHUBARB by Research Design Associates.
Jones, Christopher. TESTMASTER, software for BBC and IBM computers. London, WIDA Software, and Zurich, Eurocentres.
Naanda, Aune (1989). Using logic puzzles in foreign language teaching in Namibian Lower Secondary Schools. Unpublished M Ed dissertation, University of Bristol School of Education.
Wallace, Ruth (1989). The use of a computer reading pacer in measuring reading skills of overseas students. Unpublished M Ed dissertation, University of Bristol.