Computer-Assisted Learning and Language-Impaired Children

ABSTRACT

This paper first reviews research begun in the 1980s into computer-based remediation for language-impaired children who have difficulties with multiple-word language. The paper then goes on to consider how this work might progress in future. Software was developed to investigate the proposal that computer programs which hold written conversations with their users can be effective in language teaching and remediation. The software is described and studies of the software in use are summarised.

Although the software attracted interest at the time and although the studies suggested that the technique could be useful in language remediation, the software never became widely used. Looking back from today's perspective we consider the reasons for this and relate the software to wider issues and trends in computer-based learning (CBL). It is suggested that the approach was never taken up because the software did not fit with how CBL came to be used in schools. This leads to a discussion of how software that simulates written conversation might now progress using today's technology.

CHILDREN WITH IMPAIRED LANGUAGE

Children who are unable to interact with others through spoken language in the normal way have difficulty in developing satisfactory language skills. For most of us, spoken language is an interactive medium with an accessible structure. As children we construct hypotheses about language and test them by observing the effects of our attempts at speech on others. We are able construct our own understanding of language from the feedback and information we receive.

Children whose interaction with others through language is restricted cannot easily construct understanding in this way. For example, children with impaired hearing, especially those who at age 2 have a hearing loss of 60dB ASA or greater, have to rely on vision rather than hearing as their primary means of communication. These children have difficulty with language because they hear only loud sounds, and then only as vibrations rather than as tonal patterns. Children with brain injury affecting the expression or comprehension of language may show a serious delay in language development often referred to as specific developmental language disorder or dysphasia. Children who are deprived of adequate social contact may have language difficulties. Others with impaired language include those with learning difficulties and those said to be autistic.

Within all of these diagnostic categories there exists a group of children who show delayed or disordered development of language (despite in many cases having otherwise average or above-average levels of non-verbal abilities). It has often been proposed that for the purpose of language remediation no distinction need be made between the different diagnostic categories: that the main emphasis should be upon the remediation of language and language-related skills (Dockrell and McShane, 1993, Chapter 3). We adopt this view in the work reported here.

One distinction we do need to make is concerned with levels of language impairment, which vary considerably. Some language-impaired children show delay in developing language but subsequently recover. Others have a complete absence of language. In between there are many children who begin to develop language but show prolonged difficulties into and beyond their school years. One common hurdle occurs during the development of multiple word sentences, after the emergence of the single words. This is the group of language-impaired children we are most concerned with here: children who have difficulties in developing multiple word language.

Examples of multiple word language difficulties are clearly seen in the syntactic difficulties of hearing-impaired children and in the slow development of longer utterances in children with dysphasia or severe learning difficulties. The following examples are illustrative. For a fuller account see Bryen (1982), Conrad (1979), Leonard (1989), Quigley and Paul (1984), Wiig and Semel (1976).

Observed omissions of noun phrases, possessive inflexion, article or preposition in the utterances of children with learning difficulties:-

    "Ran down the street"       "It is Mary dress"
    "I want apple"              "I wanna go this game"

Observed errors in question formation or verb processes in deaf children's writing:-

"Who the boy saw the girl?"     "The kitten is black?" (question)
"Tom has pushing the wagon"     "The boy is a shirt"

Deaf children's misinterpretations or various clauses:-

    Sentence:       The boy was helped by the girl.
    Interpretation: The boy helped the girl.

    Sentence:       The boy learned the ball broke the window.
    Interpretation: The boy learned the ball.

We should not assume from these examples that difficulties are confined simply to the formation and understanding of single sentences taken out of context. Difficulties also occur at both the sub-sentence level, for example in size of vocabulary and in understanding terms for spatial, temporal and kinship relationships, and at the super-sentence level, for example in conversational conventions for initiating and maintaining interaction and in making the inferences necessary for contextual bridging between sentences in narratives.

LANGUAGE REMEDIATION

Approaches to language remediation can be divided into structured and natural approaches. Structured approaches, such as behaviour modification and programmed instruction, are analytical in that they explicitly address issues such as grammar and syntax. An example would be remedial materials organised into structured sets of noun-verb, preposition-noun and determiner-adjective noun phrases. Whilst this has its place and led to some interesting ideas in computer-based language remediation (e.g. Sewell et al, 1979; Bates et. al 1981), it is often criticised for its passive stylised exercises in which language is subject matter rather than tool. Natural approaches, on the other hand, expose learners to situations in which language is used purposefully as a tool. They aim to involve learners in interaction through whatever language they posses, say through conversational activities with teachers or therapists. The remedial materials may still be organised in some way, say according to developmental sequence, but the emphasis is on what language does rather than how it is structured. One programme of remediation based on the Reynell scheme (Cooper, Moodley and Reynell, 1978) involves children in playing with real-life objects, using language such as "Put the spoon in the box" and "Show me the longest pencil". These activities resemble mother-infant games such as those observed by Bruner (1983), which he believes to be very important in early language development. Natural approaches are claimed to reflect the way language is normally acquired.

Whilst both structured and natural approaches have their place in language remediation, it should be apparent that natural approaches can make excessive demands on teachers' and therapists' time. One solution, one that is highly desirable, is to involve others such as siblings and parents in remedial activities. Another possibility is to provide computer-based support.

Around 1982 we began investigating how computers might support natural approaches to language remediation. The main requirement was for software that could interact with its users through meaningful language. The obvious difficulty in this was that, for all intents and purposes, computers could interact only through written and not spoken language, and this remains true today. This seems to imply that computers can be of little help in natural approaches to language remediation which are mainly reliant upon spoken language. However, this turns out to be a much less serious difficulty than it at first appears, and in some ways it is a definite advantage. Firstly, many language-impaired children, especially those with impaired hearing or auditory processing problems, find language more accessible in its written than in its spoken form. Secondly, written dialogue with a computer can have an interactive quality normally absent from other forms of written language, with the immediacy of feedback normally present only in spoken language. Thirdly, the pace of interaction can be slower than in spoken language, allowing time for thought and reflection. Finally human-computer dialogues are not ephemeral like spoken language: a record of what has been said is open to inspection. For some children these factors make written dialogue with computers a highly supportive environment for learning about language.

LANGUAGE-UNDERSTANDING SOFTWARE

The preceding discussion supports the proposal that computer programs that simulate written conversation could be effective in language teaching and remediation. This idea was put forward by a number of computer-based learning researchers in the 1970s and early 1980s. For some it followed from seeing computer-based adventure games and other items of software that supported limited dialogue. For others the idea followed from early AI research into natural language understanding such as Weizenbaum's (1966) ELIZA program and Winograd's (1972) influential SHRDLU program. ELIZA simulated a psycholtherapist (e.g. User: "I feel so depressed"; ELIZA: "How long have you been feeling depressed?") by means of a computational trick. SHRDLU used artificial intelligence techniques to hold a discussion about a "blocks world" (e.g. User: "Find a block which is taller than the one you are holding and put it in the box". SHRDLU "OK".). According to Winograd, SHRDLU could handle complex language such as "Pick up anything green, at least three of the blocks, and either a box or a sphere which is bigger than any brick on the table". But irrespective of linguistic complexity, all these programs used language as the currency of interaction rather than as its subject matter.

In beginning to investigate how computers might support natural approaches to language remediation we devised several SHRDLU-like programs that allowed children to exchange limited, written English dialogue with a computer. It should be made clear at the outset that these programs were by no means as sophisticated as SHRDLU. Our aim was to investigate an approach to language remediation, not to conduct research into artificial intelligence. Our software was written in BASIC and ran on Tandy TRS-80 and later on BBC model B machines with 32K RAM. It was important we used the machines available in British schools in order to be able to collect data in classroom settings. Although this hardware was limited it did not create any major difficulties.

In the software we developed, the subject of dialogue was screen graphics. Each program presented a different graphic environment. One of the simplest programs was concerned with the relative positions of just two objects, a square and a cross. The square could be over, under, to the left of or to the right of the cross. Another simple program displayed three objects, a triangle, a square and a diamond, each of which could independently be coloured red, blue or green. These programs had vocabularies of around 10 words or phrases, allowing around 40 "acceptable" sentences to be constructed by the user. Other programs possessed more complex screen environments and greater vocabularies. One held a dialogue about the numbers, shapes and colours of up to 30 objects displayed on the screen, and had a vocabulary of 32 words or phrases allowing over 3,000 different "acceptable" sentences to be constructed. Programs also varied in cognitive complexity. One discussed the relative lengths of six lines, three of which were labelled "my lines" (indicating that they "belonged" to the computer) and three were labelled "your lines" (indicating that they "belonged" to the user). In dialogue with this program the normal reversal rules for possessive pronouns applied, thus the lines labelled and referred to as "Your lines" by the computer had to be referred to as "My lines" by the user, and those labelled and referred to as "My lines" by the computer had to be referred to as "Your lines" by the user.

The following examples of dialogue should make this clearer.

Program:    SQUARE AND CROSS
User:       Put the cross to the left of the square.
(The computer does this)
Computer:   My turn.
            Where is the square?
User:       To the right of the cross.
Computer:   Well done.

Program:    COLOURS
Computer:   What colour is the triangle?
User:       Is the triangle red?
Computer:   Yes.
            Well done.
            Your turn.
User:       What is blue?
Computer:   The square is blue.
            My turn.
            I will colour the diamond green.
(The computer does this)

Program: SHAPES

User:       Give me twenty one shapes.
(The computer does this, filling the first 21 available positions
with objects of random colour and shape)
Computer:   My turn.
            Do you have four blue shapes?
User:       No, I have five.
Computer:   Well done.
(The computer highlights the 5 blue shapes)

Program:    LINES
Computer:   Is my blue line longer than your green line?
User:       No, my green line is longer.
Computer:   Well done.
            Your turn.
User:       What is shorter than my red line?
Computer    My red line, my blue line and your green line.

It can be seen from the examples above that dialogues could include statements, instructions, questions and answers, either as complete sentences or as elliptical answers to questions. Also, exchanges could be initiated either by the user or the computer.

Input was constructed word-by-word or phrase-by-phrase using units of language marked out on paper overlays to a peripheral, A4-sized, touch-sensitive pad known as a Concept Keyboard. This conveniently precluded attempts to use vocabulary unknown to the programs. However its main purpose was to help focus users' attention upon language rather than typing or spelling. Input was parsed using finite state grammars defined for each program.

The software provided feedback about inputs it was unable to recognise. For syntactically unacceptable inputs the software highlighted the erroneous part of the input and provided short explanatory messages about simple omissions, insertions, substitutions and transpositions, for example "PUT ... GREEN: word missing?", "TO OVER: extra word?". In response to other errors the software provided specific messages. For example, if the user attempted to give an impossible instruction the message "You cannot do that" would appear, or where possible a more exact message such as "You have too few yellow diamonds". If the user answered a question about the red box with a statement about the blue box the computer would respond "I asked about the red box".

EVALUATIONS OF THE SOFTWARE IN USE

Two main studies were carried out using 12 items of software. The first study involved a group of six profoundly deaf teenagers with a mean chronological age of 14 years 2 months and a mean reading age of 7 years 1 month, attending a special unit for the hearing-impaired within a mainstream school. These children used the software for one school term, a period of approximately three months. Before and after this period they took part in a referential communication task administered as a pre- and post-test. The referential communication task was not computer-based, but was text-based and used similar language and concepts to that of the software. Subjects' language in the pre- and post-tests was scored for syntax errors, which gave the following results:

          Total number of errors     No. different types of error
 Subject       Pretest     Posttest          Pretest     Posttest

    1              8              0                   3               0
    2             29             7                   6               3
    3             20           23                   3               7
    4             17             9                   4               2
    5             13             9                   2               2
    6             11             5                   6               3

  Totals          98          53                 27             16

Subjects therefore made fewer syntax errors in the posttest both in number and in type, suggesting improvements in subjects' abilities to construct syntactically acceptable sentences within the domain of language covered by the software. The nature of the errors made was similar to those observed by Conrad, Quigley and their colleagues cited above.

The referential communication task involved subjects in working out the relative positions of objects placed in a grid. In the posttest subjects achieved this task using fewer sentences (in total 12 redundant sentences in the posttest as opposed to 28 in the pretest), which suggests there were improvements in subjects' abilities to use language purposefully.

The above is merely a summary of the findings. The study is reported in detail in Ward et al (1985).

The subjects of the second study were nine 10 to 13 year-olds with a mean language age of 6 years 8 months attending a residential school for speech and language impaired children. The language difficulties of this second group were associated with a variety of conditions including problems with receptive language, problems with expressive language, problems with comprehension suggesting cognitive deficiency and severe language delay caused by environmental factors. This study produced similar findings and is reported in detail in Ward (1988).

Although these studies strongly suggest that written dialogue with computers could be a powerful remedial technique worthy of further investigation, one must be careful not to read anything more into the results. The purpose of the studies was to investigate the potential of a possible new technique in language remediation. They indicate that the software can be an effective way of motivating children to work with language.

Methodological concerns such as problems with single group pretest-posttest designs and the meaning of statistical significance in small groups of subjects are therefore seen as a side issue. The studies were not intended to test beyond question whether subjects made significant gains in their language skills. Such an unequivocal demonstration of language gains would only be useful in the context of a larger corpus of computer-based learning software covering a wider subset of language than the 12 programs used here. Questions of generalisability and persistence of learning would also need to be addressed, especially the extent to which the language learned became incorporated into everyday speech. In fact when these issues are considered, it is very difficult to demonstrate the value of any kind of remediation at all Dockrell and McShane (1993). The studies summarised above suggest that it would be worth developing software to cover a wider subset of language, and only then to evaluate its benefits more thoroughly.

However, before leaving the studies it is of interest to argue through the methodological concerns at least briefly. Single group pretest-posttest designs are common in studies in naturalistic settings and also in recent thinking about the evaluation of computer-based learning software. In practice it is almost impossible to construct satisfactory control groups because large homogeneous groups of language-impaired children within one school do not (thankfully) exist, and even if they did there would be ethical and practical difficulties in denying the control group access to the software. But in any case, the sources cited in the first section of this paper show that without intervention most language-impaired children aged above 10 years will not make any measurable gains over a three or four month period. Thus the single group design is not as problematic as it might at first appear. One other difficulty, that any gains observed in the posttest were a result of experiences in the pretest, also has little weight in this context. It is difficult enough to produce any lasting gains at all, whatever the remediation, let alone gains resulting from a fifteen-minute pretest that persist and are measurable several months later.

WHY THE APPROACH WAS FORGOTTEN

Although the software created considerable interest at the time, only two of the programs were published and the technique was soon forgotten. Looking back from a distance of almost a decade we can now understand some of the reasons why this happened and can consider how the project related to issues and trends in the wider, general progress of computer-based learning.

There are probably two main reasons why the approach was never taken up by others. One is that technology moved rapidly on and the software began to look elementary. This will be addressed under the next heading. The other reason, the one that is the most important and fundamental, is that the software simply did not match the way teachers came to use educational software in the classroom.

If we look back at CBL software developed during the 1980s as microcomputers started to become widely available in schools, we find that the category of software that has had the most lasting success is that which supports rather than directs children's learning experiences. Teachers have come to prefer software that encourages work away from the computer, especially working in groups, in which the computer plays a supporting rather than a central role in the learning environment.

Teachers' preferences are illustrated by the software recommended to teachers and therapists by the British National Council for Educational Technology (McKeown 1991; NCET 1994). NCET draw a distinction between remediation in which the learner is directed by the software, and support which the learner is enabled by the software, and strongly come out in favour of support. They therefore favour software that promotes activities such as word processing, creative writing, a software simulation of a journey which uses paper worksheets, and working in groups with a floor turtle. In all these activities the computer plays a supporting role within the wider learning experience of group project work.

This preference is not confined to CBL in special education, it is evident in mainstream education too. Some teachers very quickly recognised the potency of groupwork supported by simulations, such as Whittington's (1984) groupwork based on a simulation of the raising of Sir Francis Drake's flagship The Mary Rose, or more recently "information stream simulations" which simulate events taking place over time and in which, for example, children play the roles of government and press handling a plane hijack by terrorists. Anyone who doubts the potency of these kinds of computer supported activities should read Warren (1993).

Conversely, much of the software from the 1980s which engaged children in intrinsic learning experiences such as word sequencing and spelling exercises, activities is which language was the subject matter rather than a tool, is now less favoured. The language software described in the sections above falls more into this category than the one now favoured by teachers. The main way in which a child might use the language software would be as a remedial exercise. The way forward for the language software would therefore seems to be to devise and incorporate it into situations in which children can use language purposefully in activities such as games. These tasks might be one-to-one between child and computer or involve children in pairs or small groups in some task supported by the computer.

In fact we did begin to investigate how we could use the software techniques in more purposeful situations. We developed programs that set puzzles for the user to solve. The puzzles were essentially non-linguistic, but required the use of language to solve them. One program ("Hidden Shapes") presented hidden pattern of coloured shapes for the user to reproduce. To do this, the user had to ask questions about the relative positions of the hidden shapes, and then use the information from the answers in constructing instructions for manipulating a second set of shapes. A typical dialogue might be:

Computer:   I have used the blue triangle.
User:       What is above the blue triangle?
Computer:   The green cross and the green square.
User:       Give me the green cross and the green square.
            Put the blue triangle below the green cross.
            What is below the blue triangle?
... (exchanges omitted)
User:       Have I finished?
Computer:   No, two positions are wrong.
User:       Move the red diamond to the left.
            Have I finished?
Computer:   Yes. Well done!
(The computer reveals the hidden pattern).

We also looked at how we might set goals that were extrinsic to the software. One example of this was where two children took turns in giving instructions to a program that discussed the relative sizes of three crosses. One child aimed to make all the crosses as big as possible, the other as small as possible. In these competitive situations, where making a mistake meant losing your turn, children tended to take far more care with their language.

THE POTENTIAL OF TODAY'S TECHNOLOGY

As mentioned above, we began developing the software around 1982. Technology has progressed a lot since then and, after a gap of around 8 years, we have recently started to explore how the software might have developed had the power and capabilities of today's hardware and software tools been available when we started.

Clearly, modern technology has the potential for far more sophisticated programs than those described above. When we recently re-implemented "Hidden Shapes" in Visual Basic as "Hidden Animals", it was easy to include sophisticated graphics (e.g. "What is under the green pig?"), clickable on-screen input buttons (rather than a peripheral device) and a scrollable history of the language exchanged. All these things would have been difficult to implement on the old 32K machines.

We are now beginning to consider further possibilities. Language processing techniques that once required the memory, disk space and processing speed of mainframes can now be implemented on microcomputers. These techniques are capable of handling a large proportion of the structures and concepts that have been shown to cause difficulties for language-impaired children. It should be possible to develop software that deals with subordinate clauses, pronouns, possessive inflections, tense inflections, adverbs, auxiliaries, conjunctions, confusion between "to have" and "to be") and different functions of language such as greetings, the expression of feelings. The software could hold discussions about the everyday objects and concepts used in non computer-based remedial schemes, for example objects such as furniture or fruit, actors and recipients of actions such as people or animals, and actions such as moving, looking, eating and hiding. Software could hold discussions about scenes containing these ideas, with the contents altered by the user's written instructions.

Today's technology also allows better quality of feedback to be provided. For example, because buttons on screen can be dynamically altered (unlike the paper overlays used originally) it is possible to highlight different word or phrase input units, and to vary the complexity of the available language dynamically by enabling or disabling different buttons.

However, the general approach would probably gain wide acceptance only if software was developed to cover an extensive range of language, so that it formed a substantial computer-based remedial scheme. A scheme like this could be of great value to speech therapists and teachers in their remedial work, but we are still a long way from such a scheme.

Acknowledgement

s I would like to thank Dr Andrew Rostron and Dr David Sewell of the Department of Psychology, University of Hull, with whom I collaborated in this research and who supervised my PhD project, several teachers for correspondence, discussions and access to their classrooms; my recent students who have begun to investigate further possibilities; and the many children who allowed me to observe their interactions with the software described in this paper.

References