/ 1
CHUNKING AND PHRASING AND THE DESIGN
OF HUMAN-COMPUTER DIALOGUES
1
William Buxton
Computer Systems Research Institute
University of Toronto
Toronto, Ontario
Canada, M5S 1A4
(416)-978-6320
buxton.dgp.toronto.edu
INTRODUCTION
It is no secret that the user interface of most computer systems could be improved.
Systems are often intimidating, prone to error, and require a high investment in effort
before productive work can be undertaken. A desire to make systems easier to use is a
good starting point, but we can't get very far without some theory of how to do so.
"Easier to use" is easy to say, but it suggests little about how to reduce errors and
frustration and promote faster learning. In order to make some headway in this
direction, we might best reformulate the problem as "How can we accelerate the process
whereby novices begin to perform like experts?". Underlying this formulation is an
assumption that there is a qualitative difference between how experts and novices
achieve particular goals. This assumption is supported by much of the recent literature
in problem solving and the acquisition of cognitive skills (e.g., Anderson, 1980).
Experts and novices differ in the coarseness of granularity with which they view the
constituent elements of a particular problem or task. Novices are attentive to
low-level details. For example, operational details such as finding a particular
character on the keyboard or remembering the name of a command involve problem
solving. The result is that valuable cognitive resources are diverted from the central
problem at hand.
With experts, these low-level details can be performed automatically. Hence, the size
of the chunks of the problem to which they are attentive are much larger. The skills
that permit these tasks to be performed automatically, however, must be highly learned,
usually through repetition ( Newell & Rosenbloom, 1980). The acquisition of skills,
therefore, can be characterized by developing an ability to perform ever-larger chunks of
a problem automatically.
1
Citation: Buxton, W. (1986). Chunking and phrasing and the design of human-computer
dialogues, Proceedings of the IFIP World Computer Congress, Dublin, Ireland, 475-480.
/ 2
We can now return to our reformulation of the problem at hand, "How can we
accelerate the process whereby novices begin to perform like experts?". Our premise is
that there should be as close a match as possible between the structure of how we think
about problems and the language or representation that we use in solving them. In
what follows we argue that this can be achieved by engineering the pragmatics of the
human-computer dialogue (Buxton, 1983) to reinforce the chunking that we believe
would used by an expert working in the domain. Another way of stating this is that the
dialogue structure, especially the pragmatics, can be engineered so as to maximize
compatibility (Fitts & Seeger, 1953; John, Rosenbloom & Newell, 1985) with the
problem domain.
SYNTAX: TWO APPROACHES
The design of the syntax has a major effect on the quality of the user interface of an
interactive system. It affects learnability, the frequency and nature of user errors, the
retention of skills (as with non-regular users) and the speed of task performance. A
major problem for users is the cognitive load imposed by remembering the tokens of a
command and their order (see, for example, Barnard, Hammond, Mortan, Long &
Clark, 1981).
One approach that designers have taken to avoid such problems is to limit the number of
arguments to a command. The user interface of the Macintosh computer, for example,
limits operators to having only one explicit argument. This causes problems, however,
for operations such as move which require both a direct and indirect object. To get
around this, applications such as MacWrite (Apple, 1984) replace the single command
move with two lower-level commands cut and paste. While the new primitives have
a simpler syntax, the user's mental model must be restructured to map the concept move
onto these two new primitives. Rather than simplifying the user interface, therefore, it
is possible that the single-operand-per-verb strategy simply redistributes the cognitive
loading.
An alternative design strategy exists. If move, for example, is the primitive that most
closely corresponds to the user's model, then the design problem is to use it while
minimizing the burden of remembering the arguments and their ordering. Proof-reader's
symbols offer one approach to doing so. An example is shown in Figure 1.
/ 3
Figure 1: Proof-Reader's Symbol Specifying "Move."
Contrast the directness of this with the "cut-and-paste" strategy
utilized by MacWrite (Apple, 1984).
There are at least three points worth noting about this example, especially in contrast
with the "cut-and-paste" strategy for specifying the same operation:
• the entire transaction, verb, direct object, and indirect object are all specified in a
single gesture;
• there will never be an error in syntax since the ordering is implicit in the gesture;
• the operation is specified using existing skills and does not require restructuring
of existing mental models.
PHRASING AND GESTURE
We can think about the components of the move command in the previous example as
being woven together by a thread of continuity similar to that binds together a musical
phrase. The "statement" is initiated in a state of neutrality, is articulated by a
continuous gesture, and upon closure, returns to neutral state where another phrase
can be introduced by either party. As in music, the phrase is characterized by tension
(in this case muscular) and the neutral state delimiting the start and finish by relaxation.
One of our main arguments is that we can use tension and closure to develop a phrase
structure to our human-computer dialogues which reinforces the chunking that we are
trying to establish.
In the "body-language" of haptic input, kinesthetics and muscular tension are the raw
materials of establishing a phrase structure. With the gesture comes heightened arousal
and performance (Yerkes & Dodson, 1908), and in the periods of relaxation, a clear
indication that it is aright to be interrupted, or move on to the next step.
/ 4
Figure 2: Yerkes-Dodson law relating performance to arousal (From
Kantowitz & Sorkin, 1983, p. 606)
COMPOUND TASKS
Problems that we saw previously in the syntax of a single command also appear at
another level of the human-computer dialogue. In actual applications, many of the
transactions which we perform consist of compound tasks. Selecting an electrical
component and positioning it in a circuit board layout would be an example of a
selection/positioning task (Buxton, 1982). Similarly, identifying a word by finding it
in a document and then selecting it would be an example of a navigation/selection
task (Buxton & Myers, 1986). In many such cases, we would argue that the user
models the compound task as a single entity. In such cases, having to address the
sub-tasks independently may result in an additional burden comparable to using cut and
paste instead of move. Furthermore, we claim that the use of phrasing through
kinesthetic gesture can be used to overcome this problem.
Pop-up menus provide a good example to illustrate our point. In general, one would
consider making a selection from a pop-up menu as being a single task. However, on
closer examination, it is seen to consist of three sub-tasks:
• invoke the menu : by depressing the mouse button;
• navigate to selection : by moving mouse while button is depressed;
• make selection and return : release mouse button.
In this case, the "glue" that ties the three sub-tasks together is the tension of holding the
mouse button down throughout the transaction. By designing the dialogue in this way,
errors of syntax and mode errors are virtually impossible to make since the concluding
action (articulated by the mouse button being released) is the unique and natural
consequence to the initial action (depressing the mouse button). Furthermore, the
tension of the finger holding down the button gives constant feedback that we are in a
temporary state, or mode. (There is a slight irony to this, since it is precisely in
so-called "modeless" interfaces that pop-up menus are most commonly found.)
/ 5
PHRASING AND COGNITIVE SKILLS
In their 1983 study, Card, Moran and Newell discussed how experts collapsed low-level
text editing tasks into cognitive "subroutines" that they termed "routine cognitive skills".
Anderson (1982) describes the acquisition of such skills as being based upon the
compilation and proceduralization of knowledge about the underlying sub-tasks. We
believe that phrasing can be used to organize these sub-tasks to accelerate this process.
PRAGMATICS AND THE COMPONENTS OF INPUT
If Card, Moran and Newell's routine cognitive skills are compilations of lower-level
primitives, one could try to determine the basic building blocks. One possible answer
comes from Foley, Wallace and Chan (1984). They tried to characterize human input
to computer systems from the user's perspective. In so doing, they came up with six
basic primitives:
• Select an item in 1, 2, or 3D;
• Position an item in 1, 2, or 3D;
• Orient (rotate) an item in 1, 2, or 3D;
• Path : specify a path, such as in inking in a paint program;
• Quantify : specify a numerical value;
• Text : enter text, as in word processing
On closer examination, however, we see that these primitives are not necessarily all at
the same level. Let us use the position primitive as an example.
If we use a mouse or a tablet, positioning an object in 2D can be viewed as a single task.
However, the moment that we change transducers and use a QWERTY keyboard,
specifying the same coordinates involves two primitives, namely quantify X and
quantify Y.
Figure 3. Position as an Aggregate of 2 Quantify Tasks
We see from this example that even Foley, Wallace and Chan's six primitives have a
deep structure. Whether the sub-tasks are consciously perceived, however, is very
much influenced by the gesture (and capturing transducer) used. When appropriate, a
single gesture (pointing) can be used to articulate a single concept (position).
We can build further upon the previous example. Let us look at a simple system for
transcribing common music notation (Buxton, Sniderman, Reeves, Patel & Baecker,
