In considering text, the first distinction that must be drawn is between actions, which are things that happen in the world, and action descriptions, which are text strings -- descriptions of actions in a language. The same action can be described by many different text strings, and many text strings can contribute to the description of one action. The description one chooses to give of an action depends on many things, including how much one knows of the action, the purpose of describing the action to a given audience, and the knowledge and skills of that audience.
In considering action descriptions then, a second distinction that
must be drawn is between narrative and
instructions. Roughly, narrative text describes what has happened
(enriched by scene-setting descriptions, descriptions of characters and
their motivation for acting, etc.), while instructions describe
what should be done to achieve one or more specific aims in a given
situation. How much is described explicitly and how much is left
implicit, depends in part on beliefs about the audience's knowledge
and skills. In the latter case, one relies on the audience's knowledge
and skills to fill
in what is necessary to behave appropriately [SC88,WD90,WBB
92].
For example, consider what the audience for the following (separate)
instructions must know to carry out them out correctly:
1. Carry the beakers carefully.In the first example, a hearer must derive from ``carefully'', behavioral constraints that will keep the beakers from breaking and their contents from spilling. In the second example, the hearer must derive from the phrases ``to the mirror'' and ``fix your hair'', that the appropriate location in front of the mirror is one that will enable him to see his hair clearly. In the third example, if a cold soda isn't immediately to hand, the hearer must derive from ``a cold soda'' locations where one might be located.
2. Go to the mirror and fix your hair.
3. Get me a cold soda.
It is not that equivalent inferences can not be drawn from corresponding declarative sentences, such as one might find in a narrative:
1'. Mary carried the beakers carefully.it is just that to behave appropriately, such inferences must be drawn, if the ``missing'' information is not otherwise provided.
2'. Fred went to the mirror and fixed his hair.
3'. Fred got Mary a cold soda.
Many different kinds of information are provided in instructions, sometimes integrated into a single text, sometimes separated out and placed in identifiable locations or structures within a text. These types of information include:
Open 3L door and install support rod into upper attachment.
Do not operate equipment for more than 30 minutes without cooling air to prevent damage to electronic equipment. If power is to be reapplied without cooling air, allow a 15-minute cool-down period.
Note: The unit weight is approximately 185 lbs. and requires a four man lift.
With door opened, adjust switch until roller contacts cam and continuity is indicated at pins A and B.
Do not use a blowtorch -- it is very easy to start a fire, especially in a loft.
In considering the generation of procedural instructions, there is one final distinction that must be made before noting the elements of a procedural instruction: that is whether instructions are given prior to action or in the context of action. In the latter case, Cohen [Coh84] has observed significant interleaving of fine-grained requests or commands on the one hand, and acknowledgements of understanding or requests for clarification on the other. For example, he notes
In the present study, there were at least two requests used for each assembly step in Telephone mode. Each pair of requests (identification requests followed by assembly requests, or requests to pick up followed by assembly requests) involved at least one common object being manipulated. ([Coh84], p. 110).Cohen observed that identification requests, where the speaker requests the hearer identify the referent of a noun phrase, dominate the first mention of an object in spoken instruction-giving discourse. These rarely had the form of an explicit request such as ``Find the yellow tube.'' but rather took the form of an existential proposition -- e.g. ``There's a little blue cap.''; a perceptual request or statement -- e.g. ``You'll see three very small pieces of plastic.''; or a sentence fragment -- e.g. ``Now, the smallest of the red pieces''. Subsequent reference to an object was, as in written text, through pronouns and definite noun phrases.
Instructions given prior to action rarely have this form. More often, they are given in the form of steps consisting of one or more utterances. For example,
With door opened, adjust switch until roller contacts cam and continuity is indicated at pins A and B. Verify positive switch contact by tightening bottom nut one additional turn.While there are no firm guidelines as to what a single instruction step should encompass, often steps are organized around small coherent sub-tasks (such as adjusting the switch, in the example above). A step may specify several actions that need to be performed together (possibly in some partially-specified order) to accomplish a single subtask, or several aspects of a single complex action (e.g. its purpose, manner, things to watch out for, appropriate termination conditions, etc.). The important point is that an agent must develop some degree of understanding of the whole step before starting to act, so that he knows what all to do and what all to attend to.
The issues then that one needs to consider in generating procedural instructions include the following:
);