Computer Slaying Drama Bid Probe

Now grim compassion mocks upon the deed.
Today reward the raven of their steel,
To follow thus a noontide revenue:
E'en stout imposture parts without a tryst,
Imperils thus our moonshine of delight ...

Many readers would vaguely suspect this to be one of Shakespeare's not-so-golden passages. After a few more lines from the tens of thousands to hand, the growing lack of sense arouses darker suspicions in the most unwary. Yes, yes, a computer is responsible.

The joke of throwing together randomly generated prose is a very old one. (Jonathan Swift described a mechanism with just this purpose, in Part III of Gulliver's Travels.) When developing a program to do it in a speedy and flexible way [1], I was conscious of working in an all too venerable tradition. But, with the mere mechanical difficulties offloaded to the machine, new doors began to open; without gaining any staggering new significance, the text generator ceased to be only a joke.

My aim had been to make this software fully programmable by anyone who used it and could handle a word processor. The resulting program doesn't simply generate cod Shakespeare, but reads "lexicons" which you write yourself, and which contain the vocabularies and syntax information from which the "random" output is constructed. In this way, a single program can produce bad blank verse, banal aphorisms, unlikely recipes ("Sir Clive Sinclair's Peppermint Brawn"), terrible fantasy titles ("Doomsword of the Crystalquest"), incomprehensible structuralist or high-tech jargon ... these being just a few of the lexicons supplied with the package.

The first unexpected spin-off was that this piece of frivolity simply reeked with educational worth. In preparing the lexicons – a deeply addictive pastime – one has to start recalling all that boring stuff from school about the parts of speech, and plurals, and the difference between transitive and intransitive verbs. When someone started talking eagerly about the program's use in creative writing classes, I felt like a pure mathematician whose determinedly abstract and theoretical work has just spawned a shiny new doomsday weapon.

A second insight was sparked by an earlier Realtime article on computerized style checking, promoting a reductionist approach based on word counts and sentence lengths. (One can't. Resist. The thought. That writing. Like this. Would be. Classed. As very. Easy. To read.) The text generator program does in a way allow the question to be tackled from the other end. Since its operation is admitted from the start to be perfectly random and mechanical – within the constraints of the lexicon – we argue that if the output looks convincingly like the original on which it's been modelled, this doesn't say much for the original....

Thus the emptiness of the pseudo-Bard pentameters is quickly apparent, while the awful fantasy trilogy titles are barely distinguishable from "real" offerings. The recipes are slightly unconvincing thanks to the random choice of ingredients, but routine business letters were simulated with uncanny precision until livened up with rudeness ("Dear End User, Your complaint is ill-judged and impertinent ..."). And nothing is easier than persuading the program to spew out SHOCK HORROR tabloid headlines.

The thought experiment: if a mindless piece of software can churn out prose which at a cursory reading is just like yours, can it be that your own style is, ahem, capable of improvement?

[1]: "A.I.Q.", for IBMs, compatibles, and the Amstrad PCW series; enquiries to Ansible Information Ltd, 94 London Road, Reading, RG1 5AU.