The visitor enters a dimly lit room. On a projection screen runs the text that is written by nobody. The keys of the keyboard move as if by a ghost's hand. A monotone, mechanical voice reads out the generated text, sentence by sentence.
Without the public nearby, the system writes quickly and fluently. Thunderstorm of letters. Incessantly, one word follows the other. When visitors approach, the text generator staggers, hesitates, at times grows completely silent. The system leaves the scene to the observer and invites him to strike the keys himself. If he enters text, it appears on the screen like the machine's. Poetry Machine takes up his text and associates starting with his words. The flow of texts in the interplay between the human and the machine doesn't cease.
If the user's input contains words that are still unknown to Poetry Machine, the program sends autonomous „bots" into the internet to get appropriate informations. They evaluate the material found and feed the resulting data back into the system. The search process of the „bots" can be followed on a second screen. Visited sites, their valuation and the documents found are shown.
What Poetry Machine generates is always different and new every time. Poetry Machine also surprises its programmer. It is not an author who varies his sentences and writes them instead of on paper on the hard disk of a computer. When the machine starts, its database is empty. Poetry Machine begins as tabula rasa. The program only contains routines to process text, no hardcoded datasets. Poetry Machine digests text of human authors and extracts their associative interconnections. Its main source of information are the gigantic masses of text in the internet. Here, the machine is listening to what people really say. In the resulting neural networks, words are mainly defined by their relation to their neighbours. A network on the average contains 50.000 connections between 10.000 words. The more often the system observes the connection of two words, the stronger the link between them grows. When the user enters a word and stimulates the semantic network, the energy flows along particularly strongly "tracked" semantic links and finds associated words in this way. The resulting material is inserted in syntactical frames that are also extracted out of the texts digested. The chaotic complexity of the network is reduced to the simple linearity of the text. Like pulling a piece of thread out of a knotted ball of string. The words found serve as the starting point for new associations, until the user interrupts the system with new input.