Poetry Machines / Machine Poetry
On the Early History of Computerised Text Generation and Generative Systems
David Link

Since the construction of the first computer in 1948, text is not only written and read, but also executed. Authors are now able to compose documents that produce content when run. "Poetry Machines / Machine Poetry" investigates the early history of these algorithmic artefacts in detail, traces them back to their literary predecessors, and emphasises the paradigms, contexts and phantasms that motivated and inspired them.

Computers are fundamentally alien to language. While Artificial Intelligence research in the 1960s and 1970s tried to overcome this difficulty unsuccessfully, text adventures used the same resistance playfully to enhance the suspense of the game. The book analyses variable scripts, Joseph Weizenbaum's "Eliza", Kenneth Colby's "Parry", early adventure games and Terry Winograd's "SHRDLU" down to their source code, points out their metaphorical and logical structures, and places them in a genealogy of growing algorithmic complexity. The attempts are based on the belief that language and the knowledge about the world represented by it can be fully explained and even be formalised, emphatically advocated for instance in Ludwig Wittgenstein's "Tractatus Logico-Philosophicus". Technically, optional elements are arranged in tree-like structures and generate seemingly endless variance.

An antagonistic tradition of thought connects the dadaist Tristan Tzara, Claude E. Shannon's re-discovery of the Russian mathematician Andrey A. Markov and the "Cut-Up" experiments of William S. Burroughs. It focuses on operations rather than on options and develops genuinely generative algorithms, which employ different routines to turn found material into collages and to produce effects unforeseen. The lacking machinic understanding of symbols transforms into poetry.

For principal reasons, the study of algorithms cannot proceed purely theoretically. As a concrete example of generative software, whose scope is by no means limited to the medium of text, Link gives an overview of a text generator he developed in the context of this research, "Poetry Machine". The interactive text generator is based on semantic networks and acquires information about language autonomously from the internet. The translation of the fundamental text "An Example of Statistical Investigation of the Text 'Eugene Onegin' Concerning the Connection of Samples in Chains" by Andrey A. Markov, which can be regarded as the foundation of the generative approach, is given in the appendix.

Approx. 150 pages, 23 b/w images.

Table of Contents:

1. Introduction7
Anthropomorphism in computer history. Apories of simulating the mind. Definitions of intelligence.

2. Gramophones on Graves20
Static media – Joyce's gramophone. Disadvantages and possible directions of dynamisation.

3. Variable Scripts22
"Romance Generator". Combinatorial explosion. Contradiction by instantiation. Flatness of variance. Danger of Recursion. Metaphoric subsumption. Combinatorics in Lullus, Trithemius, and Leibniz: Universal language, cryptology, mnemonics. Options versus operations. Fictitiousness of random. Shifters and unique identifiers. Methodology of mass production. Monotony of recombination: Potentiation as division.

4. "I am not sure I understand you fully" – ELIZA31
Simulation of a Rogerian psychotherapist: Joseph Weizenbaum, 1966. Keywords. Decomposition and re-assemblage. Worst case scenarios in text generators. Rhetorics of reflection. Natural numbers as analysers. The topos of the "self-made woman". Transcending natural borders. Machinic mass psychotherapy as panopticon.

5. "People get on my nerves sometimes" – PARRY44
Simulation of a paranoid patient: Kenneth M. Colby, 1971. Passing the Turing test. Pure being and nothing as basis of the computer. Phenomenology of paranoia. Promethean embarrassment. Goedelian incompleteness as reason for bugs. Programmers as professional players. Paranoid programming. Concealing absence, revealing the truth. ARPANET as panopticon. Idealisation of information. Conservation under digital circumstances. Dialogic chiasms. Strategies of paranoid deduction. Food as a metaphor in computer science. Between two empires. Variables of emotion and arbitrariness of variable names. Inversion and projection. Frankenstein. Parry's predecessors. Technique of conspiracy theories. Advantages and disadvantages of machinic psychotherapy. Dramatisation of the hidden.

6. "I smell a Wumpus" – Early Adventures74
Wumpus, Gregory Yob, 1971; Adventure, Crowther/ Woods, 1972; Zork, Lebling/ Blank, 1977. Guessing the hidden. Dramatics of initiation. Turning the computer's language incapability into an advantage. Doubling the world: descriptions as objects. Mnemonics, Hypertext. Absence of resistance and its simulation. Adventures as graphs with passwords. Recombination as method of solution. Key words and words that act.

7. "I can't explain a non-existing event" – SHRLDU86
Robot in a world of blocks: Terry Winograd, 1972. Linotype: "Platonoidisation" of language. Distribution and decentralisation of information. Paul Baran. Unimate. Shakey, the robot. History of parsers. Augmented Transition Networks. Clairvoyance and telekinesis. Absence of outer reference. Coincidence of self and world. Simulation of a task. Moving with the speed of light. Temporality in games. Uncalculable machines. Readability of the book of nature. Recalcitrance of intelligent agents.

8. Generation versus Scripts102
Typology of scripts: variable, interactive, spontaneous. Static versus generative algorithms. Tristan Tzara, William Burroughs, Claude Shannon. Outdoing the typewriter. Uniqueness as measure of information. Language independence of string functions in C. Differentialism in linguistics: Ferdinand de Saussure. Edges instead of nodes, operations instead of options.

9. "Poetry Machine" – a Text Generator Based on Semantic Networks110
Sigmund Freud: Project for a psychology, 1895. Semantic networks. Reading as filtering. Recombining semantic tokens. Syntactic frames. Concept of "networks". Writing as neural firing. Pathways: Association, determination, paraphrase. Routing.

10. Appendix: Andrey A. Markov: An Example of Statistical Investigation of the Text "Eugene Onegin" Concerning the Connection of Samples in Chains119
This fundamental text introduced the concept of Markov chains by counting the letters in Pushkin's "Eugene Onegin" and showed that the probability of a letter being a vowel was dependent on the preceding letter.

11. Bibliography134

Abstract PDF

Poesiemaschinen / Maschinenpoesie. Zur Frühgeschichte computerisierter Texterzeugung und generativer Systeme (On the Early History of Computerized Text Generation and Generative Systems, in German). Munich: Fink, 2006.
Full German publication (4 MB)