On the Sad Fate of Gertbot and the Process behind the Stein Poems
Having read over Funkhouser and Klobucar’s ELO paper draft last night (it’s linked in edde’s most recent news/update post), I was reminded of my original ambitions for the Stein Poems project, which began with the attempt to train gertbot, a Jabberwacky chatbot, using Gertrude Stein excerpts, Gnoetry output which used Stein’s writings as input, and my own “lesson plans” written by riffing off of my favorite Stein compositions.
[ASIDE: My favorite Stein compositions are "Identity a Poem," The Geographical History of the United States or The Relation of Human Nature to the Human Mind, "Business in Baltimore," "An Elucidation," "We Came. A History," "Composition as Explanation," Stanzas in Meditation, Tender Buttons, all of her lectures (Narration, Lectures in America), and what I have read of The Making of Americans.]
Early Stein Poems written with Gnoetry (“Why is there a prison” being one of these) were written to be both stand-alone Gnoetry poems as well as to serve as input for gertbot. After half a year of “educating” gertbot, though, it had only learned about 150 phrases, and it became more and more obvious that I would never succeed in a chat with gertbot ever being more than a sequence of random phrases in response to my prompts. The kind of interactive Gertrude Strein experience I had hoped for would not be a product of my efforts. I still believe that such a things is possible, and maybe I (or someone else) might get something like gertbot to work in the future. I had also considered creating a sillibot using selections from Ron Silliman’s The Alphabet as source material. I think I just find the name amusing, though. Sillibot has a nice ring to it.
I’m sure there are a number of reasons for gertbot‘s failure; in the end, I just blame my own poor conception of how the project would work within the parameters of the chatbot software. I simply do not understand well enough how chatbots function, how they “learn,” and what can be expected from a chatbot when such unusual language as Stein’s is used as the model to learn from.
After the death of gertbot, the Stein Poems project has continued as a kind of homage to Stein and to Jackson Mac Low. Mac Low’s Stein Poems (some of his last poems; a selection can be found in Thing of Beauty; several also are linked at his EPC page) are the obvious inspiration of these works, but what Gnoetry and jGnoetry can do with Stein’s compositions is very different from what DIASTEX5 and Mac Low’s edits of its “reading through” procedure can do.
Over time, my process of writing Stein Poems has gelled into a set of procedures and constraints, some loosely and others more strictly defined. As far as source text selection, I always use at least one Stein text and often choose another text/s to add to the mix. This is done according to mood, taste and sometimes simple curiosity. All of the poems are composed of end-stopped lines (periods only), although the length may vary from two to maybe twelve syllables (this is highly variable). Poems are composed by writing them one line at a time and assembling/rearranging these lines on-the-fly. Poems are also to be completed within no more than three days; I have found from working in this way for several years now that the creative impulse/idea/shape which I had in the first couple of days becomes too vague and the resulting poems tend to be failures after two or three days. Preferably, poems are written as a first draft on day one, and a second draft is finished on day two after revision. Minor revisions may happen after that. I also no longer constrain myself to only using the words I am given. I change what words I want, when I want, to achieve the effect that is most striking to me.