Reconstructions is an infinite computer-generated poem whose output conforms to the literary figure of chiasmus. The program samples a sequence of text from a variational autoencoder neural network trained on millions of lines of English poetry, and pairs that line with a reconstruction of the sequence in reverse order. The lines of poetry thereby produced are semantic and syntactic mirror images of one another, exhibiting a classic chiastic X structure: ABBA. At regular intervals, the system produces a new pair of lines, which are then displayed between the lines of the previous step, creating a chiastic structure in the arrangement not just of words within lines but among the lines themselves. As the program progresses, lines continue to appear in the middle of each pair, leading to an endlessly nested chiastic structure (ABCCBA, ABCDDCBA, ABCDEEDCBA, etc.). Older lines fade as they move away from the center but are never fully erased from the screen.
Reconstructions was made by Allison Parrish for ISEA 2020.