Jumbo Jungle

Hofstadter details the workings of his Jumbo program in chapter two of Fluid Concepts. The purpose of this program is to model the instant mental processes of assembly and transformation. Jumbo is a builder program which constructs English pseudo words from a given set of letters using pre defined rules. Hofstadter emphasizes that Jumbo builds words and does not extract or compare English words with a pre loaded dictionary file. A simple dictionary comparison would not exhibit any intelligence or creativity, only quality assurance. Hofstadter is instead trying to emulate the spontaneous and unconscious composition of coherent wholes derived from scattered parts, a sort of creative construction and regrouping of letters into words. I personally like the analogy of visual art with the construction of words. A painter starts with a defined palette of colors and although the painter most likely has a mental picture or pre determined aspect he/she wishes to construct, the painter does not know exactly what the picture will look like or the subtle nuances and artistic expression it will have. Also one must consider that the painter knows which colors go best with certain others and which will mix to make a more defined and powerful picture. Much like the painter, the Jumbo program starts with a pre determined ‘picture’ or set of rules which govern its letter combinations and then constructs words from a palette of letters, the English alphabet. The program itself has no prediction of its output but will stay within the parameters it is given. Jumbo will have a sense of which letters fit with one another or are commonly grouped together, but one would expect for some errors in arrangement of multiple letter segments that might occur. However, Hofstadter explains that Jumbo will compute many possible letter groupings in parallel and seek out the groupings that are more probable and likely than the others. Each successful grouping of letters triggers another test for grouping those letters with another set and so on, until the most likely combination is reached. Hofstadter emphasizes that there must exists multiple stages of groupings with different parameters. Bonding, combining two neighboring elements together that can still be viewed as separate is different than Hofstadter’s notion of glomming. Glomming is a set of suitable bonded items that creates a higher level structure through its relationships, representing the parts as a whole and reducing further break up of its separate parts. This is just the basic workings of Hofstadter’s sophisticated Jumbo program.

When reading about the different levels of groupings such as bonds and gloms, it occurred to me that this process very closely emulates what humans actually do when constructing words. For example, when I see the mixed up letters ‘reeht’ I first pick out the bonding group of ‘he’. Then further structure it into the glom of ‘the’ and further still group all the given letters into ‘there’. When I grouped the ‘the’ I knew I had to be on the right path because ‘the’ is a word and well known word fragment and could not break the contain letters up again. I wonder how much prediction has to do with this aspect of word construction and if it is common to get multiple words with increasing input size. Such as the case as giving Jumbo ten letters to work with instead of five and receiving multiple different full word outputs.