|
Hey, I don't disagree completely, but it's hard to decide how to answer this
kind of question. In short, we are trying to figure out a way to identify a
class of stories, based on a single example and on the observation that it has
some common elements with other stories we happen to know. The difficulty is
in the fact that there are many such "common elements" we might decide to
focus on, or ignore, and each set thereof can substantially change the stories
we identify as "similar to" our single example, the Sorcerer's Apprentice. So, for instance, we could broaden the description to include any story of
losing control over one's, or someone else's creation- therefore "covering"
stories as diverse as the legend of Icarus, Mary Shelley's Frankenstein and of
course the stories of Sañjīva, Phaethon and Abhimanyu. We might focus on the
moral dimension of the story, which draws more stark parallels to the story of
Sañjīva but also King Midas. We might choose to stay as close as possible to
the them of "magical automation" in which case, we 'll only include stories
like the original tale in Lucian's work, the Sorcerer's Apprentice many
versions, and the story of the Golem of Prague, all of which include something
like the Paperclip Optimiser AGI. And so on. I guess now I make it sound like navel-gazing, but this is actually a
legitimate problem with very real applications. For example imagine trying to
organise the stories I list above, plus who knows how many others, in
appropriate categories _automatically_ based on their narrative
characteristics. It's probably impossible to do that sort of thing with
current NLP techniques, or at least to do it in a way that would satisfy a
majority of human classifiers. Not to mention, the problem of choosing what
part of a _story_ (as opposed to what portions of _text_) to attend to when
categorising a document is also not something we can currently solve
convincingly. So it's an ill-defined, hard, classification problem. Perfect subject for a
machine learning paper :) |