Counts, not Weights: The first stage of LOCUS: filing novels with no neural network

Abstract

A reader that returns to a novel after a week away can usually recover, within a page, where it was andwho each character is. That recovery precedes any understanding of the book, and this note askswhether it can be achieved without a neural network. LOCUS is a symbolic system that learns only bycounting, from raw text, with no word lists and no labels. Given a novel, it writes its own notebook of thebook at 13% of the book’s size. On 70 novels it had not seen, the notebook names a masked character asaccurately as the full text does (0.332 against 0.322) and more accurately than raw sentences occupyingthe same space (0.265), and writing it takes a quarter of a second on one processor core. The designfollows the Repertoire Hypothesis: methods are kept in one shared memory, what each book teaches iskept separately, and what is distinctive about each book is saved so that the book can be recognisedagain. Each part of the hypothesis was tested in isolation, and 62 of 87 pre-registered predictions passed.Resolving pronouns to characters does not yet work, and nothing in this note tests understanding.