Latin and ATNs
Jan 8 2021 05:58

Last year I undertook to learn Latin1. I’d stumbled on Hans H. Ørberg’s Lingua Latīna per se Illūstrāta and it clicked2. A “natural method” text, it is a novel about Roman family life in the early empire, proceeding from the very simplest sentences (“Rōma in Italiā est”) to fully natural sentences in context. The story is compelling. The book is in Latin only, with marginal illustrations to support explaining the text. (As in early children’s books, the simple pictures are subtle and crucial parts of learning to read.) Each chapter is followed by a grammar section (in Latin, but carefully designed both to teach the grammar and to re-inforce the use of Latin) and exercises. (Answer keys are published separately.)

As I learned, I found myself introspecting on the process by which I was learning. I decided to write a computer program that learned with me. The input was to be the chapter text and the assessment questions. The output was to be answers to the questions and the evaluation metric was comparison with my answers.

I knew the program would need to break sentences into words, learn the idea of endings, declensions and conjugations; I was both expected to infer the notion of word changes, and was shown the morphology explicity in the marginal notes and the grammar section, so I would have to encode was I was taught explicitly in the program, and figure out what and how I was inferring. It would have to have some kind of memory of the facts in the story, and models of emotional and social life to fill in inferences about why the people in the story felt and did the things they felt and did. There would need to be a question answering modules. Yes, yes, this is the recapitulation of fifty years of work in natural language processing, cognitive modelling, artificial intelligence, commonsense reasoning - no, no, I didn’t imagine I was going to do it all, and all myself. But day-dreams are nice.

The first chapter being a remarkably “begat” like beginning of simple statements naming towns, countries and rivers, I decided to start with an ATN, an antique form of state transition diagram. I thought I might try to infer it from the text, but quickly gave up that notion and wrote the ATN by hand, augmenting the augmented transition network as needed for each each new sentence and sentence pattern.

  1. I’d made three previous attempts: decades ago via Frederic M. Wheelock’s venerable Wheelock’s Latin 3rd edition, my pons asinorum being chapter IX on Hic, Ille, Iste; the Quasillum latin study lists run by dedicated volunteer moderators who collates the weekly assignments of participants working through some particular text together, a low-pressure mildly social means of collective self-study; and an early partial version of Duolingo’s Latin tree. ↩︎

  2. If you decide to use the text, I recommend you consider the Dowling Method and Luke Ranieri’s YouTube lessons and other material. ↩︎

projects, latin, nlp, linguistics, augmented transition-networks, atn, computational linguistics, series latin.