Joyce, Machine Translation, and AI

Maybe five years ago, I read an old Arno Schmidt essay (originally, a radio feature) on the language in Finnegans Wake. Schmidt is a not very well-known German author, dead for a quarter century. (And most renowned for his epic masterpiece "Zettels Traum" - a book that no-one owns, let alone reads, since the only accessible edition is a DIN A3 typoscript reproduction monster that costs ~400 DM. His lesser writings (short stories, historic novellas, essays and radio features) are often quite funny in an ethnographical way, but his idiosyncratic use of orthography and punctuation makes him hard to read.) Schmidt earned his living by translating (English -> German), and in that essay he wrote about the language used by Joyce in FW, which is an extension of the stream-of-consciousness idea already present in the (much more readable) Ulysses. Schmidt noted that languages have a semantic level below the lexeme level that we take for granted. This lower level (now what was his word for it?) consists of the "roots" of words that are powerful enough to evoke associations, but not precise enough to have a unique semantics. FW is supposed to be written at this level, and the "crossword puzzler & punster" nature of the text is a consequence of this. Schmidt demonstrates his idea on a few words from one paragraph from FW (Jugurtha).

When I encountered works on machine translation (Maschinelle Übersetzung), I formed the impression that most approaches worked on the basis of grammatical sentences and semantic dictionaries, with the "holy grail" consisting of a totally descriptive german or English grammar, combined with a semantic net that contains the whole Encyclopedia Brittanica. Translation is usually depicted as consisting of several analysis steps; from the "token" level (phonetic/lexical) up to word & phrase recognition, sentence-level grammar, and context analysis ("context" being everything beyond a complete sentence). This is a nice, methodical, and analytical approach, but it hinges upon the pre-existence of "a semantics" that is to be detected. What if there is none? What if there are many possible semantices? What if a poem is to be analyzed, or a piece of FW?

You might say that machine translation is only advocated for highly formalized, technical or juristical writings, but that argument doesn't cut: Even the most technical texts contain subtilities such as emphasis by inverted word order, choice of words to improve connotations, or metaphorical use of language: How often have you read that a browser "travels" "up" a "tree", starting "at" the "root" nodes? Humans have no problems with these metaphors, use them freely, invent and extend them on the spot, and mix them without anyone even noticing. A machine translation database needs special knowledge about each metaphor (and I tend to think that the translation are often only acceptable because the source and target language share most of their "metaphor space", so that the 1:1 translation doesn't do much damage).

Now my intuition is this: the translation process should be based on the acceptance of non-unique semantics, instead of viewing them as a problem. "The semantics" of a word, a sentence, a paragraph, or even a complete text would then more correspond to a "wave function" (to use a trite new-age physical metaphor, but I don't know a better one) instead of a "data structure" (semantic net, grammar) as it is now perceived. A very technical text would have a "wave function" with a "narrow peak", signifying a low tendency of misunderstanding; an ambiguous sentence like "jack and john went swimming. he drowned." would have two peaks, signifying the two primary interpretations (jack drowned/john drowned), and many smaller peaks for the improbable, but still possible, meanings (somebody else drowned / "drowned" or "swimming" is used in a different meaning / "he drowned" is unmarked direct speech / ...).

I have, of course, no idea of how to represent such a "probabilistic semantics", nor how to build inference or composition rules for them. But the basic idea is already present in the Schmidt essay: if there is a semantics in each phoneme (if only the meaning: "this phoneme is the start of the following lexemes" ;-), composing two phonemes amounts to "adding" (or multiplying/cutting) their "semantic fields"; adding new words and punctuation will cut down on the possible interpretation space, as will the context of the previous paragraphs (which not only gives meaning to relational words (he, it) that refer "back", but also weights the interpretation of already used words).

I gather that such a system would go a far step toward true Artificial Intelligence.

I wrote this musing down after encountering a few postings (to rec.arts.books) of Jorn Barger, who has both an interest in Joyce and has been working in AI for some time. I haven't read his articles in detail, but gather that Jorn seems more to think on the lines of situation descriptions and story lines.


(24.04.95) After two day's search, I yesterday found the reference (I had mis-remembered the anthology in question ;-). The essay is called "Das Buch Jedermann" (The Book Everyone), and the terminus coined for the meanginful sub-word entity is "etym". To Schmidt, etyms form the "language of the (un)consciousness", and a stream-of-consciousness should be written as a sequence of etyms.
This page was last changed on Apr 24 1995, 17:31 by mfx@pobox.com. Comments and corrections welcome.