User:Firespeaker/Apertium-turkic talk outline

From Apertium
Jump to navigation Jump to search

Sketch for talk on Writing Turkic-language morphological transducers using HFST (for MT) on October 2nd.

Abstract

This talk will outline the development of Free/open-source morphological transducers for Turkic languages using HFST (the Helsinki Finite State Toolkit). Morphological, phonological, and orthographical challenges encountered in Turkic languages are reviewed, and working solutions are presented. Also included are reasons for developing morphological transducers, how these can benefit communities that use the languages, and the current development status of various Turkic morphological transducers.

Morphological transducers: what and why

  • slide 1: definition, example (sample input/output)
  • slide 2: use in RBMT, specifically apertium
    • apertium as MT for lesser-used/marginalised languages
  • slide 3: other uses: spell checkers, ...?
  • also mention: why we use orthography, not some transcription
    • spell-checkers
    • accessibility by native speakers (as devs and as end-users)
    • no need for pre/post-processing
    • effect: it makes it harder, but means it's useful to communities that use the language

Turkic languages

Geographical/demographic overview of Turkic languages

  • slides 4, 5?
    • a map, numbers of speakers, wikipedia presence

Morphological and phonological properties encountered in Turkic languages

(these are all to be taken as "challenges for morphological transducers")

  • slide 5: Agglutination
  • slide 6: Vowel harmony
  • slide 7: Consonantal processes
  • slide 8: "buffer" segments
  • slide 9: phonology of numerals and acronyms
  • slide 10: Cyrillic orthographical issues
  • something on morpho-syntactic issues that've come up a lot
    • no suffix can attach to "any word", "any part of speech" or even e.g. "all nouns"; often suffixes recur in very specific sorts of places; it's almost like we have dozens of POSes
      • We don't want to overanalyse(/overgenerate)
        • disambig issues
        • testvoc issues
    • Adjective classes (e.g., whether used as <attr>/<subst>/<advl>, +comparative, etc.)
    • Non-finite verb forms
    •  ?

Developing a morphological transducer

  • Important resources to start with:
    • a corpus
    • some grammars and dictionaries
    • linguistic knowledge of the language (if you want to get into it deeply)
    • native speakers!
      • ability to work with informants
      • patience!
      • cf. Chuvash (i.e., the native speakers hopefully agree on forms)

HFST and how we use it

  • slide: HFST: what and who
  • slide: our purposes: using two two-level systems together for a three-level system (?):
    • slide: overview of lexc and why it was chosen
    • slide: overview of twol and why it was chosen

Examples: how morphophonological issues above are dealt with

  • bing
  • bang
  • bam

State of affairs now with apertium-turkic