Learnability of Adjuncts

Publications

Description

An exploration of the formal learnability properties of repetition and optionality, with new ongoing work on how and whether neural models generalise from limited repetition and optionality.

Abstract

Human languages include adjuncts, which are grammatically optional
elements. Some adjuncts can also be repeated indefinitely. I
consider four learnable classes of languages and ask whether
these classes include optional and repeated elements, and what
input a learner requires in order to generalise from finite to
indefinite repetition.



Keywords: adjunct, adjective, optionality, repetition, learnability, PDFA, 0-reversible, n-gram, substitutable context free

Conclusions


  • N-gram learners:

    • For an n-gram learner to learn a language with a string
      x that
      can be repeated indefinitely in some context, the learner must
      encounter, in that context, x, xx, xxx,... up to n
      divided by the length of x, +2, x's.



  • 0-reversible learners:

    • For a 0-reversible learner to learn a language with a string
      x that is optional and
      can be repeated indefinitely in some context, the learner must
      encounter, in that context, n x's and n+1 x's for
      some n. For example, if the learner encounters The red red
      red rose
      and The red red red red rose, it will conclude
      not only that, say, The red red red red red red rose is
      grammatical, but also that The rose is.


  • Substitutable context free learners:

    • Just like for 0-reversible learners, for a substitutable learner to learn a language with a string
      x that is optional and
      can be repeated indefinitely in some context, the learner must
      encounter, in that context, n x's and n+1 x's for
      some n.

  • Clark & Thollard's PDFA learner

    • Repetition and optionality are both learnable
    • No mutual entailment in normal cases


    In human language, not all repeatable elements are optional,
    and not all optional elements are repeatable. However, this work
    does not yet include learners for human-like languages; stay
    tuned!

Sample
screenshot
b is optional; merging states with common suffixes makes it repeatable.

Other presentations, code, and publications

BibTeX
                

@incollection{fowlie2014learning,
title={Learning Adjuncts},
author={Meaghan Fowlie},
booktitle={Connectedness: papers by and for {S}arah {V}an{W}agenen},
series={UCLA Working Papers in Linguistics},
editor={Carson Schutze and Linnaea Stockall},
volume={18},
pages={},
year={2014}
}

                

@phdthesis{fowlie2017slaying,
title={Slaying the Great Green Dragon: Learning and modelling iterable ordered optional adjuncts},
author={Fowlie, Meaghan},
year={2017},
school={UCLA}
}