The influence of context on learning metrical stress systems using finite-state machines

Abstract

Languages vary in the way stress is assigned to syllables within
  words.  In this article, we investigate the learnability of stress
  systems in a wide range of languages.  The stress systems can be
  described using finite-state automata with symbols indicating
  primary, secondary or no stress.  Finite-state automata have been
  the focus of research in the area of grammatical inference for some
  time now.  It has been shown that finite-state machines are
  learnable from examples using state-merging.  One such approach,
  which aims to learn $k$-testable languages, has been applied to
  stress systems with some success.  The family of $k$-testable
  languages has been shown to be efficiently learnable (in polynomial
  time).  Here, we extend this approach to $k,l$-local languages by
  taking not only left context, but also right context into account.
  Furthermore, we consider empirical results.  Some stress systems are
  only learnable when more examples are provided than the theoretical
  minimum.  Our results show that when learning stress patterns using
  state merging, left context is more important than right context.
  Some stress systems are not learnable using either $k$-testable or
  $k,l$-local language learning system.  A more complex merging
  strategy and hence language representation is required for these
  stress systems.