<OT> New Posting: ROA-795

[roa at ruccs.rutgers.edu]

ROA 795-0106

Learning From Paradigmatic Information

Bruce Tesar <tesar at rutgers.edu>

Direct link: http://roa.rutgers.edu/view.php3?roa=795


Abstract:
Paradigmatic information is information requiring knowledge
of morphological identity across words. It consists of the
phonological consequences of knowing that a morpheme must
have a single phonological underlying form, even if it surfaces
differently in different words. There are two basic forms
of paradigmatic information. One is morphemic alternation:
the surface realizations of a single morpheme in different
morphological contexts (a context consists of the other
morphemes used to form the word). The other is morphemic
contrast: the surface realizations of two different morphemes
in the same morphological context.

Paradigmatic information is necessary for phonological learning.
This can be demonstrated with a constructed linguistic system
in which several distinct languages, with distinct mappings,
have identical inventories of surface phonological forms.
To learn the full phonology, the learner must utilize paradigmati
c information: that is the only information that can distinguish
the different phonotactically identical phonologies.

Phonotactic learning, treating each word in isolation, can
uncover some ranking information. Once morphological constituency
information is grasped by the learner, some underlying feature
values can be determined on the basis of the ranking information
provided by phonotactic learning, by using inconsistency
detection to test different values for the underlying forms
of morphemes. However, this is insufficient for learning
the full phonology when the relevant ranking information
isn't phonotactically apparent. Paradigmatic information
must be used.

Learning such phonologies is non-trivial because of the
interdependence of the morphemic underlying forms with each
other, and with the mapping. If a learner entertains several
possible underlying forms for a root, the consequences (in
terms of surface realizations) of each considered underlying
form depends upon the underlying forms chosen for the affixes
that the root combines with. The consequences of underlying
form choice for each of those affixes depends upon the underlying
form for other roots they can combine with, and so forth.
All surface consequences of considered underlying forms
are affected by knowledge of the ranking (or lack thereof).
In short, everything seems to depend upon everything.

The learner needs to balance two considerations when learning
from paradigmatic information: information content and computatio
nal efficiency. The learner needs to simultaneously process
units of data that are large enough to contain the needed
information: single words aren't sufficient. On the other
hand, the units should be small enough to be reasonably
processed. The entire data set certainly contains the desired
paradigmatic information, but considering all possible lexica
across the entire paradigm simultaneously will be intractable.

Contrast pairs are pairs of words that differ in only a
single morpheme, where the two differing morphemes surface
non-identically. Contrast pairs provide paradigmatic information,
and can determine the underlying values of certain alternating
features. This is because morphemic contrast implicates
a causal role for underlying feature values: if two morphemes
surface differently in the same environment, they must have
different underlying feature values that actually cause
the surface differences. Contrast pairs also strike a balance
between information content and computational cost: they
contain paradigmatic information not available from single
words in isolation, but only involve a small portion of
the lexicon.

Once contrast pairs have been used to set the underlying
form for a morpheme, the learner can look at words where
that underlying form is not faithfully realized on the surface,
and learn non-phonotactic ranking information. That ranking
information can then be used to determine the underlying
values of additional features. The processes of setting
the underlying feature values and determining ranking information
feed each other, leading to the learning of the phonology.
Contrast pairs provide key paradigmatic information, making
it possible to set the underlying values of some key alternating
features, and thus providing a crucial entry into the portions
of the grammar that are not phonotactically visible.

Comments: To appear in NELS 36.
Keywords: learnability, contrast
Areas: Learnability,Phonology
Type: Conference Proceedings Chapter

Direct link: http://roa.rutgers.edu/view.php3?roa=795