VORTRAG
*******
Oesterreichisches Forschungsinstitut fuer Artificial Intelligence(OeFAI)
Schottengasse 3, A-1010 Wien
Tel.: +43-1-53361120, Fax: +43-1-5336112-77, Email: sec@oefai.at
-------------------------------------------------------------------------
Adam Albright, M.A.
UCLA Linguistics Department, Los Angeles
A MINIMAL GENERALIZATION APPROACH TO RULE INDUCTION
Numerous computational models of morphology have taken on the task
of identifying morphemes and decomposing complex words into their
constituent parts. Relatively fewer models have taken on the reverse
task, of learning rules to compose novel complex forms. Before a
model can combine morphemes to create new forms, it must learn two
things: (1) the contexts in which the morphemes occur (their
distribution), and (2) the rates at which they occur (their
productivity). I present an inductive approach to learning the
distribution and productivity of rules. The model starts by
considering pairs of morphologically related forms (e.g., (present,past)), and
comparing them to discover the rules that are needed to derive one
form from the other. Comparing (jump,jump[t]) and (sip,sip[t]), the
model posits a rule suffixing [t] after stems ending in [p];
comparing further with (kick,kick[t]), it posits a rule suffixing [t]
after stems ending in non-coronal voiceless stops, and so on. This
conservative strategy, which we refer to as "minimal
generalization", can accurately discover the distribution of
morphemes, because it never generalizes a process beyond the contexts
in which it has been observed. In order to discover the productivity
of rules, the model collects simple statistics about the reliability
of rules in different environments.
After describing the basic model, I discuss a common but neglected
pattern of linguistic exceptions, in which a few exceptional forms
take the "wrong" allomorph. For example, the English verb
(burn,burnt) uses the [t] suffix, but in the context of the voiced
sound [n], which should ordinarily take the [d] allomorph. I present
an algorithm for identifying this type of exception and learning the
"true" distribution of allomorphs, even in the presence of such
exceptions.
Zeit: Donnerstag, 18. April, 2002, 18:30 Uhr pktl.
Ort: Oesterreichisches Forschungsinstitut
fuer Artificial Intelligence
Schottengasse 3, 1010 Wien.
OESTERREICHISCHES FORSCHUNGSINSTITUT
FUER ARTIFICIAL INTELLIGENCE
o.Univ.-Prof.Ing.Dr. Robert Trappl