A Man of Measure
Festschrift in Honour of Fred Karlsson, pp. 422–431
Kimmo Koskenniemi
Notes on the Two-Level Morphology
Abstract
Some 24 years have passed since the advent of the two-level morphology. One
motivation for the invention of the two-level model was computational efficiency. The
power and capacity of present computers is now more than thousand times greater than
what was available then. This article briefly relates the two-level model to the current
state of the technology and identifies some possible areas where further work in the
two-level framework would be worth while.
1. Background
Two-level morphology emerged in 1981–82 (Koskenniemi 1983) as a
branch of the finite-state transducer cascade technology which Ronald
Kaplan and Martin Kay had been developing since the 1970’ies (Kaplan
and Kay 1994). Around 1980, the available memory of computers was
about 1/10,000th of the memory of ordinary present computers, and the
speed of computation was about 1/1000th of the speed of present
computers.
In those days, the transducer technology was feasible for simple rule
sets, but it was prohibitively heavy, if several transducers had to be
composed together. The morphological (or graphemic) alternations of
English could be handled then, but for Finnish, so many rules were needed
that the results of the composition would have been too large. The obvious
incentive for inventing the two-level model was to find a more tractable
way to perform the required computations.
The two-level morphology turned out to be more than just as a
computational trick or an optimized implementation of traditional
phonological rewrite rules. It implied a different framework and a new
interpretation for morphophonological alternations. Having fully parallel

NOTES ON THE TWO-LEVEL MORPHOLOGY


423
rules instead of the usual sequential cascade of rules had consequences
such as:
1. There were no intermediate representations or results, i.e. only the
morphophonemic and the surface forms existed.
2. The rule ordering did not matter. The rules could neither feed nor
bleed other rules.
3. The rules were logical constraints defining the relation between the
two representations rather than procedural actions converting strings
to other strings.
4. Two-level rules could refer to conditions also on the surface, i.e. the
final stage of the derivation (such as “ending up between vowels”),
or even refer to relations between the morphophonemic and the
surface (such as a “after a contracted diphthong”). The location of
deletions and epenthesis can be controlled quite explicitly e.g. by
referring to the resulting morphotactic structure.

These properties were partly good news implying certain rigidity and
simplicity. Logical constraints are easy to understand and they have no
opaque interactions. All rules must stick to the truth all the way. Therefore,
one could not fix a fault in a rule by issuing another rule to patch it. The
two-level rules did not tolerate exceptions. Some complicated rule conflicts
and other interactions were also more problematic to handle in the two-
level framework than in the cascading scheme.
The reader is advised to read the Karttunen (1991) or Karttunen and
Beesley (2001) which elaborates the similarities and differences of the
finite-state implementations of the cascaded transducer rules and two-level
rules, and to the Short history of the two-level morphology by Karttunen
(2001) for a more thorough survey.
The original idea was to restrict the use of two-level rules to
alternations which were phonologically conditioned or which represented
phonologically motivated changes. Everything else, i.e. irregular and
purely morphological alternations, was supposed to be taken care of by the
lexicon component. Complicated alternations can described using the
mechanisms of a partitioned and linked lexicon; cf. the Xerox LEXC tool