Markus Walther's PhD thesis

This page describes my PhD thesis in computational phonology. The thesis itself is written in German (vi+370pp) and will be published by Niemeyer, Tübingen. In the following abstract the title has been translated to English; see the BibTeX entry for the original wording.

Declarative prosodic morphology - constraint-based analyses and computational models of Finnish and Tigrinya


Markus Walther, University of Duesseldorf (Germany)
Current theoretical work in linguistics, and particularly in phonology, witnesses a trend towards using declarative wellformedness conditions (constraints) in order to adequately describe linguistic generalizations. The Declarative Phonology paradigm (Scobbie 1991, Bird 1995) adopted in this thesis maximally exploits these in describing both the grammar and the lexicon with inviolable constraints. The specific contribution of computational linguistics in this regard is to show how constraints can be implemented in order to derive precise consequences from formal grammar descriptions in an automatic fashion. Prosodic morphology lends support to the constraint-based approach from its empirical domain: words preferably consist of wellformed syllables resp. syllable constituents, etc. Wellknown examples may be found in the rich morphology of Semitic languages such as Arabic or Modern Hebrew with supposedly `non-concatenative' word structure.

On the `atomic', segment-sized level of prosody the present work develops a new non-multiply-linked representation of geminates. Being inherently ambisyllabic and monopositional it contains information about the border of two neighbouring syllables while nowhere referring to two segmental positions. This is made possible by classifying syllable-internal structure with two prosodic features [ons] and [cod] to yield, besides the traditional roles `onset', `nucleus' and `coda', a new fourth syllable role `codaonset': the proposed representation of geminates and ambisyllabic segments alike.

It is shown in detail that this representation is linguistically adequate, in particular being able to account for the characteristic stability effects found for geminates while disposing with all additional Linking Constraint-type principles employed in previous accounts. A computational in-depth study of Finnish consonant gradation proves that this result persists in the context of a realistic phenomenon. The new perspective here is that a voicing opposition is seen as central to so-called qualitative gradation, while quantitative gradation (degemination) is entirely parallel in employing an opposition that involves the prosodic feature [cod] introduced above.

Next, the thesis turns to word-level, `molecular' prosodic effects. Here the focus is on `positionally variable' prosodic morphology, where the realization X or non-realization 0 of optional segment positions (X) - preferably vowels (V) - is controlled by syllable structure constraints. Expanding on ideas by Hudson (1986), generalized X/0 alternation is explicitly recognized as a formal primitive. Now, if Modern Hebrew g(a)m(a)r `he finished' must, according to the structural constraints, fulfill a CV(C) criterion, then *gamr, *gmar, *gmr are all ungramatical, but ga.mar is not. This leaves open what should happen to ji-g(a)m(e)r-u `they will finish': both jig.me.ru and *ji.gam.ru yield well-formed syllables. The decisive innovation of this work is to postulate a psycholinguistically plausible incremental optimization principle `Prefer non-realization as early in time as possible', i.e. from left to right, which in our example enforces the unique choice of jig.me.ru from the set of prosodically well-formed alternatives. Note that - unlike in Optimality Theory (Prince & Smolensky 1993) - optimization does not influence autonomous constraint interaction: it merely filters the small and finite set of 'candidates' that satisfy all constraints. Since the inclusion of X/0 positions widens the scope of concatenation-based analysis, the present approach proves that large, if not all parts of `non-concatenative' morphology can be formally reduced to concatenative combination only: the new perspective has a gradient relationship between ontologically nondistinct morphologies with different degrees of grammatically controlled X/0 alternation. Furthermore, the fixed succession of obligatory vs optional positions that encodes the range of possible stem shapes in an intensional fashion (modulo constraint action) is shown to be derivable from surface forms: the constraint-based description of this basic range functions as a kind of nondistinctive generalized template. Since I assume an inheritance-based approach to the lexicon, some differences between verb classes (binyanim) are nonredundantly described as subtypes of this basic range. Other differences can be attributed to inherent shape and prosodic subcategorization requirements of certain affixes, in sum rendering the theory proposed a-templatic in character.

Elaboration of this theoretical proposal proceeds by building an detailed computational model of Tigrinya verbs. Tigrinya, an Ethiosemitic language, is spoken in Eritrea and northern Ethiopia by over 6 million speakers. There are up to 410 surface forms per verbal paradigm (Berhane 1991). Apart from analysing the contribution of root, stem vowels and affixes, the model accounts for both morphologically and phonologically-defined gemination patterns as well as spirantization. While the latter phenomenon has received considerable attention in the literature - serving as an argument for multiply-linked geminate representations in the face of inalterability effects - it is shown to yield easily to the new non-multiply linked geminate representation. Also dealt with is a case of systematic stem-internal reduplication. In addition, special emphasis is placed on developing a declarative reanalysis of vowel coalescence in so-called `hollow' verbs, where the participation of vocalic radicals creates dramatic effects regarding number of syllables and changed vowel qualities: mot-e `he died' vs. geref-e `he whipped'. Here the analysis succeeds in formally regularizing the descriptive irregularities.


BibTeX entry:
@PhdThesis{walther:97,
  author = 	 "Markus Walther",
  title = 	 "Deklarative prosodische Morphologie --
		  constraintbasierte Analysen und Computermodelle zum
		  Finnischen und Tigrinya",
  school = 	 "Philosophische Fakult{\"a}t der
		  Heinrich-Heine-Universit{\"at} D{\"u}sseldorf",
  year = 	 1997
}

Markus Walther
Last modified: Wed Jun 4 13:24:28 MET DST 1997