The application of two-level morphology 
to non-concatenative German morphology 
Harald Trost 
Deutsches Forschungszentrum fiir Ktinstliche Intelligenz (DFKI) 1 
Stuhlsatzenhausweg 3, D-6600 Saarbrticken 11, FRG 
Email: htrost@ sbuvax.campus.uni-sb.de 
.Abstract Introduction 
In this paper 2 we describe a hybrid system for 
morphological analysis and synthesis. We call 
it hybrid because it consists of two separate 
parts interacting with each other in a well- 
defined way. The treatment of morphonology 
and nonoconcatenative morphology is based on 
the two-level approach originally proposed by 
Koskenniemi (1983). For the concatenative part 
of morphosyntax (i.e. affixation) we make use 
of a grammar based on feature-unification. 
tloth parts rely on the same morph lexicon. 
Combinations of two-level morphology with 
t'eature-based morphosyntactic grammars have 
already been proposed by several authors (c.f. 
llear 1988a, Carson 1988, G6rz & Paulus 
1988, Schiller & Steffens 1990) to overcome 
the shortcomings of the continuation-classes 
originally proposed by Koskenniemi (1983) 
and Karttunen (1983) for the description of 
morphosyntax. But up to now no linguistically 
~;atisfying solution has been proposed for the 
treatment of non-concatenative morphology in 
:such a framework. In this paper we describe an 
extension to the model which will allow for the 
description of such phenomena. Namely we 
propose to restrict the applicability of two-level 
rules by providing them with filters in the form 
of feature structures. We demonstrate how a 
well-known problem of German morphology, 
so-called "Umlautung", can be described in our 
approach in a linguistically motivated and 
efficient way. 
1 Work on this project has begun while i was working for 
the Austrian Research Institute for Artificial 
Intelligence in Vienna, Austria 
2 I want to thank my collegues Greg Dogil, Wolfgang 
Heinz, Tibor Kiss ~md Gfinter Neumann for tYuitful 
discussions and helpful comments on earlier versions of 
this paper. 
Conventional morphosyntactic grammars do 
not allow to describe non-concatenative parts of 
morphology declaratively. Two-level 
morphology on the other hand can deal with 
some of these phenomena like vowel or 
consonant change, but there is no sound way to 
transmit information to the morphosyntactic 
grammar. This leads to quite unnatural 
solutions like the use of diacritics for the 
representation of morphosyntactic phenomena. 
While German morphology is mainly based 
on concatenation, some non-concatenative 
phenomena do exist. The most prominent 
example is so-called "Umlautnng". Umlautung 
means that in tile surface form the original stem 
vowel is replaced by another vowel in a 
systematic manner. Possible transformations 
are a=>~i, au=>~iu, o=>6, u=>ii, and - in some 
cases - e=>i. 
Umlautung in German realizes quite 
different morphological features. With nouns it 
can mark the plural either by itself (e.g. Mutter 
=> Mtitter) or in combination with an explicit 
plural-ending (e.g. Mann => M~inner), 
depending on the inflection class. With 
adjectives it is used to mark comparative forms 
(grog => gr6ger => am gr6\[hen), again in 
combination with an ending, for verbs 
following strong conjugation it marks the 
subjunctive ii and 2nd and 3rd person singular 
of the indicative present tense. Umlautung also 
occurs in derivation in combination with a 
number of derivational particles, e.g. -lich 
(klagen => kl~iglich). In contrast to its use in 
inflection, umlautung provides for no extra 
morphosyntactic information in derivational 
forms. At last, it appears in compounding in 
combination with some "Fugenelement" 
(joining element) (e.g. Miinnerchor - male 
chorus). 
1 371 
There are two common ways to cope with 
umlautung in conventional morphological 
components for German. One is to treat all 
forms created by umlautung as suppletions, i.e. 
these forms are explicitly entered into the 
lexicon. This is linguistically inadequate, 
because it obscures the phonological similarity 
of the two forms. From a more practical point 
of view it has the drawback that in a few cases, 
e.g. forming of the diminutive with the 
derivational -chen, umlautung is still 
productive, and cannot therefore be lexicalized. 
The other solution is a special function 
replacing (and interpreting) or generating the 
umhmt in all stems which are marked for 
umlautung required by the morphosyntactic 
context (c.f. Trost & Dorffner 1987). This 
makes umlautung a special case neglecting its 
status as a regular means of morphosyntactic 
marking. 
Solutions within the two-level approach 
have also been proposed. They rely on the idea 
to represent stem vowels which exhibit 
umlautung with special characters (diacritics) 
(e.g. A) at the lexical level. These characters 
are then realized as either the regular vowel 
(e.g. a) or the corresponding umlaut (e.g. t~) at 
the surface level. The idea behind is that these 
stem vowels are lexically somewhat 
underspecified. To trigger the appropriate 
substitution, G6rz & Paulus (1988) use a 
separate data structure to control for each word 
form which of the two possible rules is applied 
to create the surface structure. Schiller & 
Steffens (1989) use still another diacritic 
symbol for this task. Flexional endings 
triggering umlautung start with the diacritic $ 
(realized as 0 at the surface level). The context 
to the right of the substitution of all umlaut 
rules requires the occurence of that $. Therefore 
the umlaut rule would fail if no such affix 
follows the stem. As a consequence, the null 
morph must be explicitly represented by $ in 
lexical strings where morphosyntactic 
information is expressed by umlautung only 
(e.g. Mutter = > Mi~tter). 
Although both solutions certainly do work, 
at least for flexional morphology, they provide 
no clean and general solution for the integration 
of umlautung in the framework of two-level 
morphology. The use of a separate data 
structure is contrary to the intuition that 
umlautung is a regular phenomenon of German 
morphology, the treatment of which should 
require no extra mechanism. And the use of the 
diacritic $ places a burden on morphonology 
which clearly belongs to morphosyntax. 
The handling of non-concatenative 
morphological phenomena within the two-level 
approach imposes two new requirements: 
• Information about the application of a rule 
needs to be transferred to the morpho- 
syntactic grammar. 
• It must be possible to restrict the application 
of two-level rules to certain classes of 
morphs. 
Accordingly, we propose an approach where 
umlautung requires no extra mechanism at all 
and where no diacritics are (mis)used to 
describe morphosyntactic features. The basic 
idea is to provide two-level rules with a filter in 
form of a feature structure which controls its 
applicability. This feature structure has to be 
unified with the feature structure of the morph 
found in the lexicon to which the rule applies. 
In case of failure the two-level rule may not be 
applied. If unification succeeds infolmation is 
transfeiTed that way from the two-level part to 
the associated morphosyntactic grammar. This 
is crucial for the treatment of umlautung 
because, as mentioned above, its application 
conveys morphosyntactic meaning. 
In the following we will describe the parts of 
our system in some detail and explain how 
umlautung can be handled using that 
framework. (Basic knowledge of the two-level 
approach and feature-unification is 
presupposed.) We will also argue that 
extending the two-level rules with filters 
facilitates the description of certain 
morphonological phenomena as well. 
The Two-Level Part 
Our implementation of the two-level part is 
similar to the one proposed by Bear (1988a, b), 
i.e. rules are interpreted directly and not 
compiled into automata. Rules consist of a left 
context, a right context and a substitution. Left 
and right contexts are regular expressions over 
pairs of lexical and surface symbols. A 
substitution consists of exactly one such pair. 
Rules may be optional or obligatory (i.e. in 
contrast to Bear there are no disallowed rules). 
By definition, all default pairs are regarded as 
optional rules with empty contexts. 
The pair of strings (lexical and surface) is 
processed from left to right. If more than one 
372 2 
optional rule is applicable at a time this shows 
an ambiguity, 1.e. there are as many 
contmuatmns as there are different 
substitutions. Obligatory rules supercede all 
optional ones (thereby pruning the tree of 
continuations). If more than one obligatory rule 
is applicable at the same time (enforcing 
different substitutions) the whole mapping must 
be discarded. The same is true if no rule applies 
at all. 
The major difference from other two-level 
approaches is the possibility to provide the 
rules with a filter. A filter is an arbitrary feature 
structure. A rule may only be applied if the 
filter unifies with the feature-structure of the 
actual morph, i.e. the morph to which the 
substitution applies. Filters are used to restrict 
the application of a rule to certain classes of 
morphs. This is in contrast to the original view 
of Koskenniemi that morphonological rules are 
to be applied over the whole lexicon regardless 
of morphosyntactic considerations. This is 
certainly true of post-lexical rules. But there is 
evidence that it is not even true for all 
morphonological rules. Take e.g. the verb 
senden (to send), which can form two different 
past tenses send-e-te and sand-te, the former 
being regular weak conjugation, the latter a 
strong stem with weak inflection ending. The 
epenthesis of schwa (or e in orthography) 
depends on the morphological class of the stem 
(weak or strong). Or take the adjective dunkel, 
where the nominalization im Dunk-e-ln (in the 
dark) is different fi'om the attributive use den 
dunkl-e-n Mantel (the dark coat) (c.f. Gigerich 
1987). Here nominalization requires schwa 
epenthesis in the stem, not at the morph 
boundary like the adjective. 
If we want to use two-level rules for the 
description of non-concatenative morphology, 
such filters are necessary anyway. Because, as 
mentioned above, we do need some means to 
convey information fi'om the two-level part to 
the morphosyntactic grammar. In the case of 
umlautung we suppose that it is triggered by the 
concatenation of a stem which is lexically 
marked for umlaut (by the occurrence of a 
diacritical character A, O, U or E) with an affix 
allowing for umlautung (i.e. carrying the 
feature \[umlautung: +\]). Therefore the filter for 
all rules concerning umlautung basically 
contains the feature-value pair which marks 
affixes \[umlautung: +/-\]. 
Umlautung must only be performed if the 
stem allows for umlautung and that feature has 
the value +. Accordingly, all two-level rules 
substituting a vowel by its umlaut have the filter 
\[umlautung +\]. Corresponding rules are needed 
which keep the original vowel in the surface 
form. They have the filter \[umlautung: -\]. All 
the above-mentioned rules are obligatory, and 
exactly one of them applies to every occurence 
of a stern marked for umlaut (see figure 1). 
Rule: Rule Status: Rule Filter: 
_ \[agl' \[lnIllautung:-\]\]\]\] 
(A--')a) obligatory \[syn\[loc\[ head\[cat:stem\] \[agr \[umlautung: +\] \] \] \] 
Fig. 1: Two-level rules for lexical a 
To select exactly one rule (in order to 
prevent a deadlock between two obligatory 
rules) the feature umlautung must always have a 
value. But, as stated above, stems are not 
lexically marked for this feature. The marking is 
effected by the morphosyntactic grammar, 
which unifies the agreement features of affix 
and stem, thereby transferring the feature 
umlautung to the feature structure of the stem 
(see figure 5). 
What is important to note here, is that the 
two-level rules themselves contain no 
information about the morphosyntactic 
interpretation of umlautung. This is only 
specified in the grammar rules. Therefore the 
same two-level rules can be used for all the 
different places, where umlautung occurs. We 
will now shortly describe the morphosyntactic 
part of our system. 
Feattwe-Based Morphosyntax 
Concatenative morphology is described in 
grammar rules following X-bar theot'y. A 
head-driven approach is adopted. The basic 
structure consists of a head, usually some sort 
of affix, and one or more complements, one of 
which must be some type of stem. We will not 
go into any detail concerning the exact format of 
the grammar rules here, because it is irrelevant 
for the treatment of umlaut. For the purpose of 
this paper it suffices to give just one example. 
We will describe the overall structure of the 
grammar using the noun Mann (man) as an 
example. Nouns are constructed from stem, 
number marker, and case marker. The number 
marker forms the head and subcategorizes for a 
stem and a case marker. The relevant syntactic 
information is collected in the agreement feature 
which is passed upwards from the daugthers. 
3 373 
Figure 2 shows (a simplified version of) the 
number markers for |flex-class: er\]. We can 
see that the plural marker triggers umlautung, 
while the singular marker does not. Both 
subcategorize for a stem and a case marker. 
-morp_h: end* 
min: +J loc head \[cat: noun-stem\] 
~-num: sg J / agr~person: 3 
1_ l.umlautung: - 
syn ' \]_\] -bar|max:+\] e r\]'\] \] ~ lJJ head|Cat: stem firs1 sy~! loc ~t flex-class: 
L _agr O I 1 
firs sy. ,odheaJCat: o Hill | ~ n°un'rmxJ|l / ! 
res\] | LagrO JJJ 
- - crest: *end* 
\] morph: \[first: #er Lrest: *er t* 
-- |bar|max:.\] / Lmin: +J / 
loc \]head \[cat: noun-stem\] / 
\[ 5 num: pl ~ / |a~rtqP ers°n: 3 \] | 
E umlautung: +J J 
syr I- ~bar \[max: +\] \]l J\] lool head'Cat: stem firs syn .\] "lflex-class: er 
_ \[agr (D 
subca: \["- I I bar|max:+\] \]1\]\]\] firs, sy Io head|Cajun.flex 
res\] _ L l agre JJJ| 
_ L Lrest: *end* J l 
Fig.2: Number markers for nouns with plural -er 
(unmarked singular and plural #er) 
Figure 3 shows the lexical entry for the stem 
Mann, which may take an umlaut (its stem 
vowel is A). The number marker takes that 
stem as a complement. The agreement 
featuresare shared between head and 
complement. As one result, the 
featureumlautung is transferred from the 
number marker to the stem. It is now locally 
available to trigger the correct umlaut rule. 
- first: mAnn\] n~orph rest: *end* J 
r b rmax: +l 11 | ar \[min: +J | 
/ h -|cat:stem \]/ / syn loc | eaa \[root: manrd | ! 
\[flox-class: erl / / L "~," kgenus: n j j J 
-subcat: *end* 
Fig.3: Lexical entry for Mann (man) 
Figure 4 shows two different case markers 
for the unmarked case and for dative plural. 
After combining with a stem, the number 
marker may now take a case marker' as its 
second complement. 
I \[\[first: #n \] 
morph \[bar \[man:: +\] Lrest: *end*J 
l syn loc \]head|cat:noun-flex\] 
/a-~ \[casus: 3\] L ~" knum: pl\] 
-morph: *end* 
syn loc 
max: 
head \[cat: noun-flex\] 
Lnum: sg 4}1 
/ Inure: pl 4}j \] 
Fig.4: Case morphs for tile unmarked case and for 
dative plural 
The grammar fragment sketched in this section 
must be viewed with care. German inflectional 
endings often combine different rnorpho- 
syntactic information, e.g. with nouns case and 
number information is sometimes expressed by 
a single morph. Explaining the unmarked case 
as a combination of stem with nullmorph is also 
somewhat problematic. A more relistic 
grammar would probably collapse parts of the 
tree into a single structure, i.e. by using 
case&number markers, which subcategorize 
only for a stem. This would, for example, 
eliminate the problem which is posed to the 
parser by allowing for the occurence of more 
than one null morphs in one position (as is the 
case for 1st, 3rd and 4th singular where both 
number and case marker are realized by the null 
morph). Nevertheless, with regard• to the 
handling of umlautung in our approach these 
problems are not relevant. 
The Integration into the Grammar 
We will now show how the two parts of our 
system work together. Take e.g. the dative 
plural of Mann (man), M~innern. For 
generation, the grammar part constructs the 
lexical string SmAnn#er#n$ (# marks a morph 
boundary and $ a word boundary), which is 
given to the two-level part. The relevant lexical 
information for the purpose of umlautung is the 
stem vowel A in mAnn , and the feature 
\[umlautung: +\] in #er. As described in the last 
chapter, by structure sharing this information 
374 4 
has already been enriched by the generation 
process providing mAnn with the feature 
\[umlautung: +\]. When reaching the stem vowel 
A the rules try to unify their filters with the 
feature structure of mAnn. Only the umlaut rule 
succeed,,;, generating the correct surface form 
$mtinnern$. 
Now one can also see why the (incon'ect) 
form Manner will not be accepted by the 
parser. The filter of the obligatory rule A = > a 
would add the feature \[umlautung -\] to the 
feature structure of mann. This inhibits the 
unification with the feature structm'e of #er. 
syn 
- \[cat: noun 1 head \[person: 3J 
\[max:+\] loc bar Lmin: 
-i- 
l 
i ! 
i \[casus \[ \] I 
agr 0 |gend~! \] I |num k J \] 
Lumlautung \[ \] l 
..I \[ \[ I  a I atn°un*< 
• loc bar Lmin:-~\] syn loc\]bar \[max: +\] 
/ ~casus \[ \] q l Lagr'~. num \[ \] \] dgender \[ \] \] 
i ag,' num \[ \] / 
l-h" .\[cat: noun\] 
syn loc/bar \[max: +1 syn lee /b~ \[miax:+J 
Lagr qgender \[ \] \] + ,d-hum \[ \] agr 1 L 
~\[umlautung \[ \] 
Figure 5: Basic tree structure created by the grmnmar for nouns 
The analysis starts with the surface fern-1 
$mdnnern$. Because no morph has been 
recognized yet, both mAnn and miinn are 
derivable (because of the default rule d = > ~). 
At that point the morph mann is found in the 
lexicon. The filter is successfully unified with 
the lexical entry, transferring the information 
that umlautung has taken place. Now a morph 
boundary must be created. A 0 is inserted in 
the surface form which is mapped to # in the 
lexical form. (There still remains the other 
possibility to look for a longer morph. This 
hypothesis can only be discarded when the end 
of the surface form is reached without finding a 
corresponding morph in the lexicon.) Next #er 
is found in the lexicon. Since that ending 
allows for umlautung (having the feature 
\[umlautung: +\]), the morphosyntactic grammar 
can combine it with the stem which has already 
been unified with the nile filter. Next the morph 
#n is recognized, completing the analysis path. 
At this point we want to remind you of the 
fact that the morphosyntactic grammar uses the 
notion of null morphs for all unmarked forms 
(e.g. in our example the singular marker). Null 
morphs are not included in the lexical string 
though. They operate solely at the level of the 
morphosyntactic grammar. Take e.g. the 
generation of Mann (nora sg). Although only 
SmAnn$ is created as lexical string, the null 
morph has enriched the associated feature 
structure with \[umlautung: -\] enforcing the 
generation of the surface string Mann by 
blocking the umlaut rule. 
Analysis works in a similar way. When 
$mann$ is input as surface string, it is mapped 
to the lexical string unchanged. It is now 
associated to the morph mann the feature 
structure of which has been unified with the 
rule filter providing it with \[umlautung: q. To 
create a legal word form it must now be 
5 375 
combined with number and case markers. 
These can only be null morphs and their 
agreement features must unify, which leads to 
the correct interpretation. 
Another example from derivation shall 
demonstrate how well the marking of stem 
vowels and the feature umlautung work 
together to define the occurence of umlautung. 
The verb klagen (to mourn) shows no umlaut in 
any of its forms. The same is true for the 
nominalization Klage. But the derived adjective 
kldglich surprisingly exhibits an umlaut. A 
closer look shows that this behaviour conforms 
to what our system predicts. The morph kIAg 
is marked as a stem which may take umlaut. 
Since all endings of weak verb conjugation are 
marked with \[umlautung ~\] no umlautung takes 
place for any of the verb forms. The same is 
true for the noun plural ending/tn. But #1ich 
comes with the feature \[umlautung +\] triggering 
the umlaut-rule to produce the surface form 
kldglich. 
Unfortunately in derivation and composition 
there are exceptions to the nile. Contrary to our 
expectations we find the adjective handlich 
derived from the noun Hand. Since the plural 
form of the noun is Hdnde the morph must 
clearly be stored as hAnd in the lexicon which 
would yield hdndlich which is incorrect. There 
m'e two solutions to this problem. One can take 
the stance that in such cases derivation is no 
longer transparent and that these words should 
be entered into the lexicon as a whole. 
The other solution would be to introduce 
exception markers with such morphs which 
block the application of the umlaut rule (say 
\[flex-uml-poss: -\] for flexion and \[deriv-umP 
poss: -\] for derivation). Instead of the single 
feature umlautung for all affixes we then need 
to mark flexional endings and derivational 
endings with the features flexional-umlautung 
and derivational-umlautung respectively. The 
rule filters become more complex too. Umlaut 
rules are equipped with the following filter: 
{ \[flexional-umlaut: + \] \[flex-uml-poss: +\] } 
\[derivational-umlaut: +\] \[deriv-uml-poss: +\] , 
the corresponding no-umlaut rules get an 
according one. All morphs not explicitly 
marked will behave like before, i.e. take 
umlautung in both cases. 
Conclusion 
We have shown a hybrid system for 
mou~hological analysis and synthesis, based on 
two-level morphology and unification-based 
376 
grammar rules. By providing two-level rules 
with a filter in the form of a feature structure the 
application of these rules can be controlled by 
the morphosyntactic grammar in a consistent 
way. The filters are also used to transfer 
morphosyntactic information from the two-level 
part to the grammar. This allows the description 
of non-concatenative morphological phenomena 
using such rules without the use of 
(phonologically) unmotivated diacritics. 
As an example, we have shown how our 
system can handle German umlautung in a 
linguistically satisfactory manner. Translation 
of the umlaut is performed by a two-level rule 
which is filtered by a feature umlaumng. The 
morphosyntactic interpretation of the umlaut is 
only performed at the level of the grammar 
rules. 
The proposed method can be applied to other 
non-concatenative phenomena as well. The 
idea of filters seems also to be a promising 
solution for morphonological phenomena 
which are restricted to certain classes of 
morphs (or words). 

References

Bear J. (1988a): A Morphological Recognizer with 
Syntactic and Phonological Rules, COLING-86, 
Bonn, BRD. 

Bear J. (1988a): Generation and Recognition of 
Inflectional Morphology, in: H.Trost (ed.), 
4.0sterreichische Artificial Intelligence-Tagung, 
Springer, Berlin, 3-7. 

Carson J. (1988): Unification and transduction in 
Computational Phonology, COLING-88, Budapest, 
106-111. 

Gigerich H. (1987): Zur Schwa-Epenthese im 
Standarddeutschen, Linguistische Berichte 112,449-469. 

G0rz G., Paulus D. (1988): A Finite State Approach to 
German Verb Morphology, COLING-88, Budapest, 
212-215. 

Karttunen L. (1983): KIMMO: A General 
Morphological Processor, Texas Linguistic Forum 
22, 167-186. 

Koskenniemi K. (1983): Twoolevel Model for 
Morphological Analysis, UCAP83, Karlsruhe, 
BRD, 683-685. 

Schiller A., Steffens P. (1990): A Two-Level 
Morphology for a German natural language 
understanding system, IBM Stuttgart, manuscript. 

Trost H., Dorffner G. (1987): A System for 
Morphological Analysis and Synthesis of German 
Texks, in: D. Hainline (ext.), New Developments in 
Computer-Assisted Language Le0Jaaing, ('rooms 
Helm, London. 
