Morphology with a Null-Interface 
Harald Trost and Johannes Matiasek 
Austrian Research Institute for ArtificiM Intelligence.* 
Scllottengasse 3, A-IOIO Vienna, Austria 
{harald, john} 0ai. univie, ac. at 
Abstract 
We present an integrated architecture for 
word-level and sentence-level processing in a 
unification-based paradigm. The core of the sys- 
tem is a CLP implementation of a nnilication en- 
gine for feature structures suplmrting relational 
values. In this framework an IiPSC,-style granl- 
mar is implemented. Word-level processing uses 
X2MoltF, a morphological component I)ased on 
an extended version of two-level morphology. This 
component is tightly integrated with the grammar 
as a relation. The advantage of this apl)roach is 
that morphology and syntax are kept logically au- 
tonomous while at the same time minimizing in- 
terface problems. 
1 Introduction 
Over the last few years there has 1)eeu a growing 
interest in computational morphology and phonol- 
ogy. A number of systems have been developed 
that deal with word-level processing. A widely 
used approach is finite-state morphology, most no- 
tably two-level morphology (for an introduction, 
see Sproat 92). Morphological components are 
sueeessflflly used for a wide range of stand-alone 
applications like sl)elling correction aM hyphen- 
ation. One obvious application is the use in NI,P 
systems geared to the analysis/generation of text. 
Surprisingly, they have not been widely al)l)lied in 
this domain up to now. 
A major reason for this is the llrolllem of 
interfacing morphology with syntax. Reflecting 
tile current trend in syntax towards lexicalism, 
unification-1)ased systems use highly structured 
feature structures as inI)ut. Translating tile out- 
put of morphologieM components into such it rel)- 
resentation has proved to be diiticult. Reducing 
interface problems is therefore crucial to success. 
*Financial support for the Austrian Research Institute 
for Artificial Intelligence is provkled by the Austrian Min- 
istry o\] Science and Research. We wouhl like to thank 
Wolfgang lleinz for vMuable comments and suggestimls 
A tight integration between word and sen- 
tence level processing also has linguistic advan- 
tages. The boundary between morphology and 
syntax is fuzzy. When processing written text the 
units nmrl>hology has to deal with are, ill it tech- 
nicM sense, not words Mt character strings sepa- 
rated by delimiters. While these strings roughly 
correspond to the words of a sentence there are 
problematic cases. In German, e.g., zu-infinitive 
or w'rbs with separMfie prefixes ;Lre written as a 
single unit in some instances and separately ill 
others. 
The prol)lem has boon recognized and seine 
possihle remedies have been prol)osed. They all 
try to minimize or to elhninMe the intel'r:tce be- 
tweell word and sontoiic(: low4 processing. One 
stop is the descriptiml of word fl)rmation ill terms 
of a unification-based gl'all/lnai' to make the result 
(~1' morphological l)rocessing dir(,ctly ~wMhd)le to 
syntax and vice w~rsa, an :g)l)roa(-h ah'eady ti~ken 
in X2Moltl,' (Trost 90, '\['rost 91), an extension of 
two-hwel nmrphohlgy. 
The harder probhml is the integration of mor- 
phol)honology which is traditionally formalized in 
it way not easily t,':mshmdlle into the fei~ture for- 
malisnx. We will show how this can he achieved 
by merging the word-level grammar of X2MolI.I,' 
into an lll'S(;-styl0 gralnnla.r, alld by adopting a 
relational view of its two-level rules. 
1l, this llal)or we ass/lille basic flmfilhu'ity with 
unification-ha.seal N I,P techniques and two-low~l 
lnorphology. 
2 Integrating Morphology into 
I IPSG 
llead-driwm Phrase Smwture Grammar (IIPSG, 
Polhu'd ,~ S;tg 87, l'ollard &. Sag in press) can 
be viewed as a mono-level Mt multi-stratal the- 
ory Of gl'anlnlitl', where different stratit relate to 
different aspects (d" linguistic informlttion, but are 
ropresonted uniformly in feature logics. As such 
it is well suited as a linguistic theory for our en- 
141 
terprise. 
ItPSG differentiates between three strata--- 
PIION, SYNSEM and DTItS. Though morphology is 
not considered in the standard approach, it sug- 
gests itself to be included as a fourth stratum by 
introducing a feature MORPH into the type sign. 
Morphotactics are easily described in terms of a 
feature based grammar. The problem is how to 
deal with morphophonology. Two proposals have 
been made to overcome this problem. 
Krieger et al. 93 encode finite state automata 
directly in the feature formalism. Since two-level 
rules can be compiled into such automata, roof 
phophonology can be straightforwardly integrate<l 
into the grammar. While this is formally elegant it 
seems to be no good solution for practical consid- 
erations. First, it is not entirely clear frmn their 
paper how the problem of null characters can be 
handled. Second, encoding large automata will 
result in a very large and unwieldy type hierar- 
chy. In general, introducing automata into fea- 
ture structures and encoding morphophonology 
directly at that level seems to be too low-level. 
Bird & Klein 93 argue against the use of 
two-level morphology because of linguistic con- 
siderations. The linguistic background of two- 
level rules--main stream segmental phonology-- 
has widely been rejected as a valid linguistic 
model. Instead, they base their implementation 
on autosegmental phonology (of. Goldsmith 90). 
This is certainly linguistically appealing. But 
there are reasons for sticking to a more conser- 
vative approach. Finite-state morphology as a 
formalism is not necessarily tied to segmental 
phonology. There are various approaches to cope 
with non-concatenative phenomena--one of tlmm 
X2MoltF (Trost 90). Also, for a nmnber of lan- 
guages complete sets of two-level rules do exist 
and can immediately be brought to bear. Finally, 
finite-state morphology has proven to be eflicient 
while the method proposed by Bird ~: Klein 93 
seems to be computationally costly. 
Like the other apl>roaches ours is also based on 
IIPSG. Itowever, we employ a different approach 
to integration. Our grammar is encoded using a 
unification engine based on constraint logic pro- 
gramming (CLP). Besides conventional attril)ute- 
value descriptions this system allows for the direct 
representation of more general relations, as they 
are required by IIPSG. This extension of the for- 
malism is used for the integration of morl>hology. 
Thus X2MoI~F is treated as one special relation 
of tile grammar. As a result, our approach is more 
modular than the others. While being fully inte- 
grated morphology can still be viewed as an au- 
tonomous component leading to a more llexible 
design. 
We will now give an overview of X2MolIF he- 
fore describing the integrated system and its im- 
plementation in detail. 
Word Level Processing 
X2MoRF 
X2MoRF differs fi'mn standard two-level mor- 
l>hology hi two hnportant respects. Continua- 
tion classes are replaced by a feature-based word 
grammar. This allows for a more line-grained 
description of morphs. It is also a prerequisite 
for a tight integration with a unification-based 
grammar. X2MoRF uses a morph lexicon where 
each morl>h has one or more feature structures 
assigned. The word grammar itself is shnple. 
Morphs have a functor-argument structure along 
the lines of di Sciullo ,~: Williams 87. Affixes are 
unary functors while stems are arguments with- 
out ally further structure, resulting in a hinary 
tree structure. 
Tile othe,' extension concerns the two-level 
rules, which are supplemented with a morpholog- 
ical filter consisti,g of a feature structure. This is 
hnportant because in morl~hol>honology only some 
rules are purely l)honologically nmtivated. Oth- 
ers are triggered by a mixture of phonologicM and 
morphological facts. Such rules cannot be prop- 
erly represented in tile standard al>proach. 
TAD,, e.g., umlaut an<l schwa epenthesis in 
(lerman: The third I>erson singular present tense 
suffix for (;er,nan w~rbs is 4, e.g., sag-t --+ sagt. 
For stems ending in a dental, schwa is inserted 
befi)re the ending, e.g., bad-t -, badct. This rule 
does not hold across the whole vocabulary though. 
Stems of the stroTIg paradignl <1o exhibit umlaut 
in 3,'dPersSgPres which blocks schwa epenthesis. 
The /inal dental of tile stem must be omitted in- 
stead, e.g., rat-t --+ riil. 
The th,'ee miles I shown in Fig. lItogether 
with the a.pl)ropriate entries in the morl>h lexicon 
(el. Fig. 7 below)--produce the re<luired behav- 
ior. In particular, these rules relate surface Nit to 
lexical $rAt÷t$ 2. X2Mold '~ can be seen as a re- 
*These rules as well as other data presented in the e×: 
amples are simplified for the purl)ose of denmnstration 
2The lexical character A may have the surface realiza- 
742 
(i) A:(L ~ _ j' \[MORPIIIMIII':AI)IUMLAIIT aolt-lt111la*tl\] 
(ii) t:O ¢=::V _ 4.:0 t 
(iii) + :e ¢=e,. dental _ +:0 \[s It\] ; \[MORPHIMtII.;ADIEPl.;NTIIESI.; -\[-\] 
Figure \[: Three extended two-level Rules 
lation between a surface string (the word i~orm), a 
lexical string, and a feature structure (tim inter- 
pretation of the word form). Relevant for sentence 
level processing is the morl)hosyntactic informa- 
tion and the stem, found as the values of paths 
MOltPlI \[MIIEAI) and MOR, Pll \[STEM resl)ectively (of. 
Fig. 9 below). This is supplemented by lexeme 
specific information in the value of SYNSF:M (for a 
detailed description see Trost 93). 
4 Implementing HPSG in a CLP 
Framework 
IIPSG employs strongly typed feature structures 
together with principles constraining them fur- 
ther. Well-typcdness requirements restrict the 
space of valid feature structures (cf. Carl)enter 
92): Every feature structure must I)e associated 
with a type, and every type restricts its associ- 
ated feature structure in that only certain features 
are allowed and the values of these features must 
be of a certain type. Appropriateness and value 
restrictions are inherited along the type hierarchy. 
The second source of constraints, in order to 
admit only linguistically valid feature structures, 
are the principles of grammar. Pollard ,~ Sag 87 
allow general implicative and negative constraints 
in the form of conditional feature structures. In 
Pollard h Sag in press principles are given only in 
verbal form. Recent work on formalizing the basis 
of IIPSG models them as constraints attached to 
types (e.g., Carpenter et al. 91). Iiowever, these 
distinctions affect only how the applicability of 
a principle is specified. More iml)ortant for our 
present purpose is the form which the constraints 
expressed by a principle may take. Besides con- 
straints enforcing simple structure sharing (e.g., 
the Head Featnre Principle given in Fig.2) there 
are also complex relational dependencies (e.g., in 
the Subcategorization Principlea). Constraints 
tions a and d. The rule ha.s an empty phonological context 
but a morphological filter. This is an example for the treat- 
meat of non-concatenative phenomena in X2MonF. 
3,in a headed phrase (i.e., a phrmsal sign whose DTRS 
value is of sort head-struc), the suncAT value of the head 
like these go beyond the exl)ressivity of l)ure Da- 
ture formalisms alone and need to be defined in a 
recursive manner. 
In order to integrate such complex constraints 
in the feature unification framework we interpret 
unitication of typed feature structures under the 
restrictions of princil)led constra.ints as constraint 
solving in the CLI' paradigm (Jafl'ar ,~ Lassez g7). 
In CI,P the notion of unification is replaced 
by the more general notion of constraint solving. 
Constraint solvers may be embedded into a logic 
l)rogramufing language either by writing a nteta- 
itlterl)reter or by urn.king use of a system which 
allows \[or the impletn(mtation el + unification ON- 
tollS\]OilS. 
The s~,cond approacls is taken by I)MCAI 
(:l,l)4 (l\[olzbaur 92), it l)rolog system whose uni- 
tication mechanlsnl is extended in such a way 
that the user may introduce interl)reted terms 
and specify their meaning with regard to uni\[ica- 
tion through l'rolog predicates. The basic mech- 
anism to a chiow, this behavior is the use of at- 
tributed variables, which may l)e qualified by ar- 
1)itrary user-defined attributes. Attributed vari- 
ables behave like ordinary l)rolog variables with 
two notal)le exceptions: when an attributed vari- 
able is to be unified with a non-wu'iable term or 
another attril)uted variable the unifi('.atk)n exten- 
sions come into play. For either case the user 
has to supply a predicate which explicitly specifies 
how the attril)utes interact and how they should 
1)e interpreted with respect to the semantics of 
the al)l)lication domain. Unilication succeeds only 
if these constraint solving clauses managing the 
cond)inati,m -el' vm'ification--af the involved at- 
tril)utes are successfid. 
The iml)hm~entati(m of typed feature struc- 
tures in our system makes use of the CLP facilities 
provided by this enhanced Prolog system. Fea- 
ture structures are imlflemented by the attril)ute 
daughter is the concatenation of the phrase's SUBCAT llst 
with tile list (in order of incre~Lslng obliqueness) of SYNSEM 
values of the COml)hmlent daughters."(Pollard & .gag in 
I)ress) 
4 I)MC, AI CLP is au enhanced version of SICStus Prolog, 
awdlahle by anonymous ftp from f tp. ai.univie, ac. at 
143 
:fs (Type, Dag, Goals), where Dag is a list of feature- 
value pairs (which is empty in case of atomic 
types) or a marker indicating uninstantiatedness 
of the substructure (feature structures are instan- 
tiated lazily). Goals is a list of delayed constraints 
(see below). Well-typed unification of two feature 
structures is implemented via the constraint solv- 
ing clauses mentioned above, taking into account 
type hierarchy and feature appropriateness (for a 
detailed description cf. Matiasek & Ileinz 93). 
Constraints imposed onto feature structures 
by the principles of grammar are stated in a con- 
ditional form where the antecedent is restricted to 
contain only typing requirements. 5 In order to ac- 
count for these conditional constraints we adopt a 
licensing view: Every node of a feature structure 
has to be licensed by all principles of grammar. 
A node is licensed by a principle if either (i) 
the feature structure F rooted in that node sat- 
isfies the applicability conditions of the l)rinciple 
and the constraints expressed by the l)rinciple suc- 
cessfully unify with F, or (ii) the feature structure 
F rooted in that node is incompatible with the 
applicability conditions of the principle. The in- 
teresting case arises when a feature structure does 
not satisfy the applicability conditions of the l)rin- 
ciple but is compatible with them. Thus applica- 
bility of tile principle can be decided only later, af- 
ter further instantiation or unification steps have 
restricted the (sub)structure rooted at that node. 
In precisely this case the application (or the al)an- 
doning) of the constraint has to be delayed. The 
delay mechanism utilizes tim Goals slot in the 
fs/3 6 attribute, which is dedicated to hold the 
delayed constraints. As an example take the well 
known Ilead Feature Principle of IIPSG (Fig.2) r. 
The conditional operator ===> is translated at 
read time via terra_expansion/2 and implements 
the delay mechanism by coml)iling l>recon<lition 
checks into the principle. These antecedent checks 
trigger either the application of the princiltle, its 
abandomnent, or its delay (by annotating the vari- 
ables wlfich are not sufficiently constrained to de- 
cide on the antecedent with the delayed goals). 
Two advantages of this approach to implement 
SThis is only a syntactic variant of attaching constraints 
solely to types (Carpenter et al. 91) and does not permit 
general conditional structnres ms used in Pollard & Sag 87. 
6pred/n is the usual notation for a n-ary Prolog 
predicate. 
VThe operators ::,,, : :, :, === are defined for typing 
of a node, path restriction, path concatenation aim value 
restriction (type or coreference) respectively. 
A VM: 
\[SYNSEMILOCICATII'E D \[\] headed- LD'I'ItS\[I \[I';AD-DTIII SYNSEMILO<':ICATII1EAI) \[~\] \] 
phrase 
Proloq: 
head_feature_principle(X) :- 
X::=headed_phrase 
===> 
X::synsem:loc:cat:head===H, 
X::dtrs:head_dtr:synsem:loc:cat:head===H. 
Figure 2: Head Feature Princil)le 
principled constraints are especially important for 
our present purpose: First, stating redundant 
typing re<luirements for embedded structures (i.e. 
type restrictions that would follow automatically 
from well-typing) forces delay of the conditional 
constraint until these sul)structures are instanti- 
ated. This device can, e.g., be used to block in: 
finite recursion in recursively detlned constraints. 
Second, the right hand part of the conditional is 
not restricted to feature logical expressions, but 
instead can contain arbitrary Prolog goals. In 
this way constraints involving relational depen- 
dencies (such as the Subcategorization Principle 
and the morl>hological relation between a lexical 
and a surface string) can be expressed within the 
feature fornmlism and there is no need for external 
devices controllh~g this interaction. Furthermore, 
the conditional constraint syntax is not restricted 
to unary licensing principles but can also be used 
to express relations, such as *s_append/3--needed 
for implementing the Subcat l~rinciple--which ap- 
l>ends two feature structure lists (Fig. 3). Note 
fs_append(X,Y,Z) :- 
fs_empty appond(X,Y,Z), 
fs nonempty append(X,Y,Z). 
fs_empty append(X,Y,Z) :- 
X::=elist 
===> Y = Z. 
fs_nonempty_append(X,Y,Z) :- 
X::=nelist 
===> X::first===F, Z::first===F, 
X::rest===XRest, Z::rest===ZRest, 
fs_append(XRest,Y,ZRest). 
Figure 3: A1)pend for feature structure lists 
that disjunctiw~ relations such as append call now 
l)e written as tile conjunction of two specialized 
cases applying conditionally. Furthermore, in- 
144 
Inp u l <=> t:O <=> \['+':0, t:t\]. 
morphrule(\[llG,43,116ILS\],\[Sc,dS,1161SS0\],SS,LCon,SCon,F) :~ 
!, Sc=48, 
morphology(\[43,116II,S\],\[48,1161SS0\],SS,\[l16ILCon\],\[HISCon\],F). 
Compiled 
Figure 4: Sample Two-l,evel R.le 
morphology (LexStream, SurfStream0, SurfPlainln, LexContext, SurfCont ext, F) : - 
instantiate(LexStream,SurfStroam0,SurfPlainln,SurfPlainOut ,F) , 
morphrul e (LexStr earn, Surf Sir earn, Surf Pla inOut, LexCont ext, Surf Cont ext, F ). 
instantiate(\[LCILCs\] , \[SCISCs\] ,SurfPlainIn,SurfPlainOut,F) :- 
valid alphabet_pair (LC, SC\], 
synchronize ( \[SC I SCs\] , S1trf PlainIn, S~irf Plain0ut ), 
lookahead (LC, LCs, SCs, Surf PlainOut ) . 
~ynchronize(\[481_\],Stream,Stream) :~- !. 
synchronize( \[Char I_\] , \[Char I Stream\] , Stream) . 
Figure 5: The morphology rela.t, iml 
finite loops due to uninstantiated variM)les can 
never occur, a cruciM requirement when integr;tt- 
ing relational dependencies into st lazy instantiat- 
ing feature formalism. 
5 Embedding X2MouF into the 
Feature System 
Originally X2MoItF was realized ~ts st separate 
morphological component interfaced to the sen- 
tence analyzer/generator only via seq.ential (lat~ 
transfer. In the case. of analysis, the feature strm> 
ture representing the word form was transmitted 
to the parser. For generation, X2MoRF' expected 
a feature structure as inlmt reproducing one or 
more word forms. This l)urely sequential architec- 
ture was not satisfactory lmcauso, of the l)roblems 
mentioned in the introduction. 
In order to achieve tight integration, we a(h|l)t 
a relational view of X2MoM;' and encode the re- 
lation between surfiLce string and lexical string <li- 
reetly without using finite state automata (for ar- 
guments suI)porting this ai)pro~Lch of. A1)ramson 
92). Ilowever, our al)l)roach extends A1)ramson 
92 in that it (i) explicitly accounts for the inser- 
tion of null characters and (ii) introduces the filter 
concept of X2MoltF into the relational approach. 
The general format of a two-hwel rule sl)eciIi- 
cation in our system is 
LCon <=> Transition <=> RCon \[:- Filter\] 
in the case of equiwdence rules, option:d rules 
are written using only single arrows (=> and <=). 
These rules aro cmnpiled into Prolog c\]a,uses 8 re- 
buting the \]oxical and surfaco character streams 
appropriately (see Fig.,1 for an example of the l- 
elision rtlle f'or (\]erlllatl). 
q'- obtain a c.rroct relathmship between sur- 
fa, c(, and h~xi('al string ewwy transition has to be 
licensed I)y st morphological rule. Transitions not 
mentioned by rules are handled by a defitult rule. 
Instantiation of contexts may not be done by the 
rules themselw,s, since this would make it impos~ 
sible to obt~dn negation via the cut-operator. In- 
stead, it is handled Sel)a.rately in a backtrackable 
f~shion. 
The central relation is the morphology predi- 
cate, (soe l"iI';. 5) nledia.tiug between lexical string, 
surf:we string (with inserted n.ll elements), the 
puro (dellullifiod) surl~w(~ string and the feature 
structure of the morl)hologica.l sign. lnstantiation 
of pairs is done del)onding on the possible lexi- 
cal con|in,rations (the lexicon bei,g represented 
by a trie-structure). The amount oflookahead is 
detormilmd by the current pair which is to be li- 
censed by morphrule. '~ Synchronization of s,rface 
and lexical string by insertion of mill characters is 
also ha.lMlml a.t the insta.ntiatlon hwel. 
Tim intogra.tion (ff tim two-hwel rohU, ion into 
KN(II~! l.hat \]e\[t Collie×Is are encoded reversed to ac- 
count for the lelt to rlg, ht traw~rsa\] af the pair of character 
s\[.r(!itlllS;. \[J('.\[t COIItCXts C~tll be rellleHIbered 3.1ld c\]lecked 
most efficiently this way. 
9This interactio, and the lexicon look,p of the feature 
structure corresponding to the current morph, which takes 
pl;~.(:e whell (:llCOllllt(~l'illg ;k IIIOl'ph bOlll/dAry iS not .'4hOWlI 
for the sake of simplMty. 
145 
the general framework of the feature based 
sentence-level and word-level grammars is now 
performed by adding this relation as a principled 
constraint at the appropriate level. 
In a definite clause style AVM notation 
this could be written as follows (given that 
morpho:l.ogy/a is a wrapper around the morphol- 
ogy relation given above, starting with empty left 
context and hiding the nullified surface stream): 
rPnON ~\]strlng 
/ 
msignLMm~AO mheod j 
|lEAD heod 
wordLSYNSEM synsem 
The actual imi)lementation as a princil)led con- 
straint in our formalism additionally takes care of 
delaying the actual enforcement of this relation in 
case the strings are not sufficiently instantiated. 
A second provision has to be made in the 
word level grammar to assnre prol)er concatena- 
tion of the lexical strings of the morl)hological 
signs being combined. Given the subtyping of 
resign into marg and mfunctor, which in turn 
has the suhtypes leftfunctor and rightfunctor, the 
principled constraints ensuring concatenation of 
a left functor with its argument are shown in 
Fig. 6. Concatenation is delayed until the ar- 
concat_right_gunctor(X) :- 
X::=rightfunctor, 
X::arg:mstring===subtype_of(string) 
===> 
X::arg:mstring===hrg, 
X::affix===Suffix, 
X::mstring===Mstring, 
concat(hrg,Suffix,Mstring). 
Figure 6: Concatenation of lexical strings 
gument's MSTRING is instantiated. Thus, infinite 
loops when concatenating are avoided. 
As an example we demonstrate how these con- 
straints interact in forming the third person sin- 
gular present tense form of the German verb raten 
(to guess). The lexical string is composed of the 
stem rat and the suffix +t. The lexical entries of 
these two morphs are given in Fig. 7. 
The two-level rules applicable for this examl)le 
are the t-elision rule (Fig.4) and two rules with 
filters licensing a-umlavt and epenthesis, given in 
rightfunctor 
I MSTRING "rAt" 1 STEM "rot" | LUMLAUT aou_umlautJ J verb_stem 
m a rg 
'STEM \[~string 
AH"IX " 4"t" 
M,,,.:A,) |PERSON 3 / 
LIIMI,AUT \[~\]umlautJ verb_form 
I-s'r,.:M \[\] l 
ARG / \[EPENq'HESI'  1/ /MI~,,,D / 1-71// 
IIl~/#/II L tJt?I'b_st¢;TII"UMLAUT u a .1 
Fignre 7: Lexical entries 
the input notation for our system (Fig.8). 
lnteractkm between syntactic and morpholog- 
ical processes takes place at the word level. The 
apl)lieation of the two-lew'J rules relating the sur- 
face string (i.e the pllON-value of the word) and 
the lexic:d-string (i.e. MORP,IMsTRINC ) is also 
triggered here. This interaction is completely neu- 
tral with respect to the direction of l)rocessing due 
to its relational nature. Parsing is performed by 
simply instantiating the PIION value. Generation 
can be achieved when MORPIIIMSTRING is present, 
which in turn is obtained by concatenating the 
lexical strings of the resigns instantiated by the 
morph grammar. 
As a result of this constraint interaction the 
structure shown in Fig. 9 is obtahmd. Features 
relevant at the syntactic level (such as PERSON 
and TENSE) are percolated fi'om MORPIIIMI1EAD 
to SYNSI,:MILoc\[CAT \[ll~,:al) via structure sharing 
constraints attached to the type word (this in- 
teraction is not shown in Fig. 9). Information 
on sul)c;d,egorization and semantic content for the 
word is obtained fi-om the lexeme lexicon using 
MORPIIIsTNM aS a key. These constraints con> 
plete the interaction between syntactic and mor- 
phological processing at the word-level. 
6 Conclusion 
We have presented a fl'amework for" tile tight inte- 
gration of word level and sentence lew.q processing 
in a unification-1)ased paradigm. Tile system is 
built upon a unification engine hnl)lemented in a 
CLP language supporting types and delhfite rela- 
tions as l)art of feature descriptions. Using this ex- 
146 
A-umlaut _ <=> h:"a <=> _ :- filter(X, \[X::mhead:umlaut===aou-umlaut\]) 
Epenthesis dental <=> '+':e <=> s_or_t :- filtor(X, \[X::mhead:openthose==='+'\]) 
Figure 8: F'ilter lhiles 
"PIlON "rift" 
"MSTRING "rAt+t" 
STEM \[~'l"a t" 
AFFIX "4-t" 
FEPENTnESE~" \] 
MIIEAD |PERSON 3 / \[TI'~NSE tense_pre~ \[ 
verb- LUMLAUT \[~ft O ll_tl llll(l It t_\] 
SOFIII MORPII 
FMS'rItIN<I ",'At" \] 
All.(; /MI1EAI) \] i>ERsON 3 I \] 
L stem J marg 
righ t-' 
word functor 
Figure 9: Result of constraint ini;eri~(:tion 
tended feature formalism, which is indel>en<h,ntly 
motivated by reqnh'enlents of standard Ill'St;, iL 
reiml)lementation of X2MoRF was integr,~ted into 
the grammar as a specialized relation. 
This architecture has computational as well 
as linguistic advantages, integrating morphology 
and morphophonology directly into the gralnmar 
is in the spirit of IIPSG, which views gramm:cr as a 
relation between the I)honological (or gr~qllielnic) 
form of an utterance and its syntactic/seniantic 
rel)resentation. This way the treatnmnt of phe- 
nomena transcending tile boundary between in()r- 
phology and syntax is also nlade l)ossil)h~. 
On the implementation si(h~, the practi('al 
problems of interfacing two inherently difl'erent 
modules are eliminated. For applh:ations this 
means that using a morl)hological component is 
made easy. Nevertheless, this tight integration 
still leaves morphology and syntaz/somantics as 
autonomous comI)onents, enal)\]ing direct use ,)1' 
existing data sets descril)ing morphopholmh)gy in 
terms of the two-level p,~ra(ligm. 
References 
Abramson If.: A Logic Programmiilg View of l(.e- 
lational Morl)hology, in Proceedings of the 15th 
COIANG, August 2?,-28, 1992, Vol.l II, pp.850-854, 
1992. 
Bird S., Klein E.: l'\]nriehing IIPSG Phonology, 
University of l'\]dildmrgh, UK, Research Paper 
I'~UCCS/I~ 1'-56, 1993. 
Carpenter B., Pollard C., Franz A.: 'the ~pecification 
and Implementation of Constraint-Based Unifica- 
tion C, rammars, Proceedings of 2 '*'I IWPT, Can- 
curt, Mexico, 143-153, 1991. 
C, arl>enter B.: The Logic of "l~ped I"calure ,S'h'uclures, 
(~amhridg,, \[hfiversity Press, Caml)ritlge Tracts in 
Theoretical (~Omlmter Science 32, 1992. 
(ilolclsmh.h J.A.: Anloseymcnhd and Metrical l'hovol- 
ogy , Basil Blacl(well, Oxford, 1990. 
llolzbaur C.: Metastructures vs. Attrilmted Vari- 
ames in the (~ontext of Extensible Unification, in 
lh'uynooghe M. and Wirsing M.(eds.), l)rogram- 
ruing Language hnplementation and Logic Pro- 
gramming, Springer, LN(JS 6:/1, pp.260-268, 19!12. 
Jatl'ar J., lmssez .I.L.: CoIisl.raint Logic Programming, 
in I'ro('eodings 14t.tl ACM IK)PI, Conf., Muifich, 
1987. 
Kriegor II.-U., Pirker lI., Nerbolme J.: Fe~ture-bmsed 
Allomorldly , Ih'oceedings of the :l 1st Ammal Meet- 
ing of the At'L, (k)hmll)us, Ohio, Pl). 140-147, 1993. 
Matiasek J.,. lleinz W.: A CLI ~ Based Approach to 
lIPS(;, ()sterrei(:hisches Forschungsinstitut fiir Ar- 
tificial h,telligen(:e, Wien, TI(-93-26, 19!):{. 
l>ollard C.J., Sag I.A.: hi formation-Based 5'ynlaz ond 
5'emmdics, Uuiw'rsity of C\]licago Press, Chicago, 
1987. 
l'ollard, (2.I, .%g I.A.: Ih'ad-l)rivcn Phrase Sh'vchtre 
(;ra,mlar, I/niw'i'sil.y of (:hicago Iq'oss and (~Sl,I 
l>uhlicat.ions, in press. 
di ,q('iullo A.-M., }Viliialns I'\].: Oil lhr \[)efinitioli of 
Word, MIT l'ross, Caulbridge, MA, 1!)87. 
Sproal. R.: Morphology alid Coniputalion, MIT Press, 
(Janibridge, MA, ACI,-MIT Series in NLI', 1!)92. 
'l'rost 11.: The Appli(mtion of Two-Level Morphology to 
Non-(ik)ncatenal.ive (i~el'lilall Morphology, hi I(arl- 
gren II.(ed.), liroce(-dings of the 13th COLIN(\], 
lh'lsinkl, F'inland, 1q).'371-376, 1990. 
Trost II.: X2MORF: A Morphological (JoDlpoli011~ 
13asi,d Oll Augniei/l>ed 'l'wo-Lew~l Morphology, ill 
Proceedings of the 12th IJCAI, Morgal/Ka/ifilianil, 
,~ttll ~/llti,eo)C,i 1 1)1).1()24-1030, 1991. 
Trost \]l.: (k)iihlg with l)erivat.ion in a. Morphologi- 
cal Conlpl)lienl., ill 61,11 (Jollference of the Eilropeail 
Chal)l.er of the ACI,, Ul.recht, pp.368-376, 199,1. 
"147 
