"DERIVATIONA12' PARADIGMS IN MORPHONOLOGY. 
Vito Pirrelli, Stefano Federici 
ILC-CNR, Par.O.La sas - Pisa (Italy) 
INTRODUCTION. 
Traditionally, paradigms were used to deal with 
inflection in inflectionally rich hlnguages. Only recently 
(Calder, 1989; Carstairs-McCarthy, I988, 1992) 
paradigms have been the object or a far-reaching 
investigaticm covering their formal and conlputational 
properties. This investigation has higbligthed the 
significance of a paradigm-based treatment of 
morphonological phenonlena and its theoretical 
implications. In this paper, we show how derivational 
processes in Morphology can be treated 
paradigmatically by using a mtn'llhonological 
network. The approach is not only theoretical 
speculation but has been subjected to the practical test of 
a computer implementation. This implementation leads, 
in our opinion, to a conceptually and conqJutationally 
eleane, treatment of Morphonology. 
1 TIlE PROBI~EM. 
There are two basic ways to relate a pair of 
nlorpllouological representations deriwitlonMly: 1) 
to take either o1' them as bslsic, arid derive tile other rronl 
it via a llrocess of rule-governed phonological change 
(Fig. I below); 2) to asstune a third representation 
(somewhat intermediate between the two) as underlying, 
and make tile other two derive I'rt)ln it either by 
substitution of some phonological segments, or by 
filling in an underspecified phonological representation 
(Fig.2 below). 
b) pro 'fawnd 
c/ 1 
pro'f ^ nditi 
Fig. I base-derivative relation in Morphonology 
Hereafter, we will refer to both I) arm 2) as 
derivational accounts of morphonological relations. 
Usually, sohition 1 stores the basic form only in the 
lexicon, while tile derivative is, sis it were, cranked out 
on demand by rule application I. On the other hand, 
sohttion 2 requires that only one underlying, abstract 
representation be stored in the lexicon: the two related 
tbrms are yielded from their common lexical source by 
rules. If b and c sire the nlorphonological representations 
to relate, the two approaches can be illustrated as shown 
in Figg. 1 arid 2. 
hi Fig. I, b arid c are taken tn be on a different 
footing: the form al the top is more basic than tile one 
tit the bottom, the arrow indicating a derivational 
relation. The IMIowing assumption is made: given two 
forms to relate, it is always possible to specify the 
direction of the arrow. 
a) 
pro'f~e:nd 
b) ~ c) 
pra 'fawnd pre'f ^ nditi 
l:ig. 2 abstract representations slnd morphonological 
relations 
This assumption stumbles upon a nmnber of 
difficulties. Let us consider a small but crucial portion 
of English segmental Phonology as classically analysed 
since Clmmsky ,'lnd llalle (1968). First, schwa is well 
known to alternate with full vowels within the 
Pllonology of English. l)erivational series such sis in 
I:ig.3 below show tllat full vowels under stress 
corresp(md to schwas in unstressed llOSilions. 
\['fowtagrgef\] / \[fowtffgra3fik\] / \[fo'ta grofi\] 
Fig. 3 derivational alternations 
This alternation is captured by a rule of vowel 
reduction: full vowels are reduced to schwa in unstressed 
position (e.g., I',el -> schwa in photography). Another 
systematic l~honological alternation in English involves 
derivational pairs such as sane-xmfity (l'sejnl/l's,'eniti\]), 
which strongly suggest a derivational rule like ej -> m. 
Ilowever, other alternations appear to go in Ihe opposite 
direction. For example in a triple like morginsl 
1Alternatively, both representations can be stored in 
{he Lexicon, and related through tile slatemcnt of some 
rcchmdancy Icxical rule (lstckendoff, 1975). 
234 
hnar£inalily/marginGlia, the following alternation in 
the underlined vowel emerges: schwa/~e/ej. If we posit 
/~e/as underlying, then we end up having to set up an 
lie/ -> /ej/ change, which is tile mirro,+-image of the 
/ej/->ke/ relation posited for tim derivational l)air 
\['sejn\]/l'smniti\] (more on this ill l)urand, 1990). 
The sohttion illustrated in Fig.2 above is a way to 
solve this apparent paradox. A third abstract segment 
/,'e:/ is assulned to be basic rehltive to /schwa/``e/ejt. 
This means that/schwa/m/ej/can be derived from/m:/ 
through application of some phonohigieal rules. A 
derivational chain of this sort can be ralher conlplex, 
since I',e:/and lej/are, phonokigically, far removed from 
each other, and ninny intermediate changes can he needed 
(as ill I'me:nl -> I'se:nl -> I'se:inl -> I'sejnl; sue tlalle 
and Mohanan, (1985) for more examples). More 
reservations on chains of lifts sort have been expressed 
recently in a series of psycholinguistic experintents, 
aimed ill probing the reality of tile derivational 
asstunptions (Jaeger, 1986; Wang and Derwing, 1986). 
ht the literature, an allogether different ai)proach 
froiI1 both derivational accounls I) and 2) al)ove has been 
suggested (Vennemann, 1974): b and e should simply be 
lisled in the Lexicon, one beside the other, On :l par: 
pro 'fawnd -..91 I-,.- pr o'f ^ nditi 
Fig. 4 mo,phonological relations and lexical listing 
Formulated in this wily, tile relation in l:igA is 
always "true on the surface", shlce there is no abstract 
representation involved ill this account. Moreover, no 
rule ordering ploblems arise (i.e., concerning tilt 
direction of the arrow). If the sohltion in Fig.4 is 
adopted, however, it is not at all clear what type of 
lexical architecture one is suggesting: i.e., it is moot 
whether b and c are sontehow related in Ihe lexicon, or 
they are simply listed together. More worringly, tilt 
notion "true on tile surface" is of little theoretical help 
for explaining sotne well-known eases of alternalion. 
Take, for example, the opposition between tile American 
English pronunciation of writer and rider, respectively 
h'ajVerl and/ra:jV'er/, where hoth /t/ and /d/ have been 
turned into aflap (\[l'\]). If lhe only difference between 
the two phonological realizations is the surface-true 
lengthening of ~ill in /ra:jVel'/, then one is nlissing the 
relation between vowel length and tile verbal b,'tses of 
the two derivatives, a relation liutt represents an 
important, productive generalization within the 
American English phonological system (l)urand, 1990). 
An American English speaker, when asked to derive a 
new agentive in -er from verbs ending in /t/, will 
produce flapping with no lengthening. In contrast, (s)he 
will produce both a lengthened vowel and a flap if tile 
base is a verb ending in /d/. Derivational acct:,tmts 
capture this generalization in an elegant way, through 
rule ordering: vowel lengthening takes place before 
flapping does, in the context "vowel followed by a 
voiced consonant", so that when either /t/ or /d/ 
disappears, lengthening has idready applied (OF failed to). 
ht contrast, accounts based on lexical lisling al+e cleaHy 
incapable of grasping this significant hidden relation, 
since the opposition between/I/and/d/is no longer true 
on the surface in pairs such as /rail 'er/ vs /ra:il 'er/, 
where the opposition is neutralized by a flalx 
To sum up, derivational accotlnts of 
nmrphonological alternations \[','Ice a ntunber of 
theoretical and practical pl'olflenm. Ill what follows, we 
will illustrate the working of a set of analogy-based 
principles and tile design of a general parallel architecture 
for their implementation. These principles dispense with 
both I'Ule ordering and Stlrfilce tllltl'tle representations hy 
using l)aradigms of alternations instead of derivational 
chains. The intplemented architecture proves to be 
accurate and COml)utationally eflicienl. 
2 TIlE GI,:NI,:RAI, MOI)i~L. 
In this section tile idea is illustrated that redundancies 
among linguistic data can he used to convey interesting 
linguistic generalizations if (.lain lll'e slored ill an 
incremental network, l.et us consider the general case 
first, exemplified by the following list of abstract cases 
and their categorial classification: 
137 x 
037 x 
33"/ x 
SUl)lmSe that tile sequence of nunterals 137 
represents tile Iorm of a linguistic ot:iect (say a word); x 
is its category. We can represent these dala by exploiting 
tile redundancies that they show at both tile lbrnml and 
categorial level of description as follows: 
3 
<=> X 
Fig. 5 a formal core and its ealegory 
The idea is to inlerconnect all forms chtssified the 
same w'ty by lelting Ihent share a cmnmon I'ornml 
core if there is any (for a rigorous definition of tile 
notion of fornml core, see Pirrelli, 1993). In Fig.5, -37 
is Ihe linmal core. A core with its category (x in the 
example at h:md) is called a nucleus. The graph it/ 
Fig.5 can be seen as a nelwork of nodes. N,::,des which 
are linked through a solid line have been wilnessed ill 
input as to.occurring in tilt same form. Nodes which are 
not linked do nol co-oeclll'. MtHeover, if two nodes 
allernale, that is if they are it/ complententary 
dislrilmlion with respect to a shared core, they occupy 
the same coltllnll ill tire figure, ht other words, nodes on 
tile same column in Fig.5 arc mutually exclusive, 
paradigm.'dically relaled alternations. 
l+et tlS considcF now ht)w lhis network e;.In be used to 
associate forms with categories. Asstmle that more data 
have heen stored in the network so that the configuration 
shown in Fig.6 overleaf is built up. In Fig.6 o is 
another possible cutegory (different from x), and 38 <=> 
o aYiotl)er nucleus. Given a network like this, network 
cores are activated hy an input string if they are 
contained by it. The nelwork output will then be the 
calcgory associated with the actiwltcd core. 
235 
236 
3 
<=> X 
4 
<=> 0 
Fig. 6 riwfl cores in a multicategory network 
More concretely, if a certain lbrm, say 437, ix given 
as input to the network, the system tries to guess the 
right category on the basis of the analogy of 437 (called 
the target string) with already stored items (called 
base strings). This is carried out through the 
lbllowing steps. Network cores are activalcd by 437: 
this is done through a simple string-matching routine 2. 
For example, 37 will be actiwtted, and its category x ix a 
candidate response of the system for the input 437. What 
happens when more than one core gets activated by the 
same input token? If the activated cores have the same 
category, the corresponding mtdtiple responses would 
reinforce each other: the only category activated is given 
as OUtl~Ut. The case of multiple responses with different 
categories (and thus potentially different outputs) is 
more complex. We can dislinguish two cases: 
I) there is no core which is fully activated 
2) there ix at least one fully activated core. 
A core is fully activated when it is entirely contained 
by the input string: in the example above 437 fully 
actiwttes the core -37. This contrasts with partial 
activation, when the core is only partially contained by 
the input string: for example, 437 partMly activates the 
core -38 of Fig.6 above, since it contains 3 in the 
second position, but not 8 in the third. Let ns consider 
case 1) above first, l\[' there is no core which is fully 
activated, then the system goes through the following 
two steps: la) for each candidate response, the system 
gauges an Activation Ratio: 
number of numerals (+1 the 
activated core contained in input 
total number of numerals in the 
actNated core 
The actiwttion ratio (AcR) is then 0 < AcR _< I. The 
case of full actiwttion (AcR= 1) is dealt with in 2); lb) 
the core with the highest AcR wins out over the others. 
Let us consider now case 2). If there exists at least 
one fully activated core, then two I'urfl~er subcases need 
be distinguished: 2a) there is only one core which is 
fully activated: the category borne out by that core is 
2 More on string-matching will be said in section 4 of 
this papcr. 
then picked up; 2b) there are at least two cores which 
are fully activated. 2a) can be seen as a degenerate case of 
1) above, when AcR = I. As to 2b), in this case the 
activation ratio is clearly no longer conclusive. One 
needs to gauge :t further ratio, called Analogy Ratio, 
whose definition follows: 
total number of numerals in the 
activated core 
total number of numerals 
in input 
&gain, the analogy ratio (AnR) is 0 < AnR _< 13. 
To stnn up, a certain pattern is analogic,'dly activated 
if the following conditions are met in tiffs order: 
I) it is activated by an input pair 
2) it has got the highest activation ratio 
3) it has got the highest analogy ratio 
So far, we have been assenting that the network 
produces cores incrementally as more data .'n'o input. In 
fact, the nclwork pfogressively exlracts retlnced cores. 
Each full form is aheady a core in its own right, 
otherwise it would never get activated. The exmtction of 
smaller cores represents the process of acquiring 
generalizations on the basis of the analogy between data 
(redundancies). In the network, an exception is simply 
an isolated case, that is a nucleus which is fully 
actiwtted only when it is witnessed in input in its 
enlircty (AnR = I). 
3 AN ANAI~OGY-BASH) NETWORK 
FOIl. Titl,\] SYNTIIES1S OF MORPIION- 
OLOGICAL ALTERNATIONS. 
Ilow is it possible to model the synthesis of the set 
of systematic alternations illustrated in section I by 
means of a network such as the one in Fig.6? Fig.7-a) 
overleaf illustrates the result of storing four 
phonological representations, namely \['vejn\] (vane), 
\['sejn\] (sane), \['vzeniti\] (vanity)and \[i'nieniti\[ (inanity), 
phls an (abstracl) catcgorlal feature string (instead 
of an atom) for each of them: respectively xo, yo, .~i and 
zj. tiacb catcgorial feature string is supposed to contain 
lexico-semanlie information (c+g., a capsule 
representation of the ineanitlg of the relevant Icmn-a 0 and 
morfJhosyntactie inR+rmation (such as grammatical 
category, gender, number, tense etc.): the content of thc 
feature string, however, will be relatively neglected here. 
The machinery of core reduction outlined above obtains 
Ibr categorial feature strings as well as for their forms. 
\[,~nitil and lejnl are forrnal cores; x, i and o arc 
categorial cores. 
In l:ig.7-b) the arrows between \[s-\] and \[-~eniti\], and 
y and j, pictures an instance of paradigm extension. 
3We assume that AnR is caletdated only for those cores 
which have been fulIy actiwtled (AcR=I). Thus, 
activated cores cannot contain more numerals than the 
input does. 
Paradigm extension is based on an int(iitive idea: if mane 
shares with vane one nucleus, then it is expected to 
shm'e all other paradigmatically-relatcd nuclei. 
tO \[' <~> Z 
Is\] ., 
b) \[in\]_ <=> Z 
Fig. 7 illorphor/ollagical alicrnalions and paradignl 
cxtcllsi()ll 
This may not always he true, but is certainly a 
governing principle irl tile theory or l)artl(tign/s. For 
example, when faced with an unknown o. cnding 
adjectival \['orlll sl.ich lls obsoleto (F, nglish 'obsolete'), ,till 
Italian speaker would also predict that obsolera is 
fenlinino singuhu, obsoleti illasctllitlc iflural, oh.volete 
|'enliniilo llhii'al, according to the paradigm of \['otir-way 
adjectives in Italian (Mallhcws, 1992). I/y Ihe silillC 
token, tile extension in ltig.'/-b) above represents the 
expectation that the nominalization of sane is \['s:cnitil, 
by paradigmatic analogy wilh \['va'niti J. 
(\]onlputalionally, this c×tonsioii is hnplonlonled as a 
COlitiiltiotis path of hilcr-n()dc c(lllilcclions (say \[r()i/i isl 
to \[~nitil, through \[ojnl and Iv\] in lqg.7--b). 
\[I\]~> <=~ 
it) 
\[r <.~> Z 
\[g\]~=> 
Fig. 8 rival paradigms in a n/orl)hollotol, ical 
notwork 
I;laplfin,~ ~ , and Icngfllcning, in the Ame.rican English 
pronunciation of writer aild rider, can be represented in 
analogical terms hy the two diagrams in l:ig.8. There, 
we have represented the immunciation of writer and rider 
in two separate analogical patterns, respectively i) 
(without lengthening), and it) (with lengthening). In 
fact, i) and it) should be merged into the same larger 
lXittcrn, llowevcr, it quickly becomes impossible to 
lficture multidinlensional links on a page. l:ig.8-i is the 
result of the exposure of the network to the following 
fotu word forms: lrajtl (write), lrajl 'crl (writer), I fajl~cl'l 
(\]'i@ter) and Ihtitl (light). l~ach word form is given with 
its catcgorial I'¢tittlre string: xk I'or \[r@\], xo for trail 'erl, 
yo for \[fajl'crl and wk for \[htjt I. Fig.8-ii is what the 
system yields when it conics across the folk~wing three 
word forms: I ga:jI-erl (guider), Iga:jdl (guide)and Ira:jd\] 
(ride). Again, forms arc given with their catel;orial 
string: jo for \[ga:j\[ "erl, jk for Iga:jdl and zk for Ira:jdl. 
The reader will note Ihat formal cores tire parasitic 
on phonological redundancies, sitlce c(:,l'eS tire 
extracted on the basis of systematic stirJ'acc-lrtlc 
analogies between morl~honological represenlati~ms. Tim 
difference between supplciion and (semi)regular 
altcrtmtions is easily captured: (semi)regular alternations 
exploit c(ncs more systematically than do SUlDlctivc 
alternations. IVlorcovcr, the structure of the analogical 
network makes it possible to express hidden 
phonological constraints as paradigms nf phonological 
alternations. This allows the system to avoid the 
questionable use of surface tllltruo., underlying 
ldlonological representations, typical of dcriwuional 
:iccotlnts (as in Fig.2 above). 
To Illzlke the lilsl Iwo points cIcarel', let tls turn Io a 
concrete instance of word synthesis by analogy. A 
typical objection levelled at representations which hcat 
alternations by listing Ihctll, is that crude lists of stored 
JtClllS (IO not lilakc it clclii" distinction boikVCCll rcgi.llar 
altcrnati(lns illl(I irregular OlleS. (liven ti raw list of cases, 
it is olyjoclcd, nolhing cnn he predicted fiom such a list, 
in llltich Ill(; SalilO way as nothing Clill I)o prodicted troll1 
- say .- the supt)lelivc allcrnation fu)lwctll. 
\[I\]~> <~~ 
it) z 
< 
Fig. 9 wc, rd synthesis through mc, rF, honoh:,gical 
paradiDns 
237 
However, paradigrnatic extension accounts for tile 
fact that flapping is a productive alternation in American 
English. 
Let us suppose that the system has to produce the 
agentive nominalization (lighter) of \['lajtl , whose 
surface form and lexicaI content bas been already stored 
in the network (Fig.9-i above). A flap (will\] no 
lengthening) would then automatically appear ill the 
place of surface \[-t\], according to tile following steps: a) 
the input conditions are represented by the lexical 
content o1' the verb light (w in Fig.9-i), and the 
categorial information "agentive nominalization" (o in 
Fig.9); b) w actiw~tes the \[1\] of \[lajtl ill the paradignl i) 
of Fig.9 above; e) tim activated core o triggers the flap 
alternation \[-aj F'er\] of word-final \[-ajt\] in Fig.9-i, and the 
alte,'nation \[a:jF'er\] of word final \[-a:jd\] in Fig.9-ii; d) 
the form \['lajFer\] is thus produced, since there is a palh 
of pradigmatic links between \[ll .'rod \[aj Ver\], while there 
being no such a path linking \[11 and the alternation 
\[a:j Fed by paradigm extension. 
It should be noted that only (semi)regular 
alternations share some (sub)core(s) in common. 
Suppletions such as go/went simply do not. 
Nevertheless, no clear-cut distinction is drawn here 
between minor alternations and irregular forms, since all 
of them m'e simply stored in an analogy-based network. 
Their difference is accounted for in terms of a gradation, 
defined by the fact that regular ahernations overlap with 
other lorms more extensively titan do snppletions. We 
believe that this solution is empirically superior to tile 
Anaximander's principle invoked by \[ludson (1974), to 
the effect that, since there exists no clear-cut distinction 
between suppletions and systematic alternations, all of 
them should be listed in the lexicon. 
4 LIMITS OF TIlE MOI)EL AND 
FURTIIER IMPROVEMENTS. 
The network has been implemented in C, and 
proved to be 95% accurate in analysis after a training on 
20,000 Italian word forms, and 75% accurate ill 
synthesi s, as reported in Pirrelli and Federici (1991). 
These performances have been obtained hy using tile 
simplified model illusm~ted here. 
It should he noted, however, that this model works 
well for the Morphology of Ihose hmguages (such as 
Italian and English) where affixation is commonly 
realized through a concatenative operation (surfixation or 
prel'ixation). In these cases, string-matching is a fairly 
simple head-and-tail operation. Clearly, this model is far 
from having a universal, cross-linguistic wdidity 
though. For example, it does not work well with cases 
of circumfixation (known also as parasynthesis), let 
alone tile Morphology of nonconcatenative l'mguages. 
Nevertheless, we contend that the set of analogy-based 
principles illustrated here hold Ibr a wider spectrum or 
hmguages than purely concatenative ones: by making 
string-nmtehing a more flexible and powerful operation, 
we can successfully adapt our model to the requirenaents 
of noncatenative languages such as Arabic, or Io the 
treatment of discontinuous affixation. What rollows 
sketchily illustrates this line of develoI-m/ent. 
(liven two strings of characters/phonemes to match 
(called tbe base and tile target string), the new string- 
matching algorithm (discontinuous string- 
matching) we are currently expe,'inaenting on is, 
informally: 
smrt fi'om the h'ft corner of both the target and base 
string and scan them rightwards; extract, down the way, 
all characterdphonemes which appear in both, and in 
the same order. 
Take tile Italian word forms \[inve'k:jarel 
(inveechiare, English 'to age' a parasynthelic verbal 
derivative of the adjective vecchio, 'old '4) and 
\[intimi'direl (intimidire, English 'to make shy, 
intinaid:ltc', a parasynthetic derivative of timido, 'timid, 
shy'), or the Arabic pair \[knlaybl ('little dog') and \[kalbl 
('dog') ill the diagrams ill Fig. 10. Ill the Alabic exmnple, 
tile shared portion of meaning (DOG) triggers tile 
extraction of a tentative core (a stem). By using a 
discontinuons string-matching, Ihe meaning I)OG is 
associaled with tile match k_lb. 'File remaining portions 
of mOlllholexical reaturcs of tile Iwo wond forms are 
respectively associated with tile substrings left out of the 
nlatcb. Note that tile two a's which are linked by a dotted 
line will no! be extracted ill Ille same pallern as k_lb, 
since they are ordered differently relative to the position 
of l (in tile top string, a \[bllows l, while in the bottom 
string, a precedes it). In the Italian example, the shared 
nmrl~hological feature i,~finitive triggers the extraction 
ofthecireumfix\[in rcl. 
inve'k:jare kulayb !/ n ,,r/ 
intimi'dir kalb 
Fig. I0 tile use of discontinuous string-matching in 
Italian and Arabic. 
l)iscontinuous string-matching captures an important 
range of phenonmna which would be treated clumsily by 
using a simple head-and-tail matching. A fnrtber 
advant:tgo of tiffs routine over even more llnconstrahled 
conceivable routines is thai it rcdtlces considerably tile 
number of c.tmdidatc nmlches. Oil tIle nlorc negative side, 
wc are not able to find matches whose order in the base 
siring is interchanged relative to the order in Ihe target, 
as in English un-re-do and re-un-pack in lVig. ll-a) 
overleaf. This by no means itnplics that plmnonmna 
such its so-called inversion or metathesis, as ill the 
Rotuman example of Fig. I I-b) overleaf, are beyond lhe 
reach oF oHr approach. The idea is that a matcl~ is 
extracted only when ,'1 base and a target string show the 
same order of interchangeable characters, as in lgg. I I-c) 
overleaf. Tim productive morphological relation 
between, say, /.,a and ap in the word 'pigeon' in 
l,~.otuman is acconnted for paradigmatically, as 
explained ill section 3 of this paper. Intuitively, we do 
not say that a sequence pa is transformed into ap nnder 
4 We depart, here, from Scalise's analysis of invecchiare 
as tile product of tile sequential application of a suffix I- 
arel and a prefix lin-\[ in this order (Scalise, 19B3). 
238 
particular circumstances. Rather, the sequence pa 
alternates with the sequence ap within a cerlahl 
paradigm. 
a) n r i d u: Z 
ri Anpm k 
b) c) 
\]\[P~ ('a pige°n') i611 ('a pigeon') 
?iap {'the pigeon') ?(pa ('amat') 
Fig. 1 1 discontinuous string-matching does not allow 
tk)l" ClOSS-lllaletaes. 
5 CONCIAJSIONS. 
The architecture illustrated here can be looked at as a 
model of the morphonok~gical competence of a native 
speaker, and functions both in analysis and generation. 
At this point, a question naturally arises: is it a lexicon 
or a gran/nlar? In our opinion, it is both: a self- 
modelling lexicon which extracts generalizations on tile 
basis of tile anak)gies between :lheady slored items, and 
uses these generalizations as repair strategies for hick of 
direct evidence provided hy an aheady stored item. This 
interpenetration of lexicon aild graillnl,:lr ofl',ais a \['cw 
advantages i,i dealing wilh tile problenl lit hand: 
- phonological concreteness of lexieal 
representations: known lexieal items are given a full 
surface phonological representation in the lexical 
network: no abstract phonological segments such as \[m:\[ 
in the phonological representation of prqfin.ld i. Fig.2 
are needed ; 
- alternations vs deletions a,ld inversions: 
changes ill the l)honologic:ll structure of at given entry 
are always expressed :is alterlultions, never as delelions; 
Ihis has the obvious implicalion that all alternating 
segments are encoded in the network and can be retrieved 
at any time: e.g., one does not say that flapping makes a 
dental disappear, bUl that IlaPlfing can appear only in 
those paradigmatic contexts where a dental is also 
present; this limits the computational power of the 
required rule set considerably, since deletive rules are 
eolnputationally most costly; moreover, by Ireating 
eases of metathesis paradigmatically, the system further 
spares tim computational price of' a tlanslornmtiona\[ 
operation such as inversion. 
-unordered rules: the context Ik~r a given 
phonological change to take place can be specified in 
terms of It whole paradigm; this formulation has lira 
irrunediate advantage of awfiding the need for rule- 
ordering: Ibr example, the fact that writer in the 
American-lblglish prontmciation has tat)lenglhened lajl 
is traditionally accounted for by ordering vowel 
lengthenirig before flapping; this move is no longer 
needed as long as one can restrict the applicatiou of 
vowel lengthening to tile paradignl of those nuclei 
which present \[a:jl ill the base Ibm) of a verb, as we 
showed above; 
- phonological concreteness of extracled 
lteneralizations: phonological generalizations are 
expressed ira terms of extracted nuclei; nuclei represent 
what is shared by two (or more) sit)red items; it Ibllows 
Ihat no phonological features appenr in extracted nuclei 
other Ihan those which occtn' in the surface 
representations from which nuclei are derived; 
-use cfl' direct evidence: as we saw, snlaller 
nuclei (morphemes) c()ia/e into play as repair strategies 
only; Ihat is, if and only if there exists no whole lcxical 
reprcsenlaticm (an already known word) which matches 
(rather Ihe inl>Ut (in analysis) or the output (in 
generation) in its entirety; otherwise, bigger nuclei 
always override smaller ()ties; this accounts for 
phenomena of Icxical blocking (Aronoff, 1976), when - 
say - the word business is never interpreled ns the 
'quality of being busy'. 
\[:rom a more dynau/ic perspective tile network 
illustrated above can be seen \[is till abstract model of the 
gradual learning of nlorphonoh)gical phemmmna by a 
speaker. In fact, the principles illustrated ahove wele 
originally developed ,'is principles inforndng machine 
language letlrning as such (Fcderici, 1990), and tested in 
dealing with some complex cases el tmsupervised 
acquisition o\[" particular linguislic capabilities (e.g., in 
\[asks o1' biqingual translation). Research carried ou\[ ill 
this paper proves that I/lorpholloh)gical phenomena llle 
within tile grasp of these learning principles. This 
represents an imporhnit wposteriori confirmation of file 
validity of the integration between models of machine 
learning and computational models of linguistic 
competence, an integration that has already shown its 
merits in cognitive approaches to language learning and 
linguistic theorizing (Pinker, 1989; Carstairs-Mc('arthy 
and Sicmherger, 198g). 
R E F' E R E N C I,'. S. 
Aronoff, M. (1976). Word Formation ill Ge,erative 
()l'(llllDl(ll; MYI' Press. 
Calder, J. (1989). Paradignuuic Morphology. in 
Proceedings of AC\[., Manchester. 
Carstairs-McCarthy, A. (19gg). Non-concatenalive 
inflection and paradigm economy. Ill lkunmond M. and 
Noonan M. lids, 77teoretical Morphology: Approachex 
in Modern Linguistics, pp. 71-77, San Diego: 
Acntlemic Press. 
Carstalrs-McCarthy, A. (1992). Current 
Morphology, l,ondon: Routledgc. 
Carstairs-McCarlhy, A. and J. P. Siembcrger (19gg). 
A processing constraint on inflectional homonymy. 
l,inguistics 26, pp. 601-17. 
Chomsky, N. alld M. Ilalle (196g). 7"he Sound 
Pattern o.fE, glish., New York: Harper and Row. 
l)uratld, J. (199(I). Generative and Non-Linear 
Phonology, I~ongman. 
239 
Federici, S. (1990). Un Sistema Connessionista 
Auto-Espandibile di Comprensione del Ling,aggio 
Naturale, Laurea Dissertation, Universita' di Piszl, l)pt. 
of Computer Sciences. 
Halle, M. and K. P. Mohanan (1985). The 
segmental phonology of Modern English. Lingtdstie 
lnquily 16, pp.57-116. 
Hudson, G. (1974). The representation of non- 
productiove alternation. In J. M. Anderson and C. Jones 
Eds. Historical Phonology, 2, Amsterdam: North- 
Holland. 
Jackendoff, R. (1975). Morphological and Sere;ratio 
Regularities in the Lexicon. Language, 51, pp. 639-7I. 
Jaeger, J. J. (1986). On the acquisition of abslract 
representations for English vowels. Phonology 
Yearbook, 3, pp. 71-97. 
Matthews P. H. (1992). Morphology (second 
edition) CUP. 
Pinker, S. (1989). Learnability and Cogtlition, MIT 
Press. 
Pirrelli, V. (1993). Morphology. Analogy and 
Machine 75"anslation., Phd Dissertation, Szdford 
University UK. 
Pirrclli, V. and S. Federici (1991). An an.'dogical 
way to language modelling: MORPHEUS. A CTA 
LINGUISTICA tlUNGARICA 41. 
Scalise, S. (1983). Motfologia Lessicale, CLESI": 
Padova. 
Vennemann, T. (1974). Words and syllables in 
natural generative glztnlmar. In A. Bruck et al. Eds., 
Papers from the parasession on Natural Phonology, 
(Chicago,CLS), pp. 346-74. 
Wang, H. S. and B. Derwing (1986). More on 
English vowel shift: the back vowel qncstion. 
Phonology Yearbook 3, pp. 99-116. 
240 
Lexicon 

