A NOTE ON MOR2H~E STRUCTURE IN GENF~TIVE 
Frances Karttunen 
The University of Texas at Austin 
Austin, Texas U.S.A. 78712 
In an early model of generative phonology 
the lexicon of a language contained entries with 
as few feature specifications as possible in the 
interest of economy. The blank feature specifications 
representing both nondistinctive features and those 
rendered redundant by sequential constraints were 
filled in by the same 9honological rules. At this 
point, the concept of ~ rules changing feature 
values was unclear. 
When the distinction between rules that fill 
in blanks and those that change feature values 
became clear, it was zmbodied in the concept 
of morpheme structure rules and P rules. The 
MS rules were further split into feature 
redundancy (segment structure) rules and 
sequ~tial constraint rules. The MS component 
bore a striking resemblence to the earlier 
"pkonotactic" sections of autonomous phonemic 
analyses, but the claim was made for I~S ~les 
that they explained what phonotactiee merely 
described. The MS rules formed a major part of 
Chomsky's "readjustment component" which rendered 
th~ output of the syntactic component fit to be 
the input to the phonological component. A fairly 
current version of ~his model is the following 
one from Harms' Introduction t__oo Phonological 
c~) 
2. 
(I) Sequential constraint rules and blank filling 
rules fill in redundant features in lexical items. 
(2) These lexical items are inserted into the 
output of the transformational component. 
(3) Phonological rules, some utilizing syntactic 
information, operat~ on these strings. 
This model takes the lexical entry to be a 
prime of sorts. Within the lexical entry, seg- 
mental features are not specified if they can be 
predicted. If all obstruents in a language are 
voiceless, voice is specified for the obstruents 
in all lexical items by the redundancy rule 
~+obs~ ~ ~-vc~ . If the only possible initial 
consonant in a triple cluster is s, then only 
~+cons~ is given in the lexical entry. The 
rest of the distinctive features for s are filled 
in by a sequantial constraint r~le and the additional 
nondistinctive ones by a redundancy rule or rules. 
Harms' model orders all redundancy rules before 
the phonological rules. This ordering excludes 
Halle's solution for the exception of i and e 
from Finnish vowel harmony, which places the harmony 
rule before the redundancy rule specifging gravity 
for i and e. ("On the Bases of Phonology", 
The Structure of Language, 1964, p. 332) 
The following diagram represents Harms' model. 
3. 
Output of the 
Transformational 
Component 
I_ 
P rules 
xical Entries 
I sc rules 
redundancy rules 
Systematic i Phoneme 
Inv____entory , 
? ~ - 
n~ry feature values J 
PhOnetic Output 
4. 
Grouped together as morpheme structure 
rules (~S rules), sequautial constraint rules 
and redundancy rules, according to proponents 
of this model, account for the acceptability 
of some strings as possible morphemes in a 
language and the rejection of others. 
If we consider the problem of the acceptability 
of strings of a language's sequential phonemes 
(English has ~ and l, but ~lin is not possible.), 
then we might justifiably characterize mor- 
pheme structure rules as generating all and only 
the possible morphemes (stems, affixes, and 
uninflected particles) of the language. Is 
this equivalent to the function of the ~S 
rules in the model above? 
There seems to be the following conceptual 
difference. The model descrioed starts with the 
lexical entry, and the sequential constraint 
rules fill in ~istinctive features. The inventory 
of systematic phonemes is realized in the output 
of the SC rules. 
It is not clear where or how the inventory 
of a languages systematic phonemes fits into this 
model° The fully specified phonemes are realized 
in the output of the ~ rules. The lexical entries 
in some sense preexist the phoneme inventory, 
because they consist of incomplete feature matrices 
which are filled in by the NS rules. How does one 
5. 
arrive at the existing and possible lexical items? 
To effect a saving of features in the lexicon, 
all fea~res that can be determined from context 
are left out of an entry. Then MS rules fill them 
in. In this way, according to Harms, "Morpheme 
structure rules can account for the fact that 
native speakers of a language agree with great 
consistency on which of several nonoccurrin~ 
forms could be admitted as new morphemes in 
their language." (Intro. to Phonol. Th., p. 88) 
Is it possible that this phonological model 
and its ~S rules can account for all and only the 
acceptable morphemes in a language? If the ~S rules 
are only blank-filling rules, a completely new phoneme 
can be added at will to a language via this model 
by simply introducing it fully specified into a 
lexical entz'y before the ~S rules. 
If, on the other hand, the MS rules are able 
to reverse features in lexical entries, then ill- 
formed entries will be corrected. In this case, 
tae lexicon may be full of impossible morphemes, 
and the NS rules act as a filter to pass only 
well-formed morphemes on to the phonological 
rules. 
6. 
The argument against this is that the 
lexicon must be as econlmical as possible. No 
features, right or wrong, which are predictable 
by rule, are specified in the lexicon. 
Where do the lexical items come from then? 
And since the model under ~iscussionapplies ~u~les 
to items from the lexicon, not to s~rings An 
general, what is the status of the acceptable 
strings which do not, at a given moment, ~xist 
in the lexicon? A test for acceptability of a 
hypothetical string might be whether the ~S rules 
could fully specify its matrices without chang- 
ing any specified features. If blanks remained 
or features had to be changed, then the string 
would be judged unacceptable in the language. This 
test would require that not only existing lexical 
entries but also hypothetical strings of feature 
matrices be input to the ~ rules. To whatever de- 
gree the model rejects or changes faulty inout, 
it is clearly an acceptor or filter rather than 
a generator. 
Unless the lexicon-~S component relationship 
is completely circular (take out of the morphemes 
what can be put in by the ~S rules, list the remains 
in the lexicon, and then fill them in again by the 
~S rules), there is no way to account for the well- 
formedness of the input to the ~ rules. 
In his elegant article, "Redundancy rules in 
phonology", Language 43.2, 1967, Richard Stanley 
clearly demonstrated the different nature of re- 
dundancy and P rules (the former predicting feature 
values, the latter changing them) and the danger 
of misusing featu~re blanks. His proposal 
was that phonological redundancy be embodied not 
in rules but in Morpheme Structure Conditions. 
The former had only the lexicon as their domain 
and were ordered before the P rules. The latter 
have all matrices in their domain and are not 
ordered with respect to the P rules. 
To quote from Stanley: 
... A grammar of each natural language will 
have, in place of a set of I~S rules, an unordered 
finite set M of I~ conditions. This set will in- 
clude, in general, conditions of sash of the three 
types. The set of all matrices m in U, such that 
m is accepted by every ~ cmndition in M, is 
well defined; we call this set M(U). 
Since each NS cmndition in M represents a 
generalization about the morphemes of the language, 
it follows that the set M(U) represents all matrices 
which violate none of taese generalizations... 
Insahort, the set M(U) is exactly the set of pos- 
sible morphemes of the language. The segment structure 
conditions in M will guarantee that M(U) contains 
only those matrices ~ose columns are systematic 
phonemes of the language; the sequence structure 
conditions in M will guarantee that no seqummtial 
constraints of the language are violated in matrices 
of M(U). The set M of MS conditions may thus be 
thought of as filtering out, from the set U of 
~o 
all matrices, those matrices which do not form p 
possible morphemes of the language, leaving the 
set M(U). (Language 4~.2 p. 428) 
An alternative model of the phonological 
component, differing from Stanley§ in that it 
views the MS component as a generator rather 
than as a filter, assumes the language's fullY 
specified systematic phonemes as primitives. The 
MS conditions are then constraints upon concatenation 
of submatrices of these matricies of phonetic 
features. 
NS 
C omp. I 
~oneme inventory \ 
istinctive features specified) 
| -- 
~S-egment redundancy conditions I l 
(fill in nondistinctive features)~ 
IConstrailnts oft co--ncatenatio----~ ~ 
I 
Output of 
the trans- 
f o rma ti onal 
c omp. L 
All and only the possible 
morphemes of the language 
I lexic°n i 
'-P ru.les 
. 
The ~S component of this model generates all 
and only tile possible morphemes of the language. 
Of these, only a subset are existing lexical items 
associated with syntactic and semantic if~formation. 
Given a hypothetical string, its accept- 
ability can be deterT~ined on two criteria. First, 
is it composed of phonemes of the language? CheEk 
the inventory. If so, are the phonemes concatenated 
in a manner permitted by the Iris conditions? If 
these conditions are met, it may be a lexical 
item. Is it associated with semantic and syntactic 
inf oinrm tion? 
10. 
The practical use of morpheme structure 
conditions in computation brin@ to mind an old 
example of the commercial aspects of phonotactics, 
namely to generate brand names for detergents, 
beauty products, etc. more seriously, it is 
desirable to have a system report immediately 
a recognizable misprint or foreign word rather than 
to search fruitlessly through dictionary storage 
for the item. ~oreover, since dictionary storage 
must be continually updated, it is important 
for a system to report possible new lexical 
items for inclusion. 
