LinguisticInquiryVolume4 Numberi (Winter, I973) 3-I6.

MorrisHalle Prolegomena to a Theory

of Word Formation*

Speakers of a language normally possess knowledge not only about the words of the
language but also about the composition and structure of the words. Thus, for in-
stance, speakers of English know that the words in (i a) are words in their language,
whereas those in (ib) are not.
(i) a. dog think write love antidisestablishmentarianism
b. svan pensare katav mile Donaudampfschiffahrtsgesellschaft
They also know that an adjective such as (2a)
(2) a. trans-form-at-ion-al
is composed of the morphemes shown and that these five morphemes cannot be con-
catenated in most other orders; i.e. sequences such as (2b) are not possible in English.
(2) b. ion-trans-al-at-fbrm
al-form-at-ion-trans
Analogous facts can be readily adduced from any other language.
A grammar is the formal representation of what a speaker must know about his
language, and it must, therefore, reflect in some fashion facts like those above, among
others. Hence the assumption has been made quite generally that a grammar must
include a list of morphemes as well as rules of word formation or morphology. The
character of these rules and their relationship to other parts of the grammar, in
particular, to the rules of syntax and of phonology, has been studied only to a very
limited extent. It is these questions that occupy my attention in what follows, and
while I am not in a position to claim that I have succeeded in achieving a break-
through in this area, I hope that I have developed enough of a structure to facilitate
discussion and to attract others into research on this topic.
It is all but self-evident that in the list of morphemes the different items cannot be
represented just as sequences of (phonetic) segments, but that they must be provided
* This study is based on my report "Morphology in a Generative Grammar", presented to the Eleventh
International Congress of Linguists held in Bologna, August 28-September I, I972. The report as delivered will
appear in the Proceedings of the Congress. I am grateful to T. G. Bever, C. Blanche-Benveniste, S. Bromberger,
and E. W. Browne III for their helpful comments and advice. This work was supported in part by grants from
the National Institutes of Health (5 Toi HDoo iii) and the National Institute of Mental Health (5 POI
MH13390).
4 MORRIS HALLE

also with some grammatical information. For example, the entry for the English
morpheme write must contain the information that it is a verbal root, that it is a
member of the "non-Latinate" portion of the list (it is by virtue of this fact that it is
allowed by the rules of word formation to combine with certain affixes and not with
others), that it is among the small class of verb stems that undergo the so-called
"strong" conjugation, etc. Moreover, the list must include not only verbal, nominal,
and adjectival roots but also affixes of various sorts.
The rules of word formation would then tell us among other things how the
morphemes are to be arranged in sequences to form actual words (allowing (2a) above
and ruling out (2b)), and it is to be expected that among them there might be rules of
considerable complexity. The nature of this type of rule will be investigated in greater
detail at a later point in this paper. At this point I am mainly concerned with the fact
that not all properties of words can readily be accounted for with such simple rules.
Particular difficulties arise in connection with the treatment of idiosyncratic character-
istics of individual words, that is, of characteristics that a given word shares with few
other words or even with none. Here are some examples illustrating the idiosyncratic
behavior of words I have in mind.

(3) a. approval recital proposal transmittal reversal

b. arrival refusal rehearsal acquittal renewal carousal betrayal
withdrawal denial survival betrothal
An examination of the individual meanings of the nouns listed in (3) shows that
in spite of many common features, there are also a great many differences among them.
Thus, while most of the nouns in (3) have the meaning 'the act of V-ing' and (if
transitive) also 'the fact of being V-ed', this formula does not hold in all cases. For
example, recitalcommonly refers to a concert by a soloist, or transmittalis used to refer to
the transfer of official documents or information, but not of other things that are
quite normally said to be transmitted.
Idiosyncratic behavior of individual words can be found also in other domains,
e.g. phonology. As an example, consider the English nouns formed from adjectives by
adding the suffix -ity:

(4) serene [sQriyn] serenity [sar&nitiy]; obscene obscenity; sincere

sincerity; severe severity; profane profanity; divine divinity
As can be seen from (4), nouns of this type are normally subject to the Trisyllabic
Shortening Rule (cf. Chomsky and Halle I968, I8i). The nouns entirety,nicety,obesity,
and probity differ from all other nouns formed with this suffix in not being subject to
the Trisyllabic Shortening Rule.
A third type of idiosyncrasy of words can be seen by comparing the deverbal
nominals formed with -al with those formed by adding the suffix -at-ion or -ion as
transformationand decision.While the restrictions on nominals in -al are quite different
 PROLEGOMENA TO A THEORY OF WORD FORMATION

from those in -at-ion, there is a number of verbal stems from which both types of
nominal can be derived; for example, the nominals listed in (5a) are formed from the
same stems as those in (3a).
(5) a. approbation recitation proposition transmission reversion
On the other hand, the nominals in (5b) do not have counterpartsin -al:
(5) b. derivation description conversion confusion permission observation
obligation omission accusation
Moreover, the nouns in (3b) lack counterpartsin -at-ion;i.e. the language lacks words
such as those in (6).
(6) a. *derival *describal *conversal *confusal *permittal *observal
*accusal (but cf. Jespersen, MEG 6.22.22)
b. *arrivation *refusation *refusion *rehearsion *acquitation
Once again it appears somewhat forced to incorporate this information in the mor-
pheme list or in the word formation rules. But if it is not to be incorporatedthere, then
one must immediately ask how this information is to be reflected in a grammar.
I propose that idiosyncrasiesof the type just illustrated be listed in a special filter
through which the words have to pass after they have been generated by the word
formation rules. The special information given in the filter under each entry is then
added to the representationof the word. In the case of semantic idiosyncrasiessuch as
those exemplified by the special meaning of nouns like recitaland transmittal the filter
would supply the appropriateindications about their semantics. In the case of phono-
logical idiosyncrasieslike those exhibited by nouns like obesity,the filter would supply
the information that the noun in question is not subject to the Trisyllabic Shortening
Rule, or, more formally,would supply the noun with the feature [- TrisyllabicShorten-
ing Rule]. Finally, "gaps" in the dictionary like those illustratedin (6) and (7) would
be accounted for by providing the "missing" words with the rule feature [- Lexical
Insertion]. In other words, the fact that English lacks the nouns *derivaland *arrivation
would be reflectedin the grammarby markingthese words, which would be generated
by the word formation rules, as not being subject to lexical insertion and therefore
incapable of appearing in any actual sentence of the language, in spite of the fact that
they are neither semantically nor syntactically or phonologically anomalous.'
1 The proposal just sketched might be modified somewhat as regards the treatment of words formed by
rules that traditionally have been called "nonproductive". One might propose that all words formed by non-
productive rules are marked by these rules as [-Lexical Insertion]. The smaller subset of actually occurring
words formed by such rules would then be listed in the filter with the feature [+Lexical Insertion]. That is,
the nouns formed with the suffix -al would all be generated with the feature [-Lexical Insertion]; the relatively
small number of actually occurring nouns of this type, like those listed in (3a), will appear in the filter marked
[ + Lexical Insertion]. In other words, it is assumed that words generated by a productive process are all actually
occurring and that only exceptionally may a word of this type be ruled out of the language. On the other hand,
words generated by a nonproductive rule are assumed not to be occurring except under special circumstances.
In this fashion we might capture the difference between productive and nonproductive formations.
6 MORRIS HALLE

In other words, I am proposing that the list of morphemes together with the rules
of word formation define the set of potential words of the language. It is the filter and
the information that is contained therein which turn this larger set into the smaller
subset of actual words. This set of actually occurring words will be called the dictionary
of the language.
The examples discussed above have been chosen from the domain that tradition-
ally has been called derivationalmorphology.As far as I can tell, facts that traditionally
have been treated under the separate heading of inflectionalmorphology must be handled
in completely parallel fashion to those discussed above. I know of no reason why the
list of morphemes should not include also the inflectionalaffixesor desinences,or why the
rules of word formation should not also include rules for positioning the inflectional
affixes appropriately or for handling such other inflectional phenomena as reduplica-
tion, stem Ablaut, etc.
It is important in this connection to realize that the three types of exceptional
behavior that have been handled above by means of the filter are not restricted to
word derivation but are found also in the inflection. Thus, one finds that particular
case forms of particular words idiosyncratically possess meanings that are in general
not those of either the base or the case. For instance, in Russian the instrumental case
of certain nouns designating times of the year and of the day has special adverbial
force that is not possessed by other nouns in the instrumental case. In particular,
letom may mean 'in summer', noc'ju 'at night', zimoj 'in winter'. However, avgustom
may not mean 'in the month of August', or obedommay not mean 'at dinner (or noon)
time'.
The second type of idiosyncratic behavior which I proposed to handle with the
help of the exception filter was phonological irregularity of the kind illustrated in (4).
Parallels to this type of exceptional behavior are found also within paradigms; i.e.
some forms in a paradigm are subject to a given phonological rule, others are not.
An interesting illustration of this is provided by the accentuation of Russian nouns,
which I have recently had occasion to study in some detail (see Halle I 972). As dis-
cussed there, a considerable number of Russian nouns must be entered in the dic-
tionary without stress marked on any of their vowels. Such words will then be subject
either to the Oxytone rule, which places stress on the last syllable, or to the Circumflex
rule, which stresses the initial syllable. It can be shown that the Oxytone rule must be
ordered towards the beginning of the phonological rules, whereas the Circumflex
rule must be ordered towards the end of the phonology. Moreover, in paradigms in
which one or more forms are subject to the Circumflex rule, the remaining form will be
subject to the Oxytone rule, but not vice versa. Finally, no form is subject to both
stress rules; words in this class are either Oxytone or Circumflex. These somewhat
elaborate facts can be handled straightforwardly by marking words subject to the
Circumflex rule with the rule feature [- Oxytone].
To see this more clearly, consider the feminine nouns of Russian that are stressless.
 PROLEGOMENA TO A THEORY OF WORD FORMATION

As shown in (7) we find in this class at least three distinct types of accent pattern. The
first, illustrated in (7a), has stress on the desinence in all case forms; the second, illus-
trated in (7b), has initial stress (Circumflex) in the nominative plural and desinential
stress elsewhere; the third group, illustrated in (7c), has initial stress (Circumflex) in
the nominative plural and accusative singular, and desinential stress elsewhere.
(7) Nom. sg. Inst. sg. Acc. sg. Nom. pl. Dat. pl.
a. kocerga kocergoj kocergu kocergi kocergam
gospoza gospoZoj gospozu gospozi gospozam
b. skovoroda skovorodoj skovorodu skovorody skovorodam
arba arboj arbu arby arbam
c. boroda borodoj borodu borody borodam
golova golovoj golovu golovy golovam
The facts illustrated in (7) will be accounted for in the following fashion. The
three types of noun forms under discussion will be entered in the morpheme list with-
out stress and none of the word formation rules will supply stress to them. This will be
sufficient to obtain the correct output for the nouns of type (7a). In the case of the
other two types of noun it will be necessary to provide the information that some of
their case forms are [- Oxytone], so that the Circumflex rule can apply to them. In
the light of the discussion above, this can be done quite straightforwardly by listing
in the filter the appropriate words, e.g. the accusative singular golovuand the nomina-
tive plural skovorodyas [- Oxytone]. In sum, the rules of word formation will generate
the inflected forms in the fashion to be expected; in most cases these will pass through
the filter without further effects. In the nominative plural and accusative singular
forms under discussion, however, the filter will supply a special marker indicating that
the words in question are to be treated in a special manner by the phonology. But
this implies that each of the case forms will appear as a special entry in the dictionary.2
Finally, paralleling the "accidental gaps" in derivation illustrated in (6) above
one finds various kinds of defective paradigms in the inflection. For instance, in
Russian there are about I OO verbs (all, incidentally, belonging to the so-called
"second conjugation") which lack first person singular forms of the nonpast tense.
Russian grammar books frequently note that such forms as (8) "do not exist", or "are
not used", or "are avoided".
(8) *lazu 'I climb'
*pobezu (or *pobezdu) 'I conquer'
*derzu 'I talk rudely'
*mucu 'I stir up'
*erunzu 'I behave foolishly'
2
A consequence of the treatment proposed here is that the absence of the fourth type of accentual pattern
(initial stress in the accusative singular only) will be regarded as a pure accident and not as a special subregu-
larity of the language which must be captured in the grammar. There are about a dozen nouns that have the
stress pattern of (7c). I know of no other interesting features that these nouns share in common.
8 MORRIS HALLE

The most recent scholarly grammar of Russian (Svedova 1970, ?988) remarks that no
ready explanation for these gaps has been offered. Thus, it has been suggested that
the first three forms cited above are avoided because they are homophonous with i.
sg. forms of other verbs. There are, however, other verbs in the language where the
identical homophony has not resulted in any gaps (cf. vozu 'I lead' or 'I cart').
Equally unconvincing is the suggestion sometimes made with regard to the last two
verbs cited in (8). It is said that these i. sg. forms are "difficult to pronounce" or
"unusual". But as is noted in the Svedova grammar note, since the language has i. sg.
forms exhibiting precisely the same behavior (e.g., vonZvu'I thrust ( a knife)' or sucu 'I
joke') this hardly is a plausible explanation. It would appear, therefore, that we are
faced here with an "accidental gap" in the dictionary. In view of what has been said
above, the natural way to handle these facts is to mark such forms as those in (8) as
being [- Lexical Insertion]. In other words, just like the forms in (6), those in (8) are
incapable of appearing in any well-formed sentence of the language in spite of the fact
that they exhibit no semantic, syntactic, morphological, or phonological abnormality.
At this point it might be useful to review briefly the proposal that has been made
above. To aid in this review I have prepared a block diagram of the proposal in Figure
i. I have suggested that morphology consists of three distinct components: a list of
morphemes, rules of word formation, and a filter containing the idiosyncratic prop-
erties of words. The list of morphemes and the rules of word formation together

Mo W ords
sFormation Xgof

output X Phonology Syntax

Figure I
 PROLEGOMENA TO A THEORY OF WORD FORMATION

define the potentialwords of the language. The set of actualwords is obtained from that
of the potential words by applying to the latter the modifications indicated in the
filter. One can think of the morphology, then, as producing a long list of words; it
is this list that is designated by the term dictionary.I should like to propose further
that the lexical insertion transformations be thought of as selecting items from the
dictionary and as entering these in appropriate slots in structures representing the
underlying constituent structure of particular sentences. It is to these underlying
representations that the syntactic transformations apply in the by now familiar
fashion and generate what has been called the surfacestructure.I shall assume that the
large body of phonological rules-rules like the overwhelming majority of those dis-
cussed in The SoundPattern of English or other works of generative phonology-apply
to the surface structure, and I shall disregard here the refinement that is necessitated
by the fact that at least some phonological rules apply as part of the transformational
cycle of syntax; cf. Bresnan (I97I) and (I972).
A few words must be said here about a problem which arises with regard to the
lexical insertion transformations. I proposed above that the lexical insertion trans-
formations have access only to words in the dictionary and, moreover, that the
dictionary contains only (and all) fully inflected forms of the language. This proposal
might appear to run into a difficulty, for the case which a given noun takes in a sentence
is normally determined by its position in surface structure, whereas lexical insertion
takes place at a much earlier stage in the derivation. This difficulty, however, is not
insurmountable. One might propose that instead of inserting a single item, the lexical
insertion transformations insert partial or entire paradigms, i.e. certain or all inflected
forms of a given "word". A perfectly general convention can then eliminate all but
the one inflected form that fits syntactically into the configuration in which the word is
found in surface structure. There are, of course, other equally plausible ways of sur-
mounting this difficulty. However, since I am concerned only to show that the diffi-
culty is not one of principle, I shall not explore here any of the other alternatives.
I do not find it surprising that the lexical insertion transformations characteris-
tically affect paradigms rather than single dictionary entries. It is well known that
paradigm pressure plays a potent role in the evolution of languages. For example it is
because of paradigm pressure that Russian has lost the consonantal alternations
k c- -, c in the nominal inflections. In fact, paradigm pressure provides a very
plausible explanation for the "accidental gaps" in the Russian conjugations illustrated
in (8) above. If paradigms can influence the evolution of language then there is every
reason to expect that paradigms must appear as entities in their own right somewhere
in a grammar. If my proposal is correct one such place would be the dictionary, from
which the lexical insertion transformations draw items for insertion into a sentence.
Note, incidentally, that if this is correct then the dictionary must be organized into
paradigms in some way and it would then no longer be equivalent to the logical
product of the morpheme list, the word formation rules, and the exception filter.
 10 MORRIS HALLE

I now turn to an examination of the character of the word formation rules. It is

hardly to be doubted that languages have words which consist of linear strings of
morphemes without internal structure, e.g. (9).

(g) serendip + i + ty vac + ant tot + al bro + ther hand + some

be + lieve
I shall assume that words of this type have the structure shown below by virtue of the
existence of word formation rules that express this fact, perhaps in the manner of
templates such as those in (io):

(io) [STEM + i + ty]N [STEM + ant]A [STEM + al]A [STEM + ther]N

[STEM + some]A [be + STEM]v
Presumably the stems in the list of morphemes will be appropriately marked so that a
given stem will be substitutable only in certain frames and not in others.
It is important to observe that the rules responsible for the formation of such
words as those in (g) will not only assign to each word the appropriate lexical
category (e.g. tell us that serendipityis a noun whereas handsomeis an adjective), but
they must also provide as much of the semantic and syntactic information concerning
the word as is general and shared by other words produced by the rule (the remaining
information, being idiosyncratic, will be provided by special entries in the exception
filter). In particular, the word formation rule must include some information about
the subcategorizational and selectional restrictions to which the words are subject.
Of a somewhat more complex structure than those above are words that are
derived from other words. Thus, in English we have nouns derived from verbs as in
arrival, refusal, or condensation;nouns derived from adjectives as in profanity, obesity;
verbs derived from adjectives as in darken,stiffen,blacken;or adjectives from verbs as in
explanatory,anticipatory,obligatory,etc. To capture such facts as these we would pre-
sumably have to have word formation rules of a form such as
(i i) [VERB + al]N [ADJ(+i) + ty]N [ADJ + en]v [N + ish]A
Word formation rules define, in part, the content of the dictionary. However, in order
for rules of the type illustrated in (i i) to operate properly, they must have access to
the dictionary, for it is only there that such crucial information as that arriveis a verb
of English will be found. In other words, it must be presumed that word formation
rules not only have available the information contained within the string on which
they are to operate, but also have access to the content of another component of the
grammar. This is an important formal difference between word formation rules and
the more familiar phonological rules.
That the content of the dictionary affects the formation of words is hardly to be
doubted. Thus the existence of arrival and confusionis one reason why English lacks
*arrivationand *confusal. However, it must immediately be noted that this cannot be
 PROLEGOMENA TO A THEORY OF WORD FORMATION

the whole explanation, for doublets such as recital and recitation,transmittaland trans-
missiondo exist side by side.
Like the word formation rules in (i o), the rules exemplified in (i i) will have to
provide information about general syntactic and semantic properties of the words
generated. In this connection the question immediately arises how a grammar is to
reflect the fact that a derived word quite commonly shares semantic and syntactic
properties with the word from which it is derived (with its source word). It must be
noted that the sharing of common properties is normally far from complete; as docu-
mented most recently by Chomsky (I972), there are numerous asymmetries as well as
apparent and real irregularities in this domain, all of which must be properly treated
by an acceptable theory of language. I review briefly the most important types of case
that appear to arise here.
The most typical situation is one in which the derived words have special syn-
tactic and semantic properties different from those of their source words. Thus, for
example, by adding the suffix -hood to nouns designating human beings, nouns are
produced that designate a state or quality such as boyhood,priesthood,etc. At the very
least the word formation rules will have to supply the information that unlike their
source words, the nouns derived with this suffix are abstract. The word formation rule
might, in this case, assign to the derived nouns the feature [ + Abstract].
Somewhat more intricate are the cases where words of one lexical category are
derived from words of another category; e.g. the noun refusal from the verb refuse.
Cases of this sort might be handled along the lines of the Base Structure Hypothesis
suggested by Chomsky (I972), which impresses me as the most plausible means yet
proposed for capturing the obvious linguistic universal that each lexical category has
characteristic syntactic and semantic properties of its own.
Related to the preceding, but requiring some further machinery in the rules of
word formation, are those cases where the same selectional restrictions apply in one
subcategorization frame in the source word, and in another subcategorization frame
in the derived word. For example, verbs derived from adjectives with the suffix -en
can take as their objects those nouns of which the source adjectives can function as
predicates. It is necessary, therefore, that the word formation rule be capable of
treating selectional restrictions independently of the subcategorization frames in
which these restrictions are embedded. In view of my very rudimentary understanding
of what is involved here, I am not in a position to make useful concrete proposals about
how this might be implemented. I trust, however, that the nature of the problem is
clear from the brief remarks above.
As has been noted repeatedly above, word formation rules function in such a way
as to involve a large number of exceptions and idiosyncrasies of all sorts. These will be
handled with the help of the exception filter in the manner sketched above.
Returning once again to the rules illustrated in (i I), it is necessary to observe that
in spite of the nested constituent structure of the words generated by the rules
12 MORRIS HALLE

illustrated it is not the case that phonological rules invariably apply to derived words
of this sort in the familiar cyclic fashion. While some derived words have nested con-
stituent structure at the stage where phonological rules apply, others do not. For
example it has been argued in The SoundPattern of English and elsewhere that words
such as those in (i2a) must be presumed to have internal constituent structure,
whereas words such as those in (I 2b) must be presumed to be formed by a linear
concatenation of morphemes.
(i 2) a. exaltation relaxation elasticity obligatory declarative assimilatory
generative
b. consultation information

The reason for this differential treatment of what appear to be very similar words is
obscure at present. I know, however, of no other plausible way of accounting for the
differences in stress and reduction in the pretonic vowel than by postulating a differ-
ence in constituent structure.
Perhaps somewhat more perspicuous is the following example from Russian. We
recall that in Russian a string of morphemes containing no stressed vowel will be
subject to the Oxytone rule which assigns stress to the last syllable. To form diminutive
nouns, Russian makes use of the rule
(I3) [N + Ak]N where N represents a noun
Nouns formed by (I 3), like all nouns of Russian, are subject to a further rule which
spells out the individual case form by adding a desinence to the suffix, i.e.
(I 4) [N + ?k + Desinence]N
Accordingly, the diminutive form of the noun gorod'town' is in the dative plural
gorod + bk + am
Since gorodis an unstressed stem the Oxytone rule will apply here and we shall get as
output
gorod + bk + am
It is obvious that the diminutive just generated is a noun and hence can by itself
occupy the N position in (I3). We find, therefore, in Russian the noun
gorod + zk + bk + am [garadockam]
Curiously, here the stress is placed not on the desinence as might have been expected
but on the first diminutive suffix. This result can be obtained with no difficulty if the
Oxytone rule is allowed to apply cyclically and the string under discussion is supplied
with the constituent structure:
[[gorod + AkIN + bk + am]N
 PROLEGOMENA TO A THEORY OF WORD FORMATION

It turns out that when z/k is suffixed to a noun which itself is formed by suffixation,
then the resulting noun has internal constituent structure; when Zk is suffixed to a
nonsuffixed noun the resulting word does not have constituent structure. This is still
far from an adequate explanation, but it suggests a direction in which to explore
further.
It has been noted above that rules of word formation must have access to the
dictionary; i.e. that certain words presuppose the existence of other words. This fact
would require that word formation rules be formal devices of considerable power. It
would, however, seem that even this increase in power is not sufficient and that ad-
ditional power is required by these rules. To see this, consider the formation of the
inchoative verbs in English with the suffix -en. These deadjectival verbs have recently
been studied by Siegel (I972) who has noted that they are subject to the phonetic
condition that their base must be monosyllabic and end with an obstruent, which
optionally may be preceded by a sonorant. It is for this reason that we have the verbs
in (I sa) but not those in (I5b).

(I 5) a. blacken whiten toughen dampen harden

b. *dryen *dimmen *greenen *laxen

It should be noted that the "words" in (I 5b) are phonetically well formed in English
as shown by the existence of the words in (i6).
(i 6) lion women Keenan flaxen

The above phonetic condition, therefore, is limited to the particular verbs under dis-
cussion. The condition, however, appears to be a condition not on the string formed by
the rule but rather on the string after the phonological rules have applied to it.
To see this, consider the verbs

(I 7) soften fasten moisten

These verbs are clearly derived from the adjectives and would have in their underlying
representation the form

(i8) soft + n fast + n moyst + n

These strings, however, violate the constraint against obstruent sequence appearing
before the inchoative suffix. It is, of course, immediately obvious that because of the

phonological rule which deletes [t] in the environment s-{ } these forms will appear
in the output with a single obstruent, thus conforming to the condition above. But if
this is indeed the correct reason for the well-formedness of moistenas contrasted with
the ill-formedness of laxen, then it must be supposed that the rules of word formation
have access not only to the dictionary but also the output of the phonology. In other
14 MORRIS HALLE

words, we are saying, in effect, that a word formation rule will produce acceptable
words if the words formed by this rule conform to certain conditions when they have
been acted upon by the rules of the phonological component. This is clearly quite unlike
the more familiar rules of phonology which will or will not apply to a string depending
only on the characteristics of the string in question.
To support further the need for word formation rules to have access to the output
of the phonology I discuss below the formation of the present adverbial participle of
certain Russian verbs. This form is composed of the verb stem which includes the root
followed optionally by any number of derivational suffixes terminating with a so-
called verbalizing suffix; a present tense marker, which is either i or e; and the present
tense adverbial participle marker n (cf. Lightner I965):

Root + (suffix) (verbalizingsuffix)+ { } +n

Russian phonology includes a process termed "transitive softening" which consists in

the replacement of labials by clusters of labial and palatalized liquid, and of other ob-
struents by palatals (t, k c-> v; d, g, z -z '; s, x -* s3).Thus we find alternations such as
those in (i 9).
(I9) past (fem. sg.) present(3. sg.)
sypala syplet 'pour'
toptala topcet 'stamp'
dvigala dvizet 'move'
rezala rezet 'cut'
pisala piset 'write'
pekla pecet 'bake'
The conditions under which "transitive softening" takes place are statable in
purely phonological terms (cf. Lightner I965), and these conditions would inciden-
tally be satisfied by the strings underlying some present adverbial participles. We would,
therefore, expect such adverbial participles as
(20) *syplja *topca *dviza *pisa *peca *reza
As a matter of fact, none of the words in (20) is actually used (cf. Svedova I971, ?I012)
because there appears to be a very strong tendency in the language to avoid present
adverbial participles of verbs that exhibit "transitive softening" in the present ad-
verbial participle but not in certain other forms (e.g. past tense).3 It should be stressed
that the restriction against forms that have undergone "transitive softening" applies
only in the present adverbial participle but not in other words derived from the same
stem. Thus, the nouns in (2 i) are perfectly good Russian words, although each one of
them is subject to "transitive softening".

3 This tendency might also be one of the reasons for the "gaps" in the paradigms cited in (8) above.
 PROLEGOMENA TO A THEORY OF WORD FORMATION

(2I) kaplja 'drop (of liquid)' vstreca 'meeting' gospoza 'lady' nosa 'burden'
sveca 'candle'

When examples of the sort just presented have been discussed the assumption has
usually been made that rules of word formation are part of the phonology (cf. e.g.
Chapin I970) and the conclusion has, therefore, been drawn that phonological rules
must be much more powerful devices than had heretofore been supposed. In particular,
in past studies of generative phonology it has been assumed that the decision of whether
a given rule should or should not be applied to a string depends solely on the com-
position of the string at the point in the derivation where the rule in question applies;
it does not depend in any way on the shape of the string at a later or at an earlier point
in the derivation. But as we have just seen, there are aspects of language that require
the power of rules which have access to several stages in the derivation at once. It is
obvious that, in general, one would not wish to replace less powerful by more powerful
devices especially when it is known that the less powerful devices are capable of
handling a very large part of the task at hand. Under such circumstances, it would be
essential to attempt to limit as much as possible the domain in which the more power-
ful devices may be invoked. There seems to be a fairly natural way of achieving this,
given the framework of grammar sketched above (cf. Figure I) where the rules of word
formation are distinct and separate from the rules of phonology. I would like to pro-
pose that the added power of having access to different stages in a derivation be
available only to word formation rules, whereas the rules of phonology be restricted,
as in the previous work, to information overtly present in the string at the point in the
derivation at which the phonological rule applies.
The proposal just made amounts to saying that word formation is a fundamentally
different process than phonology. In fact, it may well be useful to speak not of "rules
of word formation" but rather, as has been suggested by Lakoff and others, of "de-
rivational constraints that hold in word formation". In the case of word formation we
are dealing with conditions that no string of morphemes can ever violate if it is to be
admitted to the dictionary as a legitimate word of the language. While my own
investigations of word formation in different languages are not extensive enough to
allow me to place too much confidence in general impressions that I have gathered,
nonetheless it seems to me significant that I have yet to come across any clear instances
where word formation rules have to be ordered in that tightly constrained fashion that
is constantly encountered in true phonological rules. Moreover, to the extent to which
I have been able to investigate proposed instances of "derivational constraints", these
could always be captured with the help of rules that could be ordered ahead of the
bulk of the phonological rules. Hence it is likely that these "derivational constraints"
could be incorporated into the word formation component. I should like to propose
therefore that the word formation component differs from the phonology by having
completely different principles of interaction among rules. Whereas in the phonology
i6 MORRIS HALLE

this interaction is captured by means of the convention of linear order of rule applica-
tion, the interaction among word formation constraints may require a different
principle altogether; e.g. simultaneous application. Needless to say, at this point this
must remain in the realm of pure speculation.
I have proposed above that the syntactic component has direct access to the
dictionary; i.e. that the lexical insertion transformations take items from the dictionary
rather than from the list of morphemes. Although the content of the dictionary is
entirely determined by the content of the list of morphemes, the rule of word forma-
tion and the exception filter, there is no need to assume that these components are
always fully involved in every speech act. Instead it is possible to suppose that a large
part of the dictionary is stored in the speaker's permanent memory and that he needs
to invoke the word formation component only when he hears an unfamiliar word or
uses a word freely invented. While this is by no means an exceptional occurrence, its
frequency is quite low. There is a fundamental difference between the use of words and
the use of sentences. In general, one uses familiar words, words one has heard and used
before, and one does not expect to use or encounter new words, whereas one rarely uses
sentences that one has encountered before. From the viewpoint of performance one
might say that the role played by the rules of syntax and phonology differs funda-
mentally from that played by the rules of word formation. The knowledge represented
by the latter might be said to be more passive than that represented by the former. If
this were indeed the case, it might serve to explain the striking differences that appear
to exist with regard to ordering, principles of application, etc. between rules of word
formation and those of other components of the grammar.

References
Bresnan,J. (I971) "Sentence Stress and Syntactic Transformations,"Language47, 257-!8I.
Bresnan, J. (I972) "Stress and Syntax," Language48, 326-342.
Chapin, P. (I970) "On Affixation in English," in M. Bierwisch and K. E. Heidolph,
Progressin Linguistics,Mouton, The Hague.
Chomsky, N. and M. Halle (i 968) TheSoundPatternof English,Harper and Row, New York.
Chomsky, N. (I972) "Remarks on Nominalization," in N. Chomsky, Studieson Semanticsin
GenerativeGrammar, Mouton, The Hague.
Halle, M. (1972) "The Accentuation of Russian Words," Language49 (to appear).
Jespersen, 0. (1946) ModernEnglish GrammarPart VI: Morphology, Allen and Unwin, Ltd.,
London.
Lightner, T. (I965) SegmentalPhonologyof ModernStandardRussian,unpublished Doctoral
dissertation, MIT. See also idem., Problemsin the Theoryof Phonology,Volume I,
Linguistic Research, Inc., Edmonton, I972.
Siegel, D. (I97I) "Some Lexical Transderivational Constraints in English," unpublished
paper, MIT.
Svedova, Ju., ed. (I 970) Grammatika sovremennogo
russkogoliteraturnogojazyka,
Nauka, Moscow.
Department
of Linguistics
MIT
Massachusetts
Cambridge, 02139