Analogy in Semitic Non-concatenative Morphology Robert R. Ratcliffe Tokyo University of Foreign Studies 1. Analogy in linguistic theory Analogy has always been something of an embarrassment for linguistic theory. In the Neogrammarian model analogic change had to be admitted in order to account for the most glaring exceptions to the hypothesis of regular sound change (and hence in order to maintain the hypothesis at all). Yet unlike sound change, analogy was conceived as operating according to no regular principles of its own, and as being regular only in its result (to paraphrase Sturtevant's famous paradox). Since analogic change seems to provide, at least ostensibly and pre-theoretically, a salient example of the interplay of surface language forms and internal or mental language operations, one might have expected that the rise to prominence of mentalist approaches to linguistics in the nineteen-sixties and seventies might have led to a greater interest in analogy and perhaps to an attempt to formulate explicit principles governing it. In fact, and somewhat surprisingly, Generative Grammarians have taken a hostile view of the traditional notion of analogy. For Generative Grammar the problem is that conceding any role to analogy in language acquisition or grammar construction implies a weakening of the innateness hypothesis (that the rules of grammar are innate in the newborn infant and not acquired.) These two theoretical paradigms -- Neogrammarian and Generativist-- continue to shape discussion of analogy in recent literature on morphological change. Mainstream historical linguists (represented, for example, by Hock 1985 or Bynon 1993) emphasize the role of proportion in analogic change and deny any role for rules (in the sense of a psychologically real abstract linguistic knowledge). Linguists influenced by or oriented toward the Generative tradition (Kiparsky 1974 or Anderson 1992, for example), emphasize the mental basis of analogy, but tend to be at great pains to avoid conceding any innovative role to analogy, arguing instead that analogy necessarily represents extension of a pre-existent rule.1 Yet whatever theoretical approach is argued for or adopted, virtually all of the discussion rests on a relatively narrow data base of examples drawn primarily from Romance and Germanic languages. As a consequence both traditional and generative approaches rest upon two unstated assumptions which deserve to be explicitly stated and called into question: The first is that consistency or regularity in a morphological system always consists in some consistent difference or change between two words, (indeed it tends to be taken for granted that the "normal" type of morphology is addition of an affix to a base). The second is that the characterization of such differences or changes is essentially a trivial matter. That is, it seems to be taken for granted that there is one optimal way to characterize any given relationship and that all linguists would agree in general about what that is, although there may be a purely theoretical disagreement about the ontological status of such a characterization -- whether it is a psychologically real rule which is part of a speaker's grammar, or simply a metalinguistic generalization which is psychologically real only for the linguistic analyst. 2. Problems from Semitic 2.1 Five-part analogy (?) If we expand our data base to include Semitic languages, we will find that neither of these assumptions stands up to scrutiny. In the first place a salient feature of the morphology of many Semitic languages is the existence of fixed syllabic/vocalic patterns, associated with particular semantic or syntactic functions. In pattern-based morphology consistency is found not in the relationship between a base word and a derivative but in the form of the derivative. As an example of how fixed pattern morphology works, consider the so-called 'stem II,' causative/factitive verbs in written Arabic, as represented in (1). There is a clear pattern here, yet it cannot be described as a consistent proportion or relation between base and derived form, since for example, xaraja is not related to xarraja in the same way that jildun is related to jallada nor in the same way that ruxaamun is related to raxxama. (1) Stem II (causative) verbs in Modern Written Arabic2 source verb (perf. asp/tns.) form II (perf. asp./tns) CaCCaCa a. xaraja "go out" xarraja "cause to go out" daxala "enter" daxxala "cause to enter" b. jildun "hide, leather" jallada "bind a book" ruxaamun "marble" raxxama "pave w. marble" dhahabun "gold" dhahhaba "gild" c. baabun "door, chapter" bawwaba "divide into chapters" buklatun "buckle" bakkala "to buckle xaimatun "tent" xayyama "to pitch a tent" Pattern-based morphology thus forces us to recognize that the first two parts of a classic four part analogy are in fact three parts: an input, an output, and a relationship between them. In theory it is quite possible that some consistent formal feature of the output (such as syllable structure or vocalism) might be imitated and extended in language change rather than the proportion relating input and output. In fact there is evidence (some of which we will present below) that precisely such an output-based (as opposed to proportion-based) analogic strategy has been exploited in the historical development of some Semitic languages. Indeed the Arabic verbs in (1b) and (1c) might be taken as one example, although an imperfect one, since they represent the results of developments within the written language. It seems that the Proto-Semitic ancestor of the stem II (also termed "D-stem") verb was exclusively or at least primarily a deverbal derivative. Semantically it broadly had the effect of pluralizing the action of the source verb -- either indicating intensity of the action of the source verb or adding an argument to the source verb -- making an intransitive transitive or a transitive causative (Moscati et al. 1964:124). The denominative use of the pattern is perhaps also already Proto-Semitic, but it becomes particularly widespread as the way of forming new verbs from nouns in post-Qur'anic Arabic and is quite productive in modern times even for borrowed words, as the examples in (1c) show. The deverbal stem II verb can be most simply analyzed as related to its source according to a regular proportion or consistent change--that is it can be analyzed as formed by geminating the second stem consonant of the source verb (as long as the stem of the source has three consonants). However, since the denominal stem II verbs cannot be analyzed simply in terms of gemination of the second stem consonant, extension of this rule or proportion will not account for them. Rather, we must suppose that the stem II verb has been interpreted as formed according to a fixed CvCCvC pattern, and the denominal stem II verbs have been made to conform to this pattern by making whatever changes are necessary-- treating long vowels as containing an underlying glide when the stem has only two consonants (as in baab-), ignoring long vowels when the stem has three (as in ruxaam-). 2.2 Rule construction and variation in grammar The second problem which Semitic historical morphology presents for traditional assumptions about analogy is a consequence of the first. Since output-consistent morphology exists in the system alongside change-consistent morphology, it may not be clear for a given set of paired forms in isolation whether the optimal analysis is in terms of regular change or a regular output. For example the pair kalb :: kulaib "dog" :: "little dog" (diminutive) may be analyzed as a regular change-- infixation of -ai- after the first CvC- sequence, plus vowel change-- or in terms of imposition of a fixed diminutive pattern CuCaiC. This appears to be a problem for formal linguistic analysis. But evidence from language change indicates that native speakers may confront the same problem and that different speakers may derive different rules from the same surface contrast. We are thus led to propose that analogy involves first the CREATION of a rule on the basis of surface patterns, and THEN the extension of the rule to new contexts. In other words we are forced to distinguish sharply the hypothesis of mentalism from the hypothesis of innatism-- analogic change requires the hypothesis that linguistic knowledge consists partially of rules or patterns more abstract than what is superficially observable, but it also requires that that knowledge is to some extent created or derived on the basis of what is superficially observable, rather than being innate. 3. Analysis of data-- case studies Having outlined what I believe to be the theoretical prerequisites to an understanding of the operation of analogy in systems which exploit fixed output morphology, I will now turn to specific case studies of two developments from Classical Arabic (CA) to Moroccan Arabic (MA) which I believe provide evidence of the working of these principles. 3.1 The development of new 'w-' initial verbs in Moroccan Arabic Since fixed- pattern morphology works on the principle of imposing a fixed pattern on varying inputs it is sometimes the case that the input must be radically altered to fit the output pattern. In the Classical Semitic languages this is most notably the case with words whose underived forms have only one or two stem consonants. In most Classical Semitic languages there are only a small number of these. But the derived forms of such words are usually reshaped to fit the patterns of the much more populous three consonant words, by the addition of a default consonant. (At the risk of confusing synchronic and diachronic notions we might say they are reshaped on the --synchronic-- analogy of the three consonant words.) The following verbal forms in Classical Arabic illustrate the point: (2) two-consonant verbs in CA stem impf. perf. part. stem II perf. CvCvC CvvCvC CvCCvC CvC yaqaCu waqaCa waaqiCun waqqaCa "fall" yaSilu waSala waaSilun waSSala "arrive" yasimu wasama waasimun wassama "mark" cf. yawbaru wabara "be hairy" yaktubu kataba "write" These verbs are traditionally analyzed as deriving from a root with initial 'w', where the 'w' is deleted by phonological rule in the imperfect. This analysis will not hold, however, since there is no productive phonological rule deleting w in this environment (cf. waTan >> ?awTaan "nation" >> "nations"), and since verbs with stable initial w also exist as the last example in 2 indicates. In MA the CA verbs with two consonants in the stem have been regularized so that the w is maintained throughout the paradigm. However a new set of irregular verbs has arisen as a result of a regular sound change deleting syllable initial glottal stop, which also results in the loss of the vowel in a word initial ?V- sequence. The results of this change are seen for example in MA comparative adjective patterns CCaC (kbar = CA ?akbar "bigger"), and MA plurals CCaC (mwaj = CA ?amwaaj- "waves"). Hence from CA verbs containing a stable glottal stop, such as ya?xudhu/?axadha "take," ya?kulu/?akala "eat" and yar?a(y)a/ra?a(y)a, "see," a new set of two-consonant or even one-consonant verbs should develop by regular sound change. In fact these verbs have been reshaped in MA in the following ways: In the perfect MA verbs like kla/klit ("he ate, I ate" = CA ?akala/?akaltu) have in effect been supplied with a default third consonant, being reshaped on the model of verbs with a final y such as MA rma/rmit "he threw, I threw" (= CA ramaa/ramaytu). In the participle and the stem II verb, a default initial consonant 'w' is supplied. (That 'w' is not simply the regular reflex of glottal stop is clear from the comparative and plural patterns mentioned above as well as from the existence of at least one stem II verb with initial glottal stop that has developed regularly without reanalysis dda, yiddi "bring convey" = CA. ?addaa, yu?addi .) The CA verb meaning "to see" is not preserved in MA except in the moribund Jewish dialect (Harrell 1966). Other dialects of MA preserve a trace of this verb in the fossilized deictic particle /ra/. However the common verb meaning "to show" shows the same development of forming a new stem II verb by addition of a default 'w' to the left edge of the stem. (3) new form II verbs in Moroccan Arabic (Heath 1987) imper. impf. pf. participle form II perf. CA kul ya?kulu ?akala ?aakil- ?akkala MA kul yakul kla wakel wakkel CA xudh ya?xudhu ?axadha ?aaxidh- ?axxadha MA xud yaxud xda waxed waxxed CA ra yar?aa ra?aa raa?ii *ra??aa (?ar?aa) MA (ra) (yra) (ra) warra cf.: CA udxul yadxul daxala daaxil- daxxala MA dxul idxul dxel daxel daxxel CA Sil yaSil waSala waaSil- waSSala Ma wSel yewsel wSel waSel waSSel 3.2 allomorphy and ambiguity: developments in the diminutive In CA the diminutive for nouns of three mora CvCC-, CvCvC-, CvvC- (the most common forms for underived nouns) has a bisyllabic, four-mora pattern CuCaiC-. For four-mora nouns CvCCvC the diminutive pattern is the five-mora, trisyllabic CuCaiCiC, and for five-mora nouns of the form CvCCvvC, the diminutive is the six-mora pattern CuCaiCiiC. Thus the data presents the set: (4) CA diminutives base diminutive examples CvCC >> CuCaiC kalb- >> kulaib- "dog" CvCCvC >> CuCaiCiC masjid- >> musaijid- "mosque" CvCCvvC >> CuCaiCiiC sulTaan- >> sulaiTiin- "Sultan" If the data offered only these three forms, we would have to conclude that the diminutive is formed according to a regular change or consists in a regular proportion between input and output rather than being based on a fixed output pattern. This proportion might be characterized in terms of a set of rules adding the sequence 'ai' after the first heavy syllable of the input., and changing the vowel of the first syllable to 'u'. (A third rule might be needed to account for the 'i(i)' for the vowel of the third syllable of forms with three syllables. But since /i/ is the normal quality of the epenthetic vowel, we might simply assume that this is supplied by default). (5) diminutive formation rule 0 >> ai / {S(mm)__ v >> u/ {C_ However, in order to account for the diminutive patterns of two-mora nouns and six-mora nouns, formation on the basis of a fixed output has to be assumed, because in these cases extra segments are added or deleted respectively to bring the outputs into conformity with the patterns of the four-mora or five-mora diminutive patterns, respectively. (6) diminutives of "over-short" and "over-long" nouns CvC >> CuCaiy ?ax >> ?uxaiy "brother" CvCCvCvvC >> CuCaiCiC Cankabuut- >> Cunaikib- "spider" Most interestingly, some four-mora nouns, notably those with three consonants and a long vowel in the first syllable, CvvCvC, and nouns whose first mora is supplied by a transparent prefix have alternative four-mora (CuCaiC-) and five-mora (CuCaiCiC) diminutive patterns. The five-mora pattern is more common in CA and is the only productive pattern in modern written Arabic. However, a residual four-mora pattern for some of these nouns are also cited in the medieval grammatical sources. Some examples, cited by Wright (1897:282-3), are given in (7): (7) residual diminutive patterns (in parentheses) Haarith- >> (Huraith-) Huwairith- "plowman" Haamid- >> (Humaid-) Huwaimid- "sour" miCTaf- >> (CuTaif-) muCaiTif- "coat" ?aswad >> (suwaid-) ?usaiwid- "black" I hypothesize that this variation has arisen because speakers have, on the basis of the same formal data, derived a different procedure for forming the diminutive. If speakers assume that the diminutive is formed by an infixation rule, then proportional analogy would give the pattern CuwaiCiC as in (8). (8) rule: infixation of -ai: 0 >> ai /S(mm)- basis b >> ai- b >> kulaib extensions rith >> ai- rith >> Huwairith Taf >> ai-Tif >> muCaiTif If speakers assume that the diminutive is formed on a fixed pattern, then segmental material must be deleted to fit four-mora inputs onto a four-mora output pattern. I assume that the process involved here is consonant extraction (Bat-el 1986, Heath 1987) and mapping to a prosodic template (McCarthy 1979, 1981). (9) rule: mapping (stem consonants only) to pattern CuCaiC basis extensions kalb Haarith miCTaf | | | | | | / | | CuCaiC CuCaiC CuCaiC kulaib Huraith CuTaif Historically, the sequence of developments leading to this situation might have been as follows: 1) To begin with the internal diminutive was originally restricted to underived, three-mora nouns. (I have no evidence for this in the case of the diminutive, but this seems from comparative evidence to have been the range of the internal plural, whose formal parallels with the diminutive have long been recognized), see Ratcliffe 1992 and to appear. 2) The internal diminutive was then extended to four mora nouns, according to an output-based analogy (as in 9)-- yielding the four-mora CuCaiC diminutive patterns which are still attested for some of these nouns in CA. (We assume this step to have preceded the following on the basis of the principle ensconced in Kurylowicz's fourth law of analogy that the more marginal or less productive means of marking a category should be older than the primary or more productive means.) 3) The strategy of forming a diminutive of these nouns according to a CuCaiC pattern is unsatisfactory, however, because it requires loss of a significant phonetic material, possibly obscuring the connection between the diminutive and the source word. Hence a new strategy, based on proportional analogy (as in 8), is employed to create a new set of diminutive patterns for four-mora nouns. The system of CA has continued to develop in the modern Arabic dialects. It is perhaps not surprising to find that in MA, for example, a reflex of the four-mora CuCaiC pattern no longer exists as a possible diminutive pattern for the reflexes of four-mora input nouns. Yet what is perhaps surprising is that the four-mora diminutive has also virtually disappeared as a possibility for the reflexes of three-mora and smaller nouns. By regular sound changes (loss of unstressed short vowels in many environments, reduction of the diphthong 'ai' to 'i') the CA CuCaiC pattern would appear as *CCiC in MA. Yet the reflexes of classical Arabic CvvC stems (CVC, where V is one of the stable vowels 'a' ,'i' or 'u') and CvCC stems (CvCC or CCvC where 'v' is 'o' or '\') in fact form a bisyllabic diminutive CCiCeC or CCiCa, as illustrated in (10). (10) MA diminutives (Harrell 1962, Heath 1987) bab >> bwiyeb "door" bwiba kelb >> kliyeb "dog" bghel >> bghiyel "donkey" bden >> bdiden "body" kbir >> kbiber "big" bakit >> bwiket "packet" mektub >> mkiteb "pocket" begra >> bgira "cow" Undoing the effects of sound change we arrive at a subsystem of (11). (11) MA diminutive system reconstructed mmm mmmm m m m m m *CVVC >> *CuwaiCaa *CvCC >> *CuCaiyiC " " >> *CuCaiCiC *CvCiiC >> *CuCaiCiC *CvvCvC >> *CuwaiCiC *CvCCvC >> *CuCaiCiC *CvCCv(v) >> *CuCaiCvv If we contrast this with the corresponding sub-system in CA, as represented in (12), the nature of the analogic reshaping which has taken place becomes clear: (12) CA System (Wright, pp. 166-175) mmm mmmm mmmm mmmmm CvvC >> CuwaiC CvCC >> CuCaiC CvCvvC >> CuCaiyiC CvvCvC >> (CuCaiC) CuwaiCiC CvCCvC >> (CuCaiC) CuCaiCiC CvCCvv >> CuCaiCaa In brief, the diminutives of historically three-mora nouns are now also (like those of historical four-mora nouns) formed on the five-mora pattern, with default segments added to fill out the pattern. The extra segment is either a consonant which provides the onset of the final syllable or a stable vowel which provides the nucleus of this syllable. The default consonant in this case is either y or a copy of the second stem consonant. It appears then that speakers have once again employed a fixed-output analogy to derive a new diminutive pattern for three-mora nouns. The motivation for the change is the elimination of allomorphy. If speakers assume that the paired sets in (12) (other than the residual patterns in parentheses, which have presumably long been lost before the change in question takes place) are related according to a regular proportion, then there is no allomorphy. But if they assume that the diminutive is formed on a fixed pattern, then there are two principle patterns for the diminutive-- CuCaiC and CuCaiCiC-- and hence allomorphy. Since fixed pattern morphology is widely found in other parts of the morphology, it is logical that speakers would try to interpret this subsystem also in fixed pattern terms and try to regularize it by generalizing one of these patterns. Imposition of a four-mora pattern on four-mora nouns is not possible without deletion of some meaningful bit of word structure. Imposition of a five-mora pattern on three-mora nouns, however, can be accomplished through the addition of meaningless segments; and this appears to be what has happened. The change has, however, created allomorphy elsewhere-- in the form of variation in the choice of the default segment. Hence we have the paradox that analogy has created allomorphy. Indeed, if our interpretation of earlier developments in the system is correct, then this is not the first time this has happened, since the application of the proportional analogy under step (3) above leads to the apparent allomorphy in patterns which is resolved in historical times in MA. 4. Classic Problems There are a number of seemingly intractable problems in Semitic historical morphology which have emerged as problems, I believe, precisely because the phenomenon of output-based as opposed to proportional analogy has not been recognized. I address two such problems in this subsection, arguing that each case simply represent the endpoints of developments quite similar to those we analyzed above in the history of Moroccan Arabic. 4.1 Words for "one" in Semitic languages (Lieberman 1991) Lieberman (1991) has cited a pattern of apparent correspondences of w, ?, and 0 (zero) in the classical Semitic languages as an example of an irregular sound correspondence which cannot be explained by "borrowing, phonological rules or analogy," (p. 311) and which therefore forces rejection of the Neogrammarian regularity hypothesis. Briefly the data as presented by Lieberman (304-311) is as follows. There are three stems associated with the meaning "one" in Classical Semitic languages: 1) a two consonant noun form which we can reconstruct as *Had, which is the normal word for one in Syriac/Aramaic and found sporadically in Hebrew and Punic and in restricted contexts in Arabic (in the word for "eleven" HadCashar, for example); 2) a three consonant noun form with initial glottal stop *?aHad, which is the normal word for "one" in Ethiopic, Arabic, Hebrew, Ugaritic, and Sabaic; and 3) a pattern with initial w *waHad, which appears as a verbal pattern in Akkadian, Ugaritic, El-amarna, Hebrew, Old Aramaic, Syriac, Ethiopic and Arabic. In Arabic adjectival (participle) and nominal (verbal noun) formations from this stem are also found. In the Northwest Semitic languages the w is reflected (regularly) as y. (The w-y correspondence is regular and uncontroversial.) Chart (13) gives the relevant reflexes in the major languages. (13) Words for "one' in Semitic Languages (Lieberman 1991) Arabic Hebrew Syriac Ugaritic Sabaic Hadh Hadh (noun) ?aHad ?aHad aHd ?Hd (noun) waaHid (adj.) waHada yeHad yawHed yHd (verb) (ya-Hidu) To begin with, it seems from the relative distribution of the forms in question that Lieberman has confused a problem in grammatical analysis for a problem in comparative linguistics. Chart (13) shows clearly that the problematic irregular correspondences exist not across languages but across grammatical categories within each language. On the basis of the comparative method we can construct three items-- a verb and two nouns; and the correspondences in the first consonant of the reflex of each of these items is entirely regular. It is only if one assumes that each of these items is somehow the SAME or derives (synchronically or diachronically) from the same source, that a problem arises. But there is nothing in the comparative method that forces such an assumption. Of course it seems extremely likely, given the similarity in form and meaning, that the verbal and nominal forms are or were related within a synchronic derivational system at some historical level, and the question is "how?". Lieberman's solution is to take the verbal forms as the source of the nominal forms: "The numerals may well have been derived from verbs. "Simple, unadorned, derivation from the verbal system will not however account for the interchange between /?/ and /w/ which we seem to find in the Arabic words for 'one', and the apparent relationship between the cardinal number 'one' and the verbs which begin with /w/ or /y/ in the other languages" (p. 310). Thus he is led to the conclusion that an irregular sound change must have occurred and that the principle of regular sound change is therefore invalid. Being unwilling to abandon the principle so easily, I am motivated to seek another solution. The variation in the stem form of words meaning "one" reconstructable for Proto-Semitic is no greater than the variation in the stems form of words meaning "eat" and "take" in MA. It is most likely that the Proto-Semitic variation also reflects a process of adding default segments to fill out a fixed pattern. I would propose that the two-consonant Had is the original noun form meaning "one." The verbal form "to be one" as reflected in CA is still a two consonant verb ya-Hid-u, which has apparently been derived from the noun simply by adding the set of verbal prefixes and suffixes. The initial 'w' which appears in the perfect stem waHada, the participle waaHidun and other derivatives of the verb is simply the regular default consonant required to fill out the fixed patterns associated with these categories. There is another bit of evidence that suggests that the process of deriving verbs from two-consonant nouns in this way was active in Proto-Semitic. The CA verb yasimu/wasama "mark, distinguish" was possibly derived at an earlier period from a noun meaning "name" (CA /ism/) which can be reconstructed as a two-consonant stem for both Proto-Semitic and Proto-Afroasiatic. This isn't the normal verb "to name" in CA, which is yusammiy/ samma(y)a formed on the stem II pattern. But as we have noted the use of the stem II pattern as the normal denominative verb is a relatively recent analogic trend. For the nominal form with glottal stop, it is possible that Hodge's solution, adopted also by Lieberman, is the correct one--that the ?a- is a prefix associated with numbers (found also in ?arbaC "four") which has been prefixed to the two-consonant base. Alternatively the ?a- may also represent a default filler consonant added by analogy at an earlier period of the language and later reanalyzed as a stable part of the stem. Finally, we could complicate Lieberman's original problem further by noting that CA has yet another stem for "one" based on the consonants H-d-y which appears in the ordinal number Haadiy Cashr "eleventh." The form Haadiy is derived from the Had of HadCashr "eleven," and the 'y' is a default consonant supplied to fill the last C position of the ordinal number pattern CaaCiC. None of this seems at all unusual in light of evidence from MA. MA verbs like y\ws\l provide a case of reanalysis and stabilization of a default consonant. Variation like that in MA kul/kla/waak\l shows that while the principle of imposing conformity with a fixed output by the addition of meaningless segments is a widely employed strategy, there is some room for variation in the quality of the segment and the position it will occupy. 4.2 The Diminutive in Modern South Arabian (Johnstone 1973) Johnstone (1973) represents a masterful application of reconstructive method to the very difficult data of the Modern South Arabian (MSA) languages. Unraveling the highly complex synchronic and historical phonology of the languages, Johnstone first uncovers the basic diminutive patterns of each of the languages, then attempts a reconstruction of the diminutive for Proto-MSA. The basic diminutive patterns for each language, according to Johnstone, are given in (14): (14) the form of the diminutive in MSA (Johnstone 1973) Socotri Mehri/Harsusi S'heri 3C C(w)o'CEC CewEECeeC Ce'CE'C Ca(w)Ce'Chen CeCEECeen Ce'CeCe'n 4c CeCe'CeC CeCEECeeC Ce'CCE'C On the basis of these patterns Johnstone reconstructs a common MSA diminutive *C(e)weeCaaC, which he then tries to connect with diminutive patterns found in classical Semitic languages. The problem which Johnstone uncovers is that while the reconstructed vocalism of the Proto-MSA diminutive can be made to correlate with classical patterns, there is a seemingly irreconcilable gap in the syllabic patterns: "It can be postulated that the vowel of the penultimate syllable was a high, back rounded vowel (u) and that the vowel of the final syllable was a low, front, unrounded vowel (a-). However, the assumption that the earlier pattern was CuCaaC as the comparative Semitic material would then also suggest, is not entirely consistent with the patterns that actually occur.... "Both the S[ocotri] and M[ehri] patterns are better explained if it is postulated that the semi-vowel w, rather than a high back rounded vowel characterized the initial syllable(s) of the earlier base pattern. An earlier pattern C(e)weeCaaC would moreover be consistent with the S'[heri] pattern CeCEC, since w does not occur in intervocalic position in this language" (p.106). Johnstone concludes with the suggestion that the CuCaaC pattern found as an obsolete diminutive in CA and sporadically elsewhere "may be a secondary formation" (p.107). In light of the careful analysis which has preceded, this concluding remark seems somewhat careless; and Johnstone offers no proposal concerning how CuCaaC might have developed from a hypothetical *C(e)weeCaaC. Sound changes of a type to reduce a whole syllable to a short vowel are not normally postulated as part of the development of CA from Proto-Semitic. And why the MSA languages, which have undergone such massive restructuring of their phonologies, should be assumed to be more conservative than CA in this one point, requires more elaboration. The solution I would suggest is that the MSA diminutive patterns have undergone exactly the same type of analogic reshaping that has produced the MA diminutives and for the same reasons: The diminutive of the three-consonant nouns has acquired an extra syllable by addition of a default consonant, thus giving it the same syllabic pattern as the four-consonant nouns. The main difference with MA is that in MSA the default consonant is added between first and second consonant rather than between the second and third. 5. Conclusion and Implications In this paper I have argued that principles of analogy proposed on the basis of patterns of change observed in Indo-European languages are inadequate to account for patterns of change found in languages where non-concatenative, fixed-output morphology plays a prominent role. First I have argued that PROPORTION is not the only operative principle in analogy. Since proportional relationships between words (or consistent change between an input and an output) is not the only kind of regularity found in morphological systems, it is reasonable to suppose that it will not be the only kind of regularity which can be imitated and extended in language change. Some consistent feature of the output logically could be imitated and extended in analogy, and I believe that I have shown that just such a strategy has played a role in shaping the morphology of Moroccan Arabic. Second I have challenged the notion that analogic change simply represents EXTENSION of a pre-existent rule or pattern. In order to account for changes in the history of Arabic, room has to be allowed for rule-creation on the basis of surface patterns, and the consequent possibility of variation in grammar among speakers even where the surface patterns are the same. Having illustrated these principles on the basis of attested patterns of change, I have then tried to show how they can be extended to account for problems which have baffled researchers in the past. The principles discussed in this paper have a number of implications for the historical study of morphology-- most obviously within the field of historical/comparative Semitics. The traditional Semiticist approach to historical morphology is quite static: the consonantal roots and patterns of the traditional grammatical analysis are assumed to be historical primitives. There is no conception of how new roots and new patterns might arise. Yet new patterns and new 'roots' do arise in the history of Moroccan Arabic-- as a result, we have argued, of the operation of sound change and analogy. Output-based analogy combined with sound change-- explains how a consonant can be lost and a new one supplied-- creating an ostensible new root. The notion of rule-competition makes it possible to outline a process whereby a new proportion is derived from an older fixed pattern and whereby a new fixed pattern may then be derived by reinterpretation of a change-consistent process. We have shown that in the course of this process analogy may give rise to allomorphy, and we have argued elsewhere (Ratcliffe, forthcoming), that analogic change is indeed largely responsible for the most extensive and well-known case of allomorphy in Semitic-- the system of noun plural formation in SW Semitic languages. Although we have restricted our discussion to languages in which fixed output-morphology remains a salient feature of the system throughout the time period of the particular change being discussed, the analysis provides a framework which could account for how fixed-output morphology might come to be replaced by change-consistent morphology, or how fixed-output morphology might arise in a system which previously never had it. Changes of the first type are known. In Ratcliffe 1996, we have discussed one such case-- a widespread pattern of change attested in different degrees in several Afroasiatic languages whereby a fixed pattern plural gives way to a change-consistent formation based on reduplication. I wonder if the development of the small set of English verbs whose past tense ends -ought/-aught, though the present stem have various shatpes, and the three past tense modal forms in -ould (would, should, could) may represent an example of the second type-- that is a regularization of the past tense stem according to a fixed pattern. At any rate, research continues to show that fixed-pattern morphology is not nearly as rare outside of Semitic languages as might once have been supposed, and sometimes may exist in a marginal way in languages which do not exploit it productively. The diachronic aspects of such morphology merits further attention. References Anderson, Stephen. R. 1992. A-Morphous Morphology. Cambridge: Cambridge University Press. Bat-el, Outi. 1986. "Extraction in Modern Hebrew Phonology" unpublished M.A. Thesis, UCLA. Bynon, Theodora. 1993. "Analogy" in R.E. Asher and J.M.Y. Simpson, eds., The Encyclopedia of Language and Linguistics, vol. 1, pp. 110-111. Oxford, UK: Pergamon. Harrell, Richard S. 1962. A Short Reference Grammar of Moroccan Arabic. Washington, D.C.: Georgetown University Press. Harrell, Richard S., ed. 1966. A Dictionary of Moroccan Arabic: Arabic-English. Washington, D.C.: Georgetown University Press. Heath, Jeffrey. 1987. Ablaut and Ambiguity: Phonology of a Moroccan Arabic Dialect. Albany, N.Y.: State Univiersity of New York Press. Hock, Hans Heinrich. 1985. Principles of Historical Linguistics. Berlin/ New York: Mouton de Gruyter. Johnstone, T.M. 1973. "Diminutive Patterns in the Modern South Arabian Languages" Journal of Semitic Studies 18. Kiparsky, Paul. 1974. "Remarks on Analogical Change," in John M. Anderson and Charles Jones, eds., Historical Lingusitics II, 257-75. Amsterdam: North Holland. Lieberman, Stephen J. 1991. "The Regularity of Sound Change: A Semitistic Perspective," in Philip Baldi, ed., Patterns of Change, Change of Patterns: Linguistic Change and Reconstruction Methodology. Berlin/New York: Mouton de Gruyter. McCarthy, John J. 1981. "A Prosodic Theory of Non-concatenative Morphology." Linguistic Inquiry 12:373-418. McCarthy, John J. 1979. Formal Problems in Semitic Phonology and Morphology. Doctoral Dissertation, MIT. Moscati, Sabatino, Antoine Spitaler, Edward Ullendorf, and Wolfram von Soden. 1964. An Introduction to the Comparative Grammar of the Semitic Languages. Wiesbaden:Harrassowitz. Ratcliffe, Robert R. (forthcoming). The 'Broken' Plural Problem in Arabic and Semitic: A Study in the Synchronic and Diachronic Aspects of Non-Concatenative Morphology.Amsterdam/Philadelphia: John Benjamins. Ratcliffe, Robert R. 1996. "Drift and Noun Plural Reduplication in Afroasiatic" BSOAS 59/2: 296-311. Ratcliffe, Robert R. 1992. The Broken Plural Problem in Arabic, Semitic and Afroasiatic: A Solution Based on the Diachronic Application of Prosodic Analysis. Doctoral Dissertation, Yale. Wehr, Hans; ed. and trans. J. Milton Cowan. 1960. Dictionary of Modern Written Arabic, 3rd. edition. Ithaca, N.Y.: Spoken Language Services. Wright, William, ed. and trans. 1896. A Grammar of the Arabic Language, 3rd. edition. London: Cambridge University Press. Notes 1) Anderson (1992: 367-8): "A valid analogy is thus always based on a valid and (reasonably) general rule of the language. Where no rule relates two forms... they cannot serve as the base of an analogical proportion. Of course for Bloomfield the notion of a rule was limited to patterns of association among surface forms, and hence the only rules that could serve as bases for analogy were similarly relations among surface forms.... "On these views, the basic mechanism of analogy is the extension of an existing rule to cover new forms. The rule can be instantiated by a proportion among surface forms, but it is the rule itself (not the forms that make up the proportion) that govern the change." 2) Examples are taken from Wehr's (1960) dictionary. Note: (a) verb from verb, (b) verb from noun (base verb has different meaning) cf. jalada "whip", jalida, "freeze", jaluda "be tough;" raxuma "be soft", raxxama (also) "soften" dhahaba "go", (c) verb from noun (no base verb exists).