Redundancy Rules in Phonology Richard Stanley Language, Vol, 43, No. 2, Part 1 (Jun., 1967), 393-436. Stable URL: http fink jstor.orgsici?sici=0007-8507% 28196 705% 2049%3A2%3C9933 ARRIPHAE20.CO@IB29 Language is curently published by Linguistic Society of America Your use of the ISTOR archive indicates your acceptance of ISTOR's Terms and Conditions of Use, available at flip: feworwjtor org/aboutterms.htmal. ISTOR's Terms and Conditions of Use provides, in par, that unless you fave obtained pcior permission, you may not dowaload an cnt isus of @ journal or multiple copies of articles, and you may use content inthe ISTOR archive only for your personal, non-commercial uss. Please contact the publisher cegarding any further use of this work. Publisher contact information may be obtained at bhupsforwe,jstor.org/joumals/Isa itm Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transtnission. ISTOR is an independent not-for-profit organization dedicated to creating and preserving a digital archive of scholarly journals. For more information regarding ISTOR, please contact jstor-info@jstor.org. hup:therww stor orgy Sun Ape 25 08:49:57 2004REDUNDANCY RULES IN PHONOLOGY Rrcnaxo Staxuzy University of California, Berkeley ‘This work contains a discussion of redundancy rulos and the role they play in the phonologiesl component of a generative grammer. Phosological redundancy rules wore Srv given 6 clear theoretical foundation in the woRreemw aeRUCTORE os used by Halle (1086) to prodiet. redundant phonological information in rmorphemes, We will be concerned in what. follows with reformulate xd ex tending this originl theory of morpheme structure rules. We will give particslar sttontion to the formal nature of morpheme structure rues and to the use of Dianks in representing redaedancien, and wil give a solution toa longeetanding problem in this are. In spite ofthis improvement im morpheme structure Fle henry, we wil ind other motivations for inodusiag # new formal device, wom- pomux' ernucruxs coxbinions, to replace morpbeme structure rules. We vill sctompt to show that, both in the types of statements which can be made about the structure of morphemea aad in the formal nature of these statements, mor- pheme stzueture conditions are tobe preferred over morpheme structare rales. 14, Iwrnonverton. The fact that natural languages posiess a high degree of phonological redundaney is well known. Also, it has long been observed that the ‘existence af phonological redundaney is due to the fact that each language ex- hibits systematic constraints on its phoneme sequences go that not all sequences of phonemes form possible morphemes of the language. Morpbeme structure rules capture the intimste eonneetian between the phonological redundanoy of a language and the constraints it exhibits on phoneme sequences in its morphemes. In fact, a morpheme structure rule can be interpreted both as a statement of a ‘constraint on phoneme sequences and 23 an algorithm for predicting redundant feature values in phoneme soquences, The morpheme strueture rule itself is neutral as regards its interpretation. To illustrate, we observe in English that a ‘morpheme-initial nasal eunnot bo immediately followed by » consonant. This observation will be reflected in a morpheme structure rule which cays, in effect, that any segment following a morpheme-initial nasel must be [—Consonantall. Interpreted as the statement of a constraint on possible phoneme sequences in gts, this rule says that no [-+Consonantal] segment can follow a morpheme- initia) nasal. Tnterpreted as a algorithm for predicting redundant feature values, the rule shows that the value of the feature Consonantal may be lett blank in segments immediately following a morpheme-initisl nassl, since this value is predictably ‘—" (i.e, ‘minus’ in this position. Thus, a full set of morphere stroc- ture rules for a lunguago will do two things: it-will state, in terms of features, all constraints on what sequences of phonemes are possible in morphemes, and it ‘will allow each morpheme to have a representation in which redundant feature ‘values are omitted. ‘The above-mentioned rule can be ealled @ SEQUENCE eTRUCTURE RULE: it makes a statement about possihle sequenees of phonemes. We ean also make statements about single phonemes, regardless of their context. A rule giving such a statement eam be ealled 3 sz0xenT STRUCTURE RULE: it makes a statement 308894 LANGUAGE, VOLUME 48, NUMBER 2 (1007) about the feature composition of individual phonemes. To illustrate this lator type of rule we can say that, in English, each [+Nasal] segment is aleo (+ Voioed]. Like sequence structute rules, segment structure rules are, in themselves, neutral as regards their interpretation. Thus, the above segment structure rule ean be interpreted not only as the statement of a constraint on the feature composition of phonemes, saying that no (+Nasal] segmant may be {—Voieed], but also as an algorithm for predicting redundant feature valves, showing that the value of the feature Voiced may be left blank in [4+-Nasall segmonts, since this valuo is pre- dictably ‘++. A full set of scement structure rules for a language will suffice to predict all the nondistinetive feature values from the feature values which are Jistinetive. Segment structure rules bave generally been put in the phonological rules and not in the morpheme structure rules. However, we will find reason to regard all redundancy rules, whether they involve predictions within a, single segment (segment structure rules) or in a sequence of segments (sequence struc- ture rules), as being included in the morpheme structure rules. We will return to this question in detail below. Yt has been eommon for linguists to view phonological redundancy asa statis- tical entity to be treated in the way redundaney is treated in information theory (eg. Carroll 1961:201-5; Cherry 1961294, 115-20, 180-7, 272-3; Gleason 19612 873-00; Hockett 1958:8i-91; Miller and Chomsky 1963:439-43). The redun- dancy of a language is then given as a percentage, which is high if the Language is highly redundant and low if the lauguage has little redundaney. However, such ‘a percentage doce not give any information about the phonological structure of the language. It merely represents an average, statistical value of the redundancy for the whole language, and does not tell us specifically what: local constraints in phoneme sequences contribute to this redundancy. Sinee these local constraints, are of primary interest in the study of the phonological redundancy of natural Ianguage, itis clear that, we need a more detailed statement of redundancy than the statistical one. We need a statement which concentrates on the specific, ature of these local constraints, and not merely on the over-all effect that these ‘constraints have on the efficiency of & language as o cormunieation system (as this efficiency is given in astatistical measure of redundaney). Such a statement is Provided by the morpheme structure rales, 1.2. Tax Wo LevELs of REPRESENTATION. Tn a generative phonology of a language there are to levels of representation: the sverumamte PRONEMIC level and the evareuamic Puoxertc level (Chomsky 1964:68). The phonologieal com- ponent of a grammar consista of a set of PHONOLOGICAL RULES (called P rules) ‘which map representations of sentences at the systematic phonemic level onto the representations at the systematic phonetic level, The morpheme structure (M8) rales, on the other and, do not map one level of representation onto another, but rather state the redundancies thatexist at single level, thesystera- atic phonemic level. ‘Each morpheme is represented at the systematic phonamnic level ay a sxsTEN- atic Paovetc warRcx.' Each row of this matrix corresponds to a distinetive The sceumption ig that each morpheme has x atwoun underlying representation, its systematic phonemic matrls; the P rules derive the various allomorphie forms of exchREDUNDANCY RULES IN PHONOLOGY 995 feature of the language, and each column corresponds to a segment (systematic phoneme) of the morpheme. Each entry in this matrix is either ‘+ ot ‘—’. Thus, in particular, there are no blank of ‘0' entries. The systematic phonemie repre sentation of a sentence consists of a string of systematic phonemic matrices (one for each morpheme of the sentence) togethet with 4 labeled bracketing which is, the surface structure of this string (Chomsky 1964:64). The P rules apply to the systematic phonemic representation in eyelicsl fashion; the entire set of P rules is, ‘applied, in the proper order, to each etring of matrices encloved by the innermost bracket sets of the surface structure, then the innermost brackets are erasod and the process is repeated; this ia kept up watil there are no more brackets, and the string which results is the systematie phonetic representation (Chomsky 1964: 12; Chomsky and Miller 1963:313-10) 2 We are not concerned here with the P rules, but rather with the systematic phonemic matrices which form theit input. As noted above, these matrices are fully specified: they have no blank entries, This full specification is desirable, as ‘wo will see in §23 below, since the P rules are most easily and naturally defined as operating on fully specified matrices, However, we would not want to enter a morpheme in the picmonane ag its fully specified systematic matri, for, clearly, this matrix contains much that is redundant. Thus we will enter each morpheme the dictionary 2 a orovrowary Manx which is identical with the systematic phonemic matrix of the morpheme except that all redundant feature values are left blank (or, equivalently, ate replaced by the symbol ‘0’) The MS rules, con- sisting of sequence structure rules and segment structure rules, will map eaeh in- completely specified dictionary matrix onto the corresponding fully specified. systematic phonemie matrix. It is important to notice that the MS rules, as bere conceived, are exclusively redundaney rules, and that they represent redundancies at a single level, the systematic phonemic level. Thatis, dictionary matrices and systematie phonemic matrices are representations of the same level in the sense that each dictionary matrix ia simply a less fully specified, redundaney-free version of the correspond ing syetematio phonemie matrix. On the other hand, it would be totally wrong to ‘view systematic phonemic representations as redundaney-free versions of the corresponding systetnatic phonetie representations. The P rules, which map the ‘hair respective environments from thi systematic phonemic matrix. ‘This is, of course, the usual one of generative phoaclogy. It must be modified in the obvious way in the ease of auppletion. * Actually, in the works cited, not all of the P rules are regarded ae being cyelle; rather there is a set of precyetien! rules, a set of cyclical rules, and a set of posteyctieal rules. ‘The presyelical rules are just the morpheme structure rule, tha eyetieal rules denoribe the ‘main phonological processes, and the postcyelieal rules are low level phonetic rales, For us, however, the morpbemo strueture rules sro no: regarded as forming part of the P rulea Dbut are statement which form part of the lexicons thus there are just two kinds of P rules, eyolical and posteyeical. This, cogstier with the fact chat for us the input to the (eyelleal) P rules consists of fully sper‘ied matrices, constitutes she main difference between our position and that of the works cited. Our position and the justification for it will be flly igeuased below, * Chomeley (1064:86) hae uged the term CLAserrICANORY KATAIE ia the aamie aease that wwe ube the term dictionary matrie.396 LANGUAGE, VOLUME 43, NUMBER 2 (1967) former onta tbe latter, ate ia no sense redunaney rules; their role is to change Feature values, not to fill in blanke feature values. ‘The history of the ideas waich led up to the position stated in the previous paragraph is quite interesting, and an outline of its main points may be helpful. Ti seems fair to say that, in the descriptivist tradition, phonemic representations ‘were canceived of a8 containing just what was not redundant in phonetic repre- seniations. There were many attompts to clarify this eonception and to aive it s precise theoretical framework. One such attempt was the original distinctive feature theory of Jakobson, Pent, and Halle (1951) and of Jakobson and Halle (1956). A careful reading of these works shows that the phoneme representations, as there conceived, were simply less fully specified, redundancy-free versions of the phonetie representations, for to pass from the former to the latter one needed only to fill in blank feature values, never to change feature values, In fact, a Phoneme was exactly what was non-redundant (invariant or distinetive) in all its phonetic resligations, ‘Thus phonetie and phonemic representations were, in a significant sense, representations of the same level, and differed only in whether ‘or not the redundant information of that level ‘was explicitly indicated (see Chomsky 1957 for an enlightening discussion of these and related issues). ‘The discovery made by Haile (1959) was that the above conception of phonol- ‘ogy was not adequate. Specifically, he showed that in order to characterize many ‘of the generalizations that ean be made about the phonological structure of & language, # level of representation was needed which was far more abstract than, the phonemic representations referred to sbove. The representations of the level proposed by Halle were in terms of units which he ealled incompletely specified morphonemes, and the level itself can be called morphonemic to distinguish it from the earlier phonemic representations, Morphonemic representations, like ‘phonemic representations, left all redundant information blank. However, unlike phonemic representations, morghonemic representations were nat merely redun- daney-free versions of phonetic representations, for the rules whieh mapped the former onto the latter could involve changing of feature values (including in- certions, deletions, permutations of segments, ete.) as well as filing in of redun~ dant feature values. Tn fact, Halle found that no representations related to ‘phonetic representations by statements of redundancy, ot by any ‘biunique’ statements, played a role in an adequate grammar, "The rules, as there organized, consisted of an ordered et of MS rules followed, bby an ordered set of P rules, In some sense the MS rules wero viewed as redun- | +Continuant + Voiced ~Sharped ‘The segment structure rules (2) and (3) can easily be extended to a full set of segment rules for Russian. As we saw in §1.1, such a set of rules can be interpreted oth as an algorithm for predicting redundant feature values and 98 a set of ‘statements of the constraints on the feature composition of eystematie phonemes. In this case, if we interpret. the set of rules as a predictive algorithm, then we ean «give all Russian eystematio phonemes in a reduced form (as in 1b above and in Halle 1959:45, Table I-3).* Dictionary matrices will always bave systematic phonemes with at least as many blanks as these redueed forms, siace the segment structure rules suffice to prediet the fully specified systematic phonemes from these reduced forms independently of the context: ‘Note that the reduced form of o given in tb (ond in Table 18 of Halle) is marked {—Diffuel, even though thio is predictable from the faot that. all [4-Compact) segments are [-Difuse), This is unavoidable if there ia to be s branching diagram of, ‘equivalently, if ‘distinctnens i roquired. We return to thess questions below.0 LANGUAGE, VOLUME 43, NUMBER 2 (1957) i, on the other hand, we interpret the set of segment structure rules stating, constraints on the feature composition of systematie phonemes, thon this set defines exactiy the set of Russian systematic phonemes. In fact, the set of Russian systematic phonemes is the unique maximum set of fully specified eleven-tow, one-column mtrices that obey all the constrainis given in the set of segment séeucture rules, For if some such fully specified matrix which was not one of the systematic phonemes of Russian obeyed all the constraints given in these rules, then the rules would simply nat be specific enough. They could (and should) be extended to rule out this matrix, since this would effect greater savings in the dictionary representations of the actual systematic phonemes. ‘Thus the segment structure rules, though included in the grammar to characterize redundancy, provide as a hy-prodiict the definition of the set of aystematic phonemes; this obviates the need for an ad hoe definition of this set by means of a lst or 4 sepa rate set of statements of come sort. This is an important point. Tt provides a certain amount of Justification for deviating from past practice and including a set of segment structure rules in the MS rules (we will give further justification in $1.6) Notice that, since systematic phonetnic matrices are fully specified, we obvi- cusly cannot gain information about. what is distinctive in a Language by looking at these matrices. For example, if in a systematie phonemic matrix we see a [Nasal] segment which ig also specified [—Compact], we cannot infer that com- ‘pactness is distinctive for nasals and thus that there are [+Compact] nastls as ‘well. Tt is the presence or absence of a segment. structure rule [+Nasal] — [Compact] that gives us the information shout whether or not compactness is distinctive for nasal, Further, it seems natural to find such information in a set of rules, rather that having to search through the morphemes of the language for it. Finally, notiee that we cannot infer from the sheence, say, of a specification for the feature Compact in a nasal segment in a ntotionany matrix that compact ‘ess ‘s not distinctive for nasals in the language, but only that itis not distinctive in that particular environment. 14, SnqvENcE srucruse RULES, Segment structure rules give information ‘shout one ind of phonological redundaney, that of systematie phonemes in isolation. When systematic phonemes are combined into sequences to form mor- phemes, a new kind of redundancy arises, a redundancy due to sequential con- straints. For example, in Russian & morpheme-initial glide must be followed by a vowel (Halle, 1969:58, rule MSla). ‘Thus we can give the sequence structure mule? _, [+Yowatie Vo wo L ] [eas | i+ (ee nat) - Rule (4) ean be extended to a full set of sequence structure rules for Russian, at Holle does in his presentation of the MS rules (1069:58-61). (Por Halle, at (MB rules are sequence structure rules sinee he docs not give any segment strue- “The “4 which stands outide the brackets in (4) stands for morpheme boundary ig wed in Halles work), This +" ia no! to be coptced with the “of the feature values “andREDUNDANCY RULES IN PHONOLOGY 41 ture rules.) If we interpret this set of rules as a predictive algorithm, then, in dictionary matrices, roany segments can be reptesented with even fewer feature specifications than the reduced forms discussed in the preceding section. For ‘example, a segment following a morpheme-initial glide need not be specified for the features Consonantal and Vocal, since these features are predicted by rule (4). Thus segment structure rules make the predictions in systematie phonemes ‘that aro possible without considering the context, and sequence structure rules ‘make the further predictions which are possible when contextual restrictions are considered. Ts, on the other band, we interpret. the set of sequence structure rules asa state- ‘ment of constraints on systematic phoneme sequences, then this set provides a characterization of the set of ‘passible’ morphemes of Russian. This fact is clear if, as is quite natural, we regard a ‘possible? morpheme as a form which may or may not occur in the lexicon, but whose phonological structure violates no sequence structure rule (and, of course, uo segment structure rule) of the lan- guage! Thus the sequence structure rules, though included in the grammar to characterise redundancy, provide as a by-product a characterization af the notion ‘possible morphemne’; this obviates tho need fora separate set of statements to characterize this notion (ef, Halle 1964b:341-2). The justification for the use of sequence structure rules given here is similar to that given for the use af seg ‘ment, structure rules in the preceding section. The reason that the latter kind of justification has not previously heen considered sufficient eause to put the seq- ent structure rules in the MS rules will bs diseussed below. 1.5, Nevrnaiszartox. Our decision to include ata, the phonological redundancy zules in 4 single set, the MS rules, may at first glance appear incorrect, for there are certaia kinds of rules which rust apply in the P rules but whieh, at least in some of the environments where they apply, function as redundanoy rules, Rules of ‘neutralization’ have this property. A feature is said to be ‘neutralized* io an environment if the value of the featare in this environment is determined by 3 sequential constraint. It is clear that if this environment ie limited to sequences which ate contained within single morpheme and which cannot extend across morpheme boundaries, then the sequential constraint -will be stated as a sequence structure rule and will be included in the set of MS rules. Generally, however, neutralization involves sequential constraints which hold across morpheme boundaries as well as within morphemnes. In this case, any rule whieh reflects the neutralization not only makes a statement about the phonological structure of smorphomes, as do MS rules, but also, ike P rules, describes how this phonological structure is altered when mozphemes aro combined in sequence. The obvious place * Ofcourse, to obtain the desired results, we must guarantee that exch aequence structure rule reflects general aystemstie fact about the language, and not a fact which is doe merely the existence of uocidental gapa in the lexicon, Thi gusrantee ia provided by the erierion ‘of simplicity, which givaa 4 well-motivated means of separating the sesidoatat sequential ‘constraints from the aystenatic ones; according to this eriterian only the aystematie nea ‘can be stated in the sequence structure rules (Halle 1994b:340-2), This means thst the soquenre structure rules provide s mesg of distinguishing azeidental gaps in the lexicon from gape whose existence is due to & general fat about the language.402 LANGUAGE, VOLUME 43, NUMBER 2 (1967) to put euch a rule, and the place where it has heen put in past practice, is in the P mules, since there it can not only predict: the neutralized features when the neutralizing environment: is contained within a single morpheme, but can also change features to their neutralized value when they are neutralized in environ- ments which include surrounding morphemes. We will argue, however, that the proper place for such a ruleis in the MS rules. Before showing how this is possible, Jet us give a concrete example of a rule of neutralization. Consider a language which has both long and short vowels at the systematic phonemic level, but in which the feature Long is neutralized before two cot- sovants (regardless of whether the VOC sequence is divided among one, two, oF three morphemes), 50 that only [~Long] vowels appear phonetically in this position, ‘The rule of neutralization whieh reflects this constraint is (5) V — [—Long] / CC Such a rule is a redundancy rule when the VCC sequence is contained within a single tporpheme, for in this case the vowel is always in the shortening enviroa- ment aod thus need not be marked for length in the dictionary. Since we have decided to include all redundancy rules in the MS rules (eo that the output of the MS rules can be fully specified), itis clear that (5) must be an MS rule. However, we are assuming that (5) also applies across morpheme boundaries, having, for example, the effect +CVC+C+ > +C¥C-+C+. Thus (5) must apply in the P rules s well. ‘We could just leave the discussion at this point, saying thet rules of noutrali- zation, such s (5), have a double function, appearing hotb in the MS rules and in the P rules, Moreover, this repetition of rules is not necessarily undesirable; it merely represents the kind of situation, quite typical in natural languages, jn which certain constraints happen io typify both eequences within single morpheme and sequences which extend across morpheme houndaries. Compare this to the situation in which constraints applying within morphemes are totally different from thase applying across morpheme boundaries; we could make stich ‘a situation less highly valued in grammars by agreeing to attach less cost. to a P rule which is identical with come MS rule than to a P rule which imposes come entirely new constraint. This is a noseible and feasible proposal. However, it may be possible to state a rule such. as (8) just onee, in the MS, rules, and to say that by convention it retains ite effect in the P rales. Tn this case, such a convention would say that at any point during the operation of the P rules where a vowel appears before two consonants, the vowel is automatically made short. This convention would work only if the order af (8), in its original statement in the P rules, was not crucial, We will leave the discussion of neu- ‘reliantion in this ineonclusive state and go on, in the next section, to discuss & mruch clearer case where certain redundancy rules (the segment structure rules) ‘must be stated in the MS rules but still retain their effect in the P rules. 1.6. Tas rosimon ov ae szcatext steceroun nuts. In this paper we have adopted the practiee of putting the segment structute tules in the MS rules, 20 that all the redundaney rules of the language appear together in one set. ‘The fact that the segment structure rules appear in tho MS rules has two asinREDUNDANCY RULES EX PHONOLOGY 103 advantages. First, they provide a definition of the set of systematic phonemes,” ‘aad, second, they fil in all blanks in matrices before the P rules begin. to apply. ‘There is, however, sn argument in favor of putting the segment structure rules in the P rules, and it is this argument that has won out in the past. The argument is that there are redundancies in segments that arise during the operation of the P rules, and that we ean sceount for these redundancies by having the segment structure rules appear in the P rufes. Tat us consider an example. Suppose that some language has non-compsct: vowels i ¢ 04 but not 1 8 6. This restriction can be reflected in the segment structure rule: '—Consonantal (6) Compact. | -> [aRounded! a Grave Now it is likely that any epenthetic non-compact vowel will also obey the restriction stated in (8); suppose that this is the case. ‘Then we would want rule (8) to roszow any P rule that siates such an epenthesie, so that (6) could apply also to the epenthetic segments. For if (8) were stated in the MS rules, or in the P niles before the rules of epenthesis, then we would have to specify the ey thetie segments for the redundant feature Rounded and thus miss a general zation. Tt was thus argued that we rust let the segment structure rule (6) ap- pear in the P roles; in particular it must follow rules which state the epenthesis ‘of non-compact vowels. To give another example, consider a fanguage whose non-compact stops ate non-strideat and whose non-compact continuants are strident.!* This restriction could be stated in the segment structure tule: + Consonantal Compact | > faStrident] Continuant, If this language has @ P rule whereby non-Compact eontinuants become stops in some enviroument, then we would like to atate this rule as (8): (8) [enact] -r!-Conon / ere 47) But this is only possible if tho segmont structure rule (7) follows the P rule (8). "Thin is discussed in §1.8. It i clear that if we adopt the uncal practice of letting the seg ‘ment structure rules be scattered through the P rules, each rule not appearing antl iia needed, then this means of dofining the set of systemstic phonemes is not available, For ia thio case the segment séructure rules no longer formally represent reduadancies and con stenints the systematic phonemic level. There could be no mativation for defining unite ‘Che aystemstic phonemes) of that level in terma of thie geatlered eet of rules. agit Nivtvae of thn ns en mentioned ay above ad il be ivused fly t 8. ‘Such a language would have the non-compact conzonsnte ¢d# ¢, but nat ¢ ar 5, whlch sre [~Continusat, +-Strident) and thus violate (7). This language could still have 3 and J, ‘which, though also [—Continusnt, +Staident), ore (+ Compact)404 LANGUAGE, VOLUME 49, NUMBER 2 (1997) If (7) had been stated in the MS rules, or had preceded (8) in the P rules, then ‘we would have to make (8) more complicated, stating it a¢: @ [orem _, [~Continuant) , (i the appropriate — Compact Strident environment) ‘This means, in effect, that we have had to state the restriotion on the features, Continuant and Strident in non-compact obstruents twice, once in (7) and once in (Q), and thus again we miss a generalization, To avoid this, it was argued, ‘we must Jet the segment structure rule (7) appear in the P rules, whero it must, ‘moreover, follow rules such as (8). {it seams that exsaaples such as the above two show that the sorment structure rules must appear in the P rules, and ths thet they cannot be put in the MS rules, There is, however, a serious conflict involved in this practice of putting segment structure rules in the P rules in order to state redundancies in segments which arise there, According to such a practice, a segment structure rule such tas (7) must rorsow any P rule which, like rule (8), changes a non-compact continuant to a stop (or any rule whieh changes a stop to a continuant), or which {introduces an epenthetic non-compact stop oF contintiant. But it's also necessary that (7) preceoe any P rule ‘R’ thet needs information about the feature Stri- dent in non-compact consonants, since the stridency is inserted only by (7) and since insurmountable problems arise if blanks oceur in matrices during the P rales (cf. §33). Yet these two conditions ean both be met only if rules like (8) always provede rules like ‘R?, and clearly there is no way to guarantee, or even. any reason to expect, that this will be the case, The conflict involved ean be suamarized as follows. Yor some purposes we want the segment structure rules to state their redundancies mazsx in the P rules, so that these redundant features can be used by other P rules. But for other purposes we want the seg- ‘ment structure rules to state their redundancies tare in the P rules eo that they an state the redundancies in segments introduced or changed by the P rules. Because of this confict, neither the usual practice of letting the segment strue- ‘ture rules he seattered through the P rules, nor the practice we advocated earlier of letting the segment structure rules appear only in the MS rules, can be a¢~ cepted as it stands, ‘Phore is & way, however, of keeping the segment structure rules in the MS rules while still allowing them ta state redundancies in the P niles. We merely need to adopt the convention that the output of each P rule is automatically subjected to the segment structure rules Such @ convention would imaply that ‘There are various ways in whish such a convention could be male preiae. One way is the folowing. Any P rile of the forta X—* [Fs a, oF |/¥—L (bors X, ¥, and 2 are bundles of fostunes, nd esch 9 a either “+1 or“) i interpreted se applying in 5 etepa. First, the feature Fm assigned the value fin any arament which hes the features in X; then any segient strusture rule which haa Fy 00 the left ofthe arrow is applied, # ap= plleable, Them the feature Fis assigned the vale inthe output ofthe rt step, and 28 rent etruccure rules invelving oP are applied This is kept up until the feature fe is as- signed ‘o Fe and seqracas structure run iavelving #4Fs are applied, Acnording to thie tauvention, che order in which the feazures age suated in the rightaide of a P rule may affect ‘the ouput of the P rile, a fact which may be of some sigaificanes.REDUNDANCY RULES IN PHONOLOGY 405; segment structure rules represent constraints that, hold throughout the P rules as well as at the systematie phonemic level; the above examples, being typical of natural language situations, show that this is just. wha we find ‘We will assume such ¢ convention in what follows. We thus have all redandaney niles in one place in the grammar (in the MS rules), 50 that the input to the rules consists of fully specified matrices. Moreover, some of these redumdaney rules (the segment structure rules) have their domain of influence extended to ‘the output of (sore) P rules. The examples in this seetion show the inadequacy of theories without this convention. Yet it is, at this point, not at all clear how the details of such a convention are ta be worked out. An extremely interesting ‘and fruitful line of research would be to obtain data bearing on this question from a variety of languages. FURTHER DETAILS 2.1, PHoneme HEDUNDANCY ROLES. We have stressed that MS niles repre- sent redundancies at the systematic phonemic level; they say nothing about redundancies st the systematic pbonetic level. In fact, the theory has no re- dundancy rales at the systematic phonetic level, rules which correspond to the (MS rules at the evstematic phonemic level (but cf. the diseassion in §1.2). More- over, the complexity and depth of ordering which in general cxists in the P rules shows that it may be dificult to construct any natural set of phonetic redundancy rules, even in principle. This is becauso phonetic representations are only ob- ed as the output of this highly structured set of rules (the P rules), 80 that in general a relationship among the elements of these representations eannot be simply stated in terms of those elements alone, but must be referred hack to the systematic phonemic representations which underkie them, ‘As an example, suppose that in a language with vowels ¢e a 0 u, back vowels are fronted but not unrounded in the environment CY when the V is ¢ or i. Tf a later P rule drops thee or the é in some environment, then @ 6 and é will, contrast with the five underlying vowels at the systematic phonetic level, but in a rather special enviroment which ia difficult, perhaps imposible, to char- sterize ab the aystematic phonetic level—namely before a C which at the system- atic phonemic level was followed by an ¢ or ane, Thus the distributional facts cannot be stated in a natural way at the systematic phonetic level. In general there will be mote ot less simple distributions of elemente at the systematic phonemic level (given in the MS rules), and more or less simple statements 4 Te acemn certain tha: we do not want al} P rules, pecially the Inte ane, to have theie ‘output subjected ia the segment structure rules. Tt i atll an open question how we are to identity, ima non-ad-hoe manner, just which P rulee ere 9 subjected. “Nove thatthe argument that the segment struetre rules should be put i the P rules, snd chore as late as possible, onthe grounds that then derivations would became ‘mare eco- nomical” since redurdantfearure values would be carried slong for shore tie, would be meaninglems. More economical gratamars are ta be pre‘erred only wien the kind of economy fchieved ia Diat dened by an evaluation measure, and x perticular evaluation measure i= choses anly becauge of «relatively higher degree of explanatory power schieved in gram- ‘iors evaluated highly by this measure. Clearly, delaying segment structure rules merely to abtain 2 more economical derivation explains nothing406 LANGUAGE, VOLUME 43, NUMBER 2 aver) (in the P rules) which map representations at this level onto representations at the systematic phonetic Jevel, but any direct statements of distribution of clements ai the systematic phonetic level are likely to be complex and to in- volve loss of generalizations, 2.2. Orpen. Iv is possible, in fact usual, to present the MS rules as an ordered set. Rule B follows rule A in this ordering just in case B refers to fetture values inserted by A; in patticulat, this means that all segment structure rules follow all sequence structure sules, However, since MS rules ate redundaney rules, which only insert feature values and never cbange them, this ordering is of a rather trivial sort. 4 non-trivial kind of ordesing can only be obtained in rules which, like P rules, ean change feature values; if a P rule R. changes the value of & feature f, different results will obtain if a rule affeoting feature f is placed before R. from those occurring if it is placed after R. Moreover, these offects of ordering of P rules are quite necessary, so that this ordering makes a significant claim about the nature of phonological structure in natural languages (cf, Halle 1968:338, 343-4, Chomsky 1964:70-5). However, if a set of rules, such as the MS rules, never change feature values, then different, orderings will or give different results. In fact, it should be clear thet the MS rules ean be presented as an unordered set with the genoral instruction “keep applying rulea over and over and in any order until all blanks are fled”. Further, they actually ‘ust be so presented, since to make the (stronger) claim that, they are ordered gives the false impression that this order plays some role (as it does in the P rules). We will see in §4 that MS conditions, which we introduce to replace MS rules, provide a natural formalism for stating just what the above discussion shows that it is necessary to state; ic. a set of tmordered statements about the structure of morphemes. 2.3, InseRT10N AND DELETION RULES. Notice that we have not allowed MS rules of the forms > x/A_B ot x — §/A__B; ie, the formalism, as we have interpreted it, does not permit us to use insertion rules or deletion rules as MS rules. This is consequence of our requirement that eacb dictionary matric be ‘exactly like the corresponding systematic phonemic matrix except that redundant features in the latter aro left blank in the former; in particular, the two matrices have the same number of colomans. ‘This requirement, is made simply because we know of no cases whete regularities of morpheme structure can be stated ‘best in terms of insertion and/or deletion rules.* Of course there may be inser- tion and deletion rules which operate in a grammar, but in most cases it is clear that these are P rules, since they must follow other rules known to be P rules. Tt hes sometimes been suggested that insertion rules in the MS rules eauld be used to make specious savings of feature specifications in the dictionary, and ‘that this constitutes a problem for the theory. For example, suppose we have & language with no consonant, clusters. If the most. common vow in this language Tn fact, the theory af morpheme structure advanced in {4 below is incapable of even tating insertion or deletion rules which are to apply soveny in invea-morphemie environ- ‘mente, ‘Thuis, this theory reproronta a claim that euch insertion or deletion rulea cannot fecur in natural langdages. Nat only does this ceoms intuitivaly correct, but ithas not bees ‘refuted by any oases which havo eome to aur attentionREDUNDANGY RULES IN PHONOLOGY 407 is ¢, then we could leave ¢ aut of the dictionary wherever it occurs between ‘two consonants, and have an MS rule + ¢/C_O. This, it is claimed, will save features in the dictionary, even though this saving is wholly unnatural and undesirable, However, thia claim is correct only if the MS rules are ordered and hhad the insertion rule as the first rule, for otherwise the other MS rules could not make use of the legitimate generalization that no consonant clusters o¢cur. But we stw reason in §22 to require that the MS rules be unotdered. Thus the undesirable insettion tule is not allowed. More generally, the requirement that the MS rules be unordered is sufficient to disallow any insertion or deletion rules. 24, DisttNorion mErweeN MS nuLES AND P nutes. In our presentation of the formalism involved in a. theory of generative phonology, we have made very clear the distinction hetween MS rules and F rules. The former are exclusively redundancy rules and are totally different, in their function and their position jn the grammsr, from the latter, whose function is to derive the phonetic repre- sentations from the systematic phonemic ones. However, it shotild be noted that this distinction has not always been so sharyly made. In fact, when MS rules were first introduced, they were viewed as being very neétly on a par with the P rules; there was one set of rules in the phonological component, the early rules of which were MS rules and the leter of which were P rules. Further, there are many cases in the literature where one is at a oas to decide whether a rule is intended to be an MS rule or a. P rule. In this paper, however, we have ‘argued that MS rules are quite distinct from P rules, both in their logical form and in their linguistie function, and that it is hoth necessary and desirable to maintain this distinetion. In fact, phonological descriptions would be much more valuable if they contained a set of MS rules clearly distinguished from the P rules. ‘We will show in §4 that the distinction between MS rules and P rules is s0 great that they are even best handled by different format devices, instead of both being handled by ‘rules’. P rules will remain rules, but we will introduce MS ‘conditions’ to replace MS rales. 2.6. Brancitna rsgzams. Halle (1959282, §1.53, and 46, fig. T-1) uses a BRANCHING DIAGRAM to represent the seaments of Russian, In bis terminology, these segments are called ‘ully specified mornhonemes'; for us they correspond to systematic phonemes in which the feature values predicted by the segment structure rules are left blank. Bach descending path from the top node to a terminal node in a branching diagram represents the distinctive (non-redundant) feature values of « systematic phoneme of the language. However, itis important to notice that not all ways of choosing the non-redundant feature values leave open the possibility of constructing a branching diagram (Halle 1950:35). A. brancbing diagram can only be constructed if there is a feature f; which is nou- redundant in every segment; if there is a featare f, which is nop-redundant in all +f seaments and a feature fg which is non-redundant in all fy segments; if there are features fy, ff, and f; which are non-redundant. in all -His, —fs Hf, and —f, segments respectively, ete. We will see below thet there are situ- ations in natural languages which do not exhibit this kind of distribution of408, LANGUAGE, VOLUME 43, NUMBER 2 (1967) non-redundant features, In general, however, generative grammars of langriages hhave been chosen go that there 18 branching diagram for the systematic pho- nemes. There are several reasons that branching diagrams have been regarded ‘as important: (1) Giving segments in a branching diagram appears to be the ‘most direet meane, involving the feweet feature speeifeations, which will gusran- tee that each pair of segments is pretivcr in the sense that for any pair of diferent: segments there is at least one feature f such that one member of the pair is specified -Hf and the other —f. (2) The branching diagram gives a hierarchy of features which can be interpreted os meaning that the features at high nodes (auch as Consonantal) are in some sense more basic than the features at low nodes (such as Voiced). (3) The branching diagram gives 2 way of formalizing ‘the notion of archiphoneme. It is for these reasons that branching diagrams have been given in grammars. However, we will show that each of them is open to question. Representing segments in @ branching diagram may be quite natural if it is assumed that segments are to be distinct. However, distinctness is only one of the possible formél requitements for guaranteeing that the segments aro all Kept apart, ie, are all distinguishable. In fact. we will give independent evidence (5§8.5, 36) in favor of adopting eriterion of distinguishability of segments which is weaker than distinctness, If this evidence is accented, then the notion of branch ing diagram must be abandoned, for non-distinet. segments cannot be represented in such a diagram, ‘The second point. is more important, sinee there is obviously some kind of hierarchical relationship among the features which must somehow be captured in the thoory. However, an attempt to eaptare this hierarchy in a branching diagram serms somewhat strange in light of the fact that there may be con- siderable freedom in the way this branching diagram is constructed far a given set of systematic phonemes; a different choice of redundant feature values in this set will leed to a different branching diagram and thus to a different hier- archy of features, Obviously we would choose, if possible, that branching diagram. which gives the hierarchy we feel is right; but this just means that we know beforchand what hierarchy we want and are simply choosing to represent it in branching diagram, In. this case, though, we might just as well describe the hierarchy separately since it. has nothing essential to do with the branching diagram. ‘What we want, of course, is to find some characteristics of features and their interrelations that force us, on independent grounds, to assume a PARTICULAR hierarchy of features. That is, we want to find some definite formal property of featares, perhaps stated in terms of the different ways in which different fea- ‘tures behave in the P rutes or in the MS rules, whieh points in an unambiguous way ta the existence of a specific hierarchy. Only in this way ean a feature hierarchy come to represent, more than a vainio set of intuitions about what features are more basie than others, and thus only in this way ean we discover a yeal nxckssiTy for incorporsting the notion of feature hieracchy in the gram- Some of the ideas in the above paragraphs were suggested to me by Chin-W. Kim.REDUNDANCY RULES IN PHONOLOGY 409 ‘What such a formal property might be is an open question at this point. We have seen, however, that the relation of redundant to non-redundant feature specifications in segments, which relationship can sometimes be stated in 4 branching diagram, ig 2 formal property of features that is Nor, in itself, well enough determined on independent grounds to justify its use in characterising a hierarehy. Further, the fact that high-level features can sometimes be predicted in terms of low-level fostures makes it seem thst the relation of redundant to non-redundant features cannot be expressed at all in terms of branching, diagram, ‘To illustrate this last point, consider the following example. Tp many languages Which use the feature Strident, the only [-}Strident] segments are true eonso- nants. This means we can leave the features Coneonantal and Vocalie blank in, [+Sicident] segments and predict them by a segment atructure rule . —Voealic (10) Ue Stedentl = [eens] Before the application of such a rule, of course, the {+Sizident] segments may not be distinct from vowels, liquids, or glides which happen to have all other features in common with them (unless we specify all vowel, liquids, and glides 5 [—Strident] which, early, we don't want to do). But this lack of distinetness is no problem: there are still enough features specified to keep all segmenta distinguishable (for further discuscion, see §3.5). ‘The third reason for having branebing diagrams, mentioned at the beginning of this section, was that it was assumed that they formalize the notion of arehi- phoneme. This assumption, however, is false. The original idea was that just ‘8 the systematic phonemes (the ‘fully specified morphonemes? for Halle) are represented at the bottom, terminsl nodes of the diagram, so archiphonemes: are represented at higher, non-terminal nodes. For example, the nodes labelled 10 (the nodes for the feature Sharped) by Halle (1959:46) represent archi- phonemes of consonants dros which tbe feature Sharped has been extoacte, eg, the archiphoneme /t t,/, Similarly the nodes labelled 9 (Voiced) represent srehipbonemes of consonants from which both the features Shared and Valeed have been extracted, e.g. the arehiphoneme /t t, d d,/. However, it ig easy to see that the branching diagram wrongly limits what the possible archiphonemes ure. For example, letting the feature Sharped label the bottom nodes excludes the possibility of having archiphonemes with just the feature Voiced extracted, Yet obviously the notion of archiphoneme apples as well to those segments jrora which the festre Voiced baa been extracted as to those from which the feature Sharped has been extracted, Thus branching diagrams fail to capture the notion of archiphoneme because of a quite basic and unavoidable fact of their structure. ‘THE ZORMAL APPARATIS 3.1. Inprorsr USE oF BLANKS. We have made three assumptions about MS rules which differ from those made in past works. We have sasumed that all redundancy rules are included in the MS rules so that their output, which is the#10 LANGUAGE, VOLUME 44, NUMBER 2 (1967) input to the P roles, is fully specified (ef. §2.2); we have assumed that MS rules are exclusively redundancy rules, and never change feature values (ef. §1.2); finally, we have assumed that the MS rules are to be presented as an unordered set (ef. §2.2). Arguments in favor of sdopting these assumptions have been given in the sections referred to. In the present section, which is a discussion of the formal apparstus involved in rules and their spplication, we will show thai insurmountable formal problems érise if these assumptions are Not made; i {or purposes of illustration, we will abandon these assumptions, and will diseuss ‘the phonoiogical component 28 it is usually presented. Specifically, we posit that the phonologies! component consists of a single ordered set of rules which ‘maps strings of morubemes io their dietionary form onto the phonetic repte- sentations of the sentences comaprised by these strings. An initial subset of these rules is the set: of MS rules, while the rest are the P rules. MS rules apply only ‘within single morphemes, and may change feature values if desired. Finally, some redundaney rales may be stated in the P rules, so that the output of the MS rules neod not consist of fully specified matrices. Our coneern will be the difficulties which blanks feature values esuse in this, the usual conception of phonology. Recall that the set of MS niles applies to each morpheme's dictionary matrix, filling in ‘0 entries in this matrix and producing a more fully speoified rastrix: the systematic phonemic matrix of the morpheme.* ‘Thus we have chosen to indicate formally the fact that a feature value is redundant in the systematic phonemic matrix of 2 morpheme by replacing that feature value by ‘0" in the dictionary matr'x of the morpheme and by including an MS rule which changes that ‘0' to the proper feature value. However, in using this syste we must take care that the feature values remain swan, and that ( in a dietionary matrix is pever allowed to function as & third value, distinet trom ‘+? and ‘—', The correctness of any empirical claim that distinctive features are binary is, of course, not at issue here. The point is simply that, once we decide to use a binary system, we must be formally consistent. Unfortunately, itis all too easy to be formally inconsistent by letting ‘0’ funetion as a third feature value, and this hhas often been done unknowingly in the writing of generative grammars, What is important is Ubat we keep the meaning of 0" clesely in mind. It is not a feature value, but merely a mark which indieates that the feature value of the entry in which it appoare has nob yet beon filed i ‘There are two general ways in which ‘0’ ean schieve an improper status. One way is to let the oxy difference between two dictionary matrices be that: one has a ‘++! entry (or a ‘—" entry) where the other has a ‘0? entry; in this ease the 0 entry is being used to keep the two matrices apart, contrary 10 its intended use. The other way to let ‘0! achieve improper status is to formulate rules which ‘mention ‘0' in their enviroament; clearly a rule must never depend for its proper application on thore being ‘0? entries 2¢ yet unfilled in matrices. In either of the above eases, “0” would be functioning as a feature value. Not only is this contrary to our intention to have a binary system, ut it would even be improper 21 Roth the ayibel and the absenes of a aymbl. [6a blank, are used in the Keerature ‘to indicate redundant feature value, We will nae them inverckangeably.REDUNDANCY ROLES IN PHONOLOGY: au as a ternary system, since only ‘++' and ‘—', and not blanks, are counted in the evaluation procedure. When ‘0? is given improper status, it is often said that « ‘specious simplifea- tion’ or ‘specious saving’ has been made. Specific examples of these situations will be given in §23. Such distortions of the binary system are, of course, to be avoided, and to avoid them we must take great care in how we formulate MS rules and how we interpret them as applying to dictionary matrices. Consider able attention has been given in the past few yesrs to this problem. The goal is, 4a formalism for MS rules which allows one to state all legitimate generalizations ‘but which prohibite one from making any specious simplifications: Le. a forma- lism for MS rules which does not permit ‘0? to function as a value distinet from ‘47 and ‘=’, This goal has been partially achieved by the general adoption of conditions on MS rules stch as the ‘distinctness’ condition and the ‘well-formed- ‘ness’ condition, but these conditions have several defects. The present section is devoted to a discussion of these issues. 3.2. Ruum pmewvrntox, We have said that MB rales are redundancy rules: ‘that they are instructions for filling in the proper values (‘+' or ‘—") of features which have been left blank in dictionary matrices. Tn this section we will give a precise definition of MS rule, and will show just how such a rule is to fill in feature values. Following standard practice, we will say that « MS rule R con- sists of tivo parts: & STRUCTURAL DescRIPTION SD(R) and « STRUCTURAL, CHANGE SC(R). SD(R) and SC(R) may be thought of as being a pair of incompletely specified, disjoint matrices with the same number of columns, The kinds of nétrices we are concerned with here have entries ‘+", ‘—", or no entry (blank), and bave a fixed oumber n of rows, each of which will correspond in an actual grammar to a distinctive feature, To say that a matrix is incompletely specified is simply to say that some of its entries are blank. To say that two matrices are disjoint is to say that one has a specification (a “+" ot ‘—-") in 2 position only if the other bas no specification (a blatck) in the eorresponding position. To illustrate these terms consider a nule R where SD(R) and SCR) are as follows: qa) (aa) (a) —I-fl412 LANGUAGE, VOLUME 43, NUMBER 2 (1967) ‘We will sometimes use this notation, but will often write R. in « form in which SDR) and SCR) are kept more clearly distinct: sp [*f] a t- «yy sont] ‘The difference hetween the formulations in (13) and (14) is, of course, purely notational. MS rales, as they have been formulated in actual practice, are often such that SC(R) lies entirely in one segment; that is, as in the above example, features are added in just one segment. However, there scems to have been no attempt to make this a formal requirement in MS mules (analogous to the requirement in, ‘the phrase structure rules that a single symbol be rewritten at a time), and in fact it has proved eonvenient to add features to more than ome sesment with a. single rule. Thus Halle (1959:60-1, Rules MS Lid, Lle) seems to indicate that, ules involving additions in two segments are necessary."* It is for this reason ‘that we have set up our formalism so as to allow such rules, The disadvantage of this, of course, is that we have a quite powerful formal device which is needed in a relatively small number of cases of a limited sort. Note that, by requiring that SD(R) and SC(R) be pissouvr matrices, we are unable even to formulate any rule which mentions the same feature an both sides of the arrow, as in (15) = ale xf (as) Loo ee "Thay are not logiesly necessary, of course cinco any MB rule of the form Al) oo 4 ich” wy (whore A, B, G, aud D atand for bundles of feature valves) cen be reformalaed in two rsles tL (8) la) 1B) ~ 1 and 1 cs) DI ‘The point is that such « reformubation reste in a loss of genecality, since the envicanment. SDR) = (Al [Bl mut be stated twiee. Note, however, that the reformulation of (X) aa the tra rules of (W) ia possible in general only if we are talking sbout MS rale—rules which do rot change feszure values, If the rules were F rulee, ther ‘hie reformlation would not be ‘ezelble if the fre rule in (¥) cheaged features in A in euch a way as to ronder the wesopd rule in (¥} inapplicable,‘REDUNDANCY RULES IN PHONOLOGY ana And, indeed, in actual practice, rules of the above form have not been used as ‘MS rules.* 3.8. Ruve areuicdmon. Informally, an MS rule is an ‘if-then’ condition on matrices: each rule R says ‘if a matrix M meets condition SD(R), then it must also meet condition SO(R)’, or ‘if 4 matrix M meets condition SD(R), then we fre to fill in blanks in M until it also meets condition SC(R)’. To say more precisely how MS rules apply, we must introduce some preliminary definitions. ‘Tet ussay that, given (éwo completely or incompletely specified) matriees Mand IN WSICH HAVE THE SAME NUMBER OF COLUMNS, then: (1) M is a sca-warmrx of N if, whenever M has a specifieation (a ‘4? or a‘) in @ position, N has the same specification in the eorresponding position, tbut not necessarily conversely. (2) M is oismver from N (equivalently, N is distinet from M) if for some specification in M, N has the opposite specification in the corresponding pasi- tion, (8) M nd N are moowrananun if they are neither distinet nor in a sub- ‘matrix relation. ‘When a matrix M is subjected to a rule R, we define the areracanrsrry oF R 10 M 4s follows: (4) Tf SD(R) is s sub-matrix of M, then M meets the conditions of R; thus R avrutms to M, and (the minimum required number of) features are added for changed in M until SC(R) is also a sub-matrix of M. (5) If SD(R) ie distinct from M, then M does not moot the conditions of Rs thus R ras 10 appt to M, and M passes through unchanged. In these two cases the definition of rule applieation is the obvious and natural one, Regarding the third possibility, thet SD(R) is incomparable with M, it ie not clear whether we should say that R applies to M ot that R fails to apply to M. Suppose that we say in goneral that a rule R. anus to apply to (i.e. has No effect on) a matrix M if SD(R) is incomparable with M. Call this the sup- MATRIX INTERPRETATION OF RULB APPLICATION: a mule R applies to M just in ‘ease SD(R) isa sub-mairix of M. Then given the three eonsocutive miles Cd (+) [-f] «16) 1 L L f-s) +el +a wwe see that segments blank for feature i will end up different both from sep- ments matked +f and from segments marked —f after the application of these rules; segments blank for f end up —g, segments marked +f on —f end up +6. But then, a6 far as these rules ate concerned, segments blank for feature f are weimen +f xon ~f; i.e. when blank, feature f funetions se though it had a. third valuo, different from both ‘+" and ‘~, It is easy to sex how the uso of rules having the property of the rules in (16) can lead to specious simplifications of the kind mentioned above, For example, if we add a fourth rule Le, however, that the requirement of dicjointvess does not completely rule out the ey that an MS rule can change fesiute values; eee (21) im $35.44 LANGUAGE, VOLUME 43, NUMBER 2 (1987) f=s] 1 (+4 to follow the above three, thon the three ‘dictionary matrices? as fe) (oa) Loa) ‘when operated on by the four rules of (16) and (17) become aay (S] Gal Gal But then, clearly, ‘0’ bas heen used ss a third feature value. Thus this interpre- ‘ation of rule application (the sub-matrix interpretation) is not the desirable one. ‘Suppose, on the other hand, we decide to say that a rule noes apply to (Le. does have an effect on) a matrix M ff SD(R) is incomparable with M. Call this the oistivcrvess WTERPAETATION OF RULE aPvLIcaTION: a rule R applies to 1M just in case SD(R) is not distinct from M. Then, given the three consocutive rules an Lolo ten A (20) I | 1 {c4] Gel GAL we see again the blank (in feature f) is funetioning as a third value: segments blanke for feature f end up (+, +l, segments marked ‘+' for fend up (+, —hl, segments marced ‘—' for f end up [—, +h], after the rules in (20) have applied. Te is easy to see how the use of rules having the property of the rules in (20) can lead to improper use of blanks; an example similar to the one just gived i readily constructible. It follows that, under the distinetness interpretation of rule application, as well as under the sub-matrix interpretation of rule applica- tion, we are unable to avoid improper use of blanks. In short, when rules have ‘to recognize blonks in any way,” the road is open for these blinks to function as a third value, and we see that this is tene under either of the two possible interpretations of rule application Thus there is No adequate way of deSning the applicability of a rule R to a matrix M if SD(R) and M are incomparable, ‘To avoid this dificulty, it ecems we must require that, whenever @ matrix M is subjected to a rule R, we must be able to tell from the spmcteten features in M whether SD(R) is a sub-matrix of M or is distinct from M; ie., whenever a ‘matrix Mis subjected to a rule R, it must not be tho ease that SD(R) and M are incomparable. This requirement has come to be referred to 38 the well- forme:iness condition, “That ia, when we have to dovide for a rule Rand a matrix M, where SDR) and M are Iwcomparsbie, whether Rix to spply to Mor aot 2 For aa cay atatemont of ths problem (in which, we might add, no conclusions are reached), ee0 Lightner 1963REDUNDANCY RULES IN PHONOLOGY 458 344. Tr wetr-soRunowess conDITmON. If there are MS rules Ry, Ra oy Foe in a language, then given a dictionary matrix D, the corresponding systematic phonemic matrix is formed by subjecting D to Ri, by subjecting the result to Ry, ete. Clearly, each time a matrix M is subjected to a rule R, we have to be able to decide in some way whether or not R.is to apply to M, and the theory hhas to give us a constant, formal basis for making this decision. A set of dic- tionary matrices and MS rules is said to he wett-ronuen if, whenever we sub- ject a matrix M to a rule R; in the course of this process, SD(R,) is distinet either from M or a sub-matrix of M?* ic. ¢ grammar is wellformed if the strvATIoN even antsrs that a matrix M is subjected to a rule R; where SD(R;) and M aze incomparable in the sense of §333 (3). This well-formedness condition was not sgiven explicitly in Halle 1959, but has been conformed to in practice and has ‘been explicitly stated sines 1959, (For s different but equivalent version of this, condition, see MeCawley 1964.) The reason for having this condition is that, a8, Wwe have seon, specious simplifications can arise in grammars which are not ‘wall formed. Tho problem involves the decision (which must be made once and for all in the defnition of rule application) whether or not a rule R is to apply to a matrix M if $D(R} and M are incomparable; we have seen that either way of making this decision leaves open the possibility of using MS rules to make specious simplifications. ‘This problem does not arise in well-formed grammars, since, by the definition of well-formedness, this decision does not have to be made: Tt is on the bavis of the above kind of reasoning that the well formedness condition has been generally adopted, at least in practice. ‘I is a simple matter to cheek a set of MS rules for well-formedness: we need only to check something about how a.sraall finite number of rules (the MS rules) apply to each of a finite number of matrices (the dictionary matrices). More- ‘over, wellformedness does not need to be a condition on how P rules apply if, as we suggested in §12, these rules operate on strings of matrices which arealways fully specified Suppose, however, we adopt the usual practice of letting the segment structure ralles be scattered through the P rules. Then the P rules operate on strings of ‘aatrices which have blanks in them, making it necessary to require that the P rules, as well as the MB rules, be subject to the well formedness condition. This ‘can be seen clearly in the following two examples, in which we arrive at specious simplifications (using each of the possible interpretations of rule application) in P rules whicb do not meet the well formedness condition. The examples ate actually eonerete illustrations of the principles involved in (16)-(20) above, *N¢ will alwaysbe the outputof rule Rr except when i= 4, in which ense M will bea die- ‘tionary matriz.—Note that “well-iormed’ samen s apeciSe, precisely dedined formal prop- erty of grammars; iin mor s general property which aozuekow relate the itive mea fing of the term. * Hie ection 2.15 (p. 57) seems to indionte that this condition was implioity required * Ip might be argued thet it doesn’t matter if the formalin permits the uso of rulea hich lead to specious simplidcations as long as we never actually make use of such rules ‘This, however, is o misunderstand the purpose of a formal theory, which is, afterall, to Drovide the most srvcivic and ataxz-eraucreesn formalin with which iti possible to {rent the phenomena in quention, Only then enn x0 explanation be reached of why the phe ‘nomena are organized in one way rather than another.416 LANGUAGE, VOLUME @, NUMBER 2 (1907) Note that, in these exzmples, we specifically assume that the sogment structure rules gre in the P rules, and thus tbat the input to the P rules contains blank entries, Hxample A: Suppose we assume the sub-matrix interpretation of rule appli- cation, and suppose that in some language all [—Consonantal] segments are (+Voiced} and that, further, there is 6 P rule that affects only [+ Voiced, -+Con- sonantal] segments, If this P rule applies before the segment structure rule whieh makes all [—Consonantal} segments [4-Voiced], and if we don’t require well- Formedinees, then all we need to state in the structural deseription (SD) of this P rule is [4-Voiced]: the rule will then apply to just the [+Voiced, -+Con- sonantall segments, sinee the [—Consonantal] segments, though really (4+Voieed|, have not yet been specified as {+Voiced]. But this is a specious simplification, an improper use of blanks. ‘The formatiam should require us to state [+-Voiced, +Consonantall in the structural description of this P rule (whieh is just what we do have to state if the whole grammar, including P rules, is well-formed). Tf we only state [+-Voieed, then we are persona on the fact that the {—Con- sonantal] segments are blank for voicing at this point. Tf these segments hed been specified beforehand with the correct. value of the feature Voiced (i.e. [-+Voiced), the P rules would give the wrong results, This is exactly the same as saying that ‘blank’ (for the feature Voiced in {—Consonantal] segments) is, functioning as a third value, distinet from ‘+' and ("5 Example B: The same kind of improper use of blanks in the P rules can be illastented if we assume the distinctness interpretation of rale application. Consider language in which the feature specification (—Voiced] is never pre- dietable by a segment strueturo rulo; this would be true if, for example, for every [-Voieed] segment there slways occurred a corresponding [++ Voiced) segment in the same environment. Suppose further thst all vowels, liquids, and glides are {-Voiced). Consider now a stage of the P tules before the (+Volced! speci fieations have been inserted (by a segment, structure rule) in vowels, liquids, and glides, If, at this stage, we want a rule R which applied to all the [+Voiced} segments (which constitute a natural class), then SD(R) will be [+Voiced], 28 expected. Yet if we wanted a rule R’ which applies to all the {—Voiced] eon- sonante plus all the vowels, liquids, and glides (which constitute a much less natu- ral class), we would still need only one feature specification in SD(R'), namely [-Voiced]. This is an intolerable situation: in any formalism where it can arise, the concept of natural class, otherwise so elegantly characterized in distinctive eature theory, cannot be captured. The problem, of course, is that a rule such 1c might be argued that it i correct to state a P rule such es the above, with ouly [+ Voiced! in the stretural description of the rue, oa the bacis of the following principle: a rule R should apply otfy to cheae seementa which are non-redundant}y specified for the features given in SD(Q). As we ean gee from the above example, this principle eould be put inca affect hy sasuming the eub-tuatsix interpretation of rule application and by putting the segment structure rules at the end of the P rules, However, this would foros us to have f.ayatem in which blank function asa third feature valu, disizet from 4" and ‘~'. More- ‘over it seems tha: ia any’ aotual grammar there are just a many cxaes ia which we want to Ihave a rule R apply to segmenta which are nabexpannzr SDR).REDUNDANCY RULES IX PHONOLOGY aur fs RY applies to too many segments. Tis structural description SD(R’) is {-Voiced], and yet it applies to vowels, liquids, and glides, which are really [+Voiced), as well as to [Voiced] consonants. It osm do this simply benause these. (+Voiced] specifications, being redundant, ate not inserted until ater. ‘The result is that s rule such aa R’, which supposedly represents generalizations about segments which have the feature specifications given in SD(R’), now applies to many segments which will never have these specifications, but: which ‘happen at this point to be blank for these specifications. Such a result, destroys ‘completely the meaning of phonological rules. From these examples and many more like them, we see that, as long as there ‘are redundancy rules {the segment structure rules) scattered through the P rules, the problem of blank specifications arises in the P rules. ‘fo solve this problem, we could, of curse, require well-formedzess in the P rules, just as in {33 we solved a similar problem in the MS rules by tequiting well-formedness. But then, to cheek a grammar for well-formedness, we taust not only see how MB rules apply to dictionary matrices, but we must algo see how P rules apply to the systematic phonemic tepresentations of whole sentences, Since there is no ‘upper bound on the length of sentences, it is elear that there is, in genersl, no finite procedure which will tell us whether a given grammar is well-form: ‘eg; we never know when some transformational rule is going produce & ¢on- figuration of systematie phonemie matrices which is incomparable with the ‘environment of some P rule. It seems much more naturel to make the require- ment that; the input to the P rules consist of fully specified matrices, for then the problem of wellformedness does not arise in the P rules, Well-formedness is then simply a condition on how MS rules apply and, as already mentioned, is easily checkable, However, the problem of specious simplifications obtained through the im- proper use of blanks is not solved by the well-formecness condition. This eon- dition, though it. does prevent all improper use of blanks, is actually too strong. ‘We will not show this directly, but will rather show that a weaker condition on grammars, the distinctness condition, is iteelf too strong. 3.6. Tu orsriNcrEss conpImtoN, The well-formeduoss condition has not bbeon widely discussed. It has been much more common to assume the DisTiveT- = Tusbould be noted that iia posable to define condition on the P rules and systematic tuntrices which ia finitely checkable and which salves the problems which the well-formed- neas condition is designed to solve. This condition i slightly stronger than the wall-ormed- ness condition. ‘The well-iormeduess condition requires thet no sequence of aystematic phonemic matriceg WHICH Ia AGhOAILY GENERATED RY THE SYATAGUHC COMPONENT OF HE ‘onaauan be ineosparkhle with the structural description ofthe frst P rule, that n0 con- figuration of matrices obtained by applying the fis: P rale to one of Taxes sequences be incomparable with the structural description of the second P rule, ete. The new condition ‘would requiee that the shove conditions be met not oxy forall configurations of aystematic ‘phonemic matrices etually geaerated by the grammar, but for aut vossints configurations of systematic phoneme matrices. However, such @ condition ia too rostrictive; it merely ‘bypasses the problom of blank specifications, and doeg nat come co terme with Wt. Tn fact we vill soe Inter that oven the well formedncae condition i418, LANGUAGE, VOLUME 43, NUMBER 2 (1067) NESS CONDITION on grammars as & means of preventing improper use of blanks (cf. Halle 1959:82, Chomsky 1965:81)" This condition requires that the dio- tiouary matrices be pairwise distinct, ie, that there be no pair Ds, D2 of distionary. matrices seh that Dy is not distinet from Ds: It is itmportant, to notice that ‘distinct’ is being used here in the technieal sonse defined in §33, and not aa an informal equivalent of ‘different’ oF ‘distinguishable’ In partienlax, the distinet- ness condition is not to be takon sa being a condition which, a prior, dictionary matrices must meet. In fact we will show below that: this eandition is too strong. For the moment we merely observe that every actual graramar that meets the well formedness condition must also meet the distinetnoss condition; for, assum- ing well-formedness, it is apparent that: any non-distinet pair of matrices must ‘undergo exactly the same rules, and so could not emerge different from the MS rules. We aleo observe that a grammar which meets the distinctness condition need not meet the well-formedness condition; for example, the four rules of (16) and (17) violate well-formedness when they apply to the second and third ‘matrices in (18), even though these matrices are distinct. Thus, woll-formedness is a strictly stronger condition than distinctness. Finally, we note that, the distinctness condition is easily checkable; we simply nood ia eheek something about each pair of members in a finite set, the set of dictionary matrices. ‘The distinctness condition was originally adopted to rule out the following particular kind of imnroper use of blanks. Suppose we have s language in which the feature Tense is predictable in consonants, [+Voiced] consonants being [—Tense] and [—Voieed] consonants being [+Tensel. Then we can let the ont difforenes hetweon peirs of systematic phonemes liketand d in diotionary matri- 08 be a8 Follows: td (aly loarly, we sre not using enough information to keep f and d apart in (21). Yet, if wo use the distinctness interpretation of rule application and the rules in (22), ‘we can predict the correct values of the features Voiced and ‘Tense in ¢ and d in as ‘The distinctness condition is not to bo confused with the ‘distinetnose Interpretation ‘of rule spplication’ ($2.2); both concepts, however, are baaed on the definition of ‘distinct! pair of matcioes given at the beginaing of §83. Halle uses be term ‘different from’ ‘The distinetness condition is correct in aot ruling out « situation where non-diatinct szourwra become distioct by the application of the MS rules if these segments oceur in distinct matrices Tn tact, thig ig exactly the gitustion when neutralized segments are repre- tented by archiphonesoes, _ "= This ‘grammar’, of course, does not obey the well-formedness ooaditian, since it wag choaon expreadly to violste the weaker distinctnogs condition. This will bo truc ofall the ‘examples of this section.REDUNDANCY RULES IN PHONOLOGY 419 (4Consonantal]—> [+Tense} (22) [senna] > [Tense] [foment] > (voit ‘This intolerable situation is ruled out by the distinctness condition, since t and d ‘were not kept distinet in the dictionary matrices. However, it is important to notice that there are other ways of ruling out this kind of ease. For example, we could require that MS rules not be allowed to change feature values; since the secoud rule in (22) involves changing features in an essential way, this require- ‘ment would cuffs. Altematively, we could require that MS rules must use the sub-matrix interpretation of rule spplicatioa, rather than the distinctness inter- pretation asin (22), Infact, we will show im §3.6 that these latter two requirements are essentially correct, whereas the distinctness requirement is not, It is truo that the distinctness condition happens to rule out many types of ‘improper use of blanks in the MS rules (such as the one just discussed). However, Lnlike the stronger well-formedness condition, it is not sufficient to rule out Aut improper use of blanks. This follows since (using either of the two possible inter- ppretations of rule application) the kind of specious simplification obtained in {36)-(20) could still be obtainod if the matrices involved were merely subparts of distinct matrices. The same point can be made by viewing the rules in exam- ples A and B of the preceding section as MS rules. The reason that the distinctness condition fails is that it has not come to terms with the problem, discussed in $83, that there 18 no adequate definition of how rules apply if the rules must recognize blanks in any way. ‘Thus the distinctness condition is not strong enough. However, in another sense, the distinetness condition is too strong. These two facts indicate that it is a wholly inadequate condition to place on grammars. ‘There are many ways of showing that distinctness is too strong, In fact we have already pointed out (§2.2) that ‘high level’ features may be predictable in terms of ‘low level’ features, and that this prediction can only be made at the expense of violating distinctness; rule (10) is an example of such a prediction. ‘The eecence of this kind of ease ean be illustrated by the following example, Consider the matrices We) a (2a) fay and the rules itl > (-t en tA) [=8) ‘Then, using the sub-matrix interpretation of rule application and the rules (24), ‘the non-distinct matrices (23) become the distinet matrices40 LANGUAGE, VOLUME 43, NUMBER 2 (1987) +f But in no sense do we have here an improper use of blanks. That is, we aro not ‘using, blanks in keeping the matrices in (28) apart, but rather the specifications +f; in-the one and +f: in the other, Ta other words, we would a priori rule out 2D), but not (23), as being instances where mattices are being kept apart with too little information; yet in eseh case the matrices involved are uot distinct, ‘Thus, the distinetaess condition rules out desirable and undesirable cases indis- cciminately. Tt is possible to give a concrete case which illustrates the fact, suggested by the shove remarks, that the distinctness eandition is too strong. Consider Modem. Standard Russian (MSR). According to Halle (1950), s grammar of MSR. in- cludes the following two MS rules: ay 4 an + [ogame] | [vase] w [ ] Eee) [Berges] + ee] Rule (26) says that a marpheme initial glide must be followed by a vowel. Rule @27) says that & morpheme-finsl glide-consonant cluster mnst be preceded by vowel. But now consider the forms iu and jis, which certainly are possible ‘morphemes in MSR. What would the dictionary matrices for such forms be? Whatever thete matticos are, we would like to be able to use rules (26) and (27) to make savings in them. If we did, then these matrices would be: ® [¥) Consonantal =] Voestie = [2 [= (We ignare all other features, which are the same for each form since j and i differ only in the featare Voealic.) However, the two matrices in (28) are not distinct, Since there aro roxsons to give rule (26) first, we conclude that to maintain dis- tinctness we must write the dictionary matrix for ij as: (ay Consonantal sald VocalioREDUNDANCY RULES IN PHONOLOGY 41 But then we miss being able to make use of the generalization that ant morpheme-final glideconsonant clusters, oven those in threo segment mor phemes, are preceded by a vowel, I rulo (27) had been first, a perfectly analogous problem would have arisen, ‘This example supports the conclusion that the distinetness condition is too strong, since this oondition prevents us from making savings in the dictionary where they seem desirable, We are therefore led to reject both the distinetness condition and the stronger well formedness candition as being inadequate devices for preventing the improper use of blanks in the MS rules. The distinctness con- dition is, of course, expecially inadequate, sinoe in some eases it is too restrictive while in others it is not resiriotive enough. Clearly, however, som condition is necessary. In the next section we will see that there is a tather natural and obvi- ‘ous condition that has so far been overlooked. 3.6. Tae TROR GENERALIZATION CONDITION. In stating the new condition we ‘will assume that, the sub-matrix interpretation of rule application is used exclu- sively in the MS rules. Evidence in favor of this interpretation, rather than the distinetness interpretation, will be given below. We will aleo assume that all the redundancy rules aro in the MS rules, so that the systematic phonemic matrices, which emerge from these rules are fully specified. The problems involved when this assumption is abandoned have been fully discussed (§41.6,2.4). Finally we will, ‘ascume that each dictionary matrix D is simply los fully specified version of the correspanding completely aporifiod systematic phonomie matrix S;i.0., we assume ‘that in each case D is a sub-matrix of 8, These assusoptions are merely corollaries of our derision to identify the concepts of redundancy rule and MS rule. ‘A grammar is suid to meet the TRUE GENERALIZATION cONDITION if each MS rule represents a trae generalization about the (fully specified) systematic pho- nomic matrices of the language. That is, given any (fully specified) systematic phonemic matrix $ and any MB rote R, it raust not be the case that SD(R) is a sub-matrix of 8 when SCR) is not, for in such a ease the rule R. would be making 4 false statement about 8. ‘The motivation for such a condition becomes clear when we recall that MS. rules are intended to state redundancies in systematic phonemie matriees; we are simply requiring that each such siatement he true of these matrices, Surly such. a requirement must be implicit ia any stetement of redundancios, and to violate it would be highly unnatural. In fact, put in this form, the requirement js 20 obvious thst it hardly neods to be stated at all? Despite the simplicity and naturalness of the true generalization condition, it and it alone suffices to prevent all specious simplifications and improper use of blanks, This can be illustrated in many ways, In grammars meeting the tras generalization condition, a rule can never DEPEND for its proper application on The reason the naturalness and usefulness of this requirement have not been noticed previously ie probably due to the fact that redundancy roles and featore changing roles have generally been intermingled in the P rule, and have nor been formally distinguished However, once the suggestion of this paper ia adopted and the redundancy rales are can sidered as a geparste set of rules (the MS rules) quite distinct from the P rule, thea it becomes obvious chat the tne generalization ceatitemens is merely par? of what we mead by redundancy rules (ef tho historieal discussion in $12).422 LANGUAGE, VOLUME 43, NUMBER 2 (1967) there being blanks in matrices at tho stage at which it applies, since wo require that the rule also apply correctly to the ayetematio phouemie matrices, in which there are no blanks. Thus we note tht this condition rules out the improper use of blanks illustrated in §34, example A" In §23, (16)-(19), we gave an example, using the sub-matrix interpretation, where ‘0’ functioned as a third feature value. ‘The true generalization condition clearly rules out such a case, sinee the second rule in (16) does not represent a true generalization about the metrices in (19). In §8.5 we showed by means of several examples that the distinetness condition ‘was too etrong. The true generalization condition does not share this defect. In particular we note thet the non-distinet matrices of (23) are now allowed, as desired, since the rules of (24) represent true generalisations about the matrices of (25). Also, we may now have the matrices of (28) in the Russian example, which are the desirable ones but which are not distinet; this follows since the rules of (28) and (27) each represent legitimate generalizstions about Russian systomatie phonemic matrices. Finally, notice that the true generalization condi- tion does not allow the kind of improper use of blanks lustrated in (21) and (22), since the first rule in (22) is not true of all consonants, ‘In summazy, the true generalization condition preventaimproper use of blanks, but is not go strong that it prevents us from malting certain generalizations. The condition is also easily checkable; we sinaply need to take each systematic pho- nemie matric of the language and check to sce that no MS rule makes a false statemant about this matrix. Tis easy to seo that, if the true generalization condition is met, MS rules will not change feature values, since each dictionary matrix is a sub-matrix of the corresponding fully specified systematie phonemie matrix, and the MS rules, which fill in the former to produce the latter, must make true statements about the latter. It is natural t ask if the true generalization condition might bo t0- placed by the (weaker) requirement that MS rules do not change feature values. ‘The following example shows that this is not possible, Consider the matrices A and B below: AB AB Fle += (30) aitl+ + oye i If we use the sub-matrix interpretation of rule application and the rules [+g] {-b] and (—f] -» +e, +h], applied in this order, then we see that, we ean let A” and BY be the diotionary matrices for A and B, However, the first rule applies properly only because the second has not applied and feature g is still blank in BY. Thus we have a specious simplification: we are able to get away with too ‘s"Thore we used the submatrix interpretation of rule application ard obtained specious lmplifeations in the segment structure rules. In that example, chase rules were viewed a8 rites; but in che present section wa have returaed to aur aggummption ‘hat aceh rates, bing redundancy rales, are in the MS rules, The essence of the example is the same ia clther ease.REDUNDANCY RULES IN PHONOLOGY 23 general a statement of the first rule, which should have to be stated as(+f, -+e]—> 1-2), This shows that merely preventing rules from changing feature values is ot, sufficient to prevent specious simplifications, for the above rules do not change feature values, However, the first rule violates the true generalization ‘condition sinee itis not true of the segment B. Thus this condition rules out the above ease, as desired. Though itis not necessary for rules to change festare values in order to obtain, the specious simplification in the last pansgeaph, itis neceasary thet the two rules be ardered with respect to one another. If we require norm that MS rules he un- ‘ordered an that they do not change feature values, then no specious simplifiea- tions are possible, Moreover, it is not hard to see that those two requirements together are equivalent to the truo generalization condition. This, in tum, is just way of saying that the MS rules aro exclusively redundancy rules, which was precisely what we argued in §1. Thus we have given two independent kinds of justification for this position, a formal one (in this section), and a linguistic one Gn $0. ‘Note further that the true generalization condition gives an obvious and natural eriterion for determining when two dictionary matrices are different, ‘enough to be kept apart; two dietionary matrices are properly distinguishable just in case the MS rules render them distinct, where ‘distinet’ has the formal ‘meaning of §23. This follows since MS rules meeting tho true generalization con- dition can never make improper use of blanks in rendering non-distinet matrices distinet. In closing this section, we will give evidence supporting our decision to use ‘the sub-matrix interpretation of rule application, rather than the distinetnoss interpretation. Recall thet, in the sub-matrix interpretation of rule application, 2, rule R applies to a suatrix M just in ease the environment SD(R) of R.is a sub- matrix of M. This interpretation mesns that R. would still have applied to M if the blanks in M had been filled in beforchand in any arbitrary way; in other words, tune APPLICATION oF R 8 INDEPENDENT OF ANY UNGPECIFIED FEATORRS 18 M, Such an interpretation seems intuitively more natura) than the distinctness interpretation of rule application. Jn this latter interpretation, rule R applies to a matrix M just in ease SD(R) is not distinet from M, Thus a rule R might apply toa matrix M even if M would be distinet from SD(R) if it had all its redundant features filled in. All that is required is that M not be distinet from SD(R) at the point of application of R. ‘However, os wenoted in §34, exaraple B, to have a rule R pply to tuattices which are distinet (even though redundantly 60) from SD(R) is an entivey wunatural situation. Tt not only places an artificial importance on the onDER in which re- dundant features are filled in, but also destroys the crucial idea thet a rule R represents a generalization about, and thus applies to, exactly those matrices which have the features of SD(R). To illustrate this point further, consider the following example of improper uso of blanks which would be attainable under the linctuess interpretation of rule application, even if the txe generalization condition were met. Suppose thore is a language whoce strident, segments are ‘exactly the contiauant consonants; and suppose we mark the eantiquant conso-44 LANGUAGE, VOLUME 43, NUMBER 2 (1967) ‘ants in the dictionary [QContinnant, +Strident] (i.e. as blank for continuaney), fand the stop eongonants as [OContinuant, OStrident] (Le, we do not mark the stops for either continuaney or strideney). ‘Then if we have rules (1) [Strident] + (—Continuant] ) [—Continusnt] — {~-Strident] @) [4 Continuent] — [+-Strident] () [4Strident] [4+ Continuant) (each of which represents a true generalization about consonants), and if wo use the distinctness interpretation of rule application, we can have the following derivation: (31) D 1 2 3 4 et et stot ee Continuant = [=] Fe] (l= wm gger GS Ge] EE EE] HE Here the entry under ‘D’ represents the values of the features Continuant and Sicident in the segments s and t in the dictionary; the entry under ‘I’ repre- ‘sents the same information after the first rule in (31) has anplied, ete. Thus we see that, even if the true generalization condition is met, the distinctness interpre- tation of rule application allows us to keep two segments apart in the dictionary using only one specified feature value—clearly an improper use of blanks, This supports our decision to use the sub-matrix interpretation of rule application, in which such an improper use of blanks is not possible, MORPREME STRUCTURE CONDITIONS 4.1, Tor cenenat arproscer. ‘The point of this paper so far is that we must take seriously the idea that the MS rules of a language are the rules which state the redundancies at the systematic phonemic level, We have neen that the MS rules are most naturally viewed as an unordered set of statements which represent ‘tre generalizations about the fully specified systematic phonemic matrices, and ‘which may be used (as an unordered set) to fill in the redundant feature values in dietionary matrices, each of which is 2 less fully specified version (a sub-matrix) cf its corresponding systematie phonemic matrix. In this section we will eee thet the most natural way of formulating such statements is not in terms of rules at all, and we will show how MS rules ean be replaced by a new device, MS coxor- ‘rows; we use rules (ike the P mules) only to map one level onto another, and ‘eonditious' to state redundancies at a given level. MS conditions not only avoid, ina natural way, all the formal problems discussed in §2, they also allow us to state situations which arise in netural languages but which are not easily stated in terms of mutes. ‘In many ways, MS rules and MS conditions are very similar. In particular, we ‘ean talk about segment structure CONDITIONS and sequence structure ConITyoNs in the same way that we talked about the corresponding kinds of notes in §1.3REDUNDANCY RULES IN PHONOLOGY 425 and §14, The discussion in §§1.5, 1.6 will also carry over essentially without change into MS conditions, and we will rotain the decisions made there. Briefly, the proposed system consists of the following two parts. First, there js an UNORDERED set M of MS conditions which defines, in a msaner speeified in detail in §§4.2-46 a sot M(U) of Fuzur spctrien matzices, where each matrix jn M(U) has a numbor of rows equsl to tho number of distinetive features in the language in queetion,® and whore each matrix in M{U) has i ealurans for some 1g i 9, where b is the longth, in segments, af the longest morpheme of the Ianguage. ‘The set M(U) contains tho systematic phonemic matrix of every ‘morpheme of the language; it alsa contains matrices for forms which, though not in the language, violsie no constraints of the language. In brief, M(U) contains ‘matrices for all the ‘possible’ morphemes of the language (cf. §1.4). ‘The second part of the proposed system is a PROCESS OF SBLECTION. For each morpheme m, the incompletely specified dictionary matrix Da of m sBrners, in a manner described in {47, the completely specified systematic phonemic matri Sq of m from the set M(U}. Thus the process of selection provides 2 method for filling in redundant information in dictionsry matrices. Recelling the discussion in §1.1 regarding the relation between redundaney and constraints, we ean suim- ‘marize the proposed system by saying that the MS conditions give statements of constraints, while the process of selection uses these statements to prodict re- dundant feature values. Thus, statement of constraints and prediction of re- dundancies, though intimately related, are given as separate processes, We will now desaribe the system of MS conditions in detail, frst showing how the MS conditions define the set MU), and then showing how the dictionary matrix of each morpheme selects the corresponding fully specified systematic Phonemie matrix of the morpheme from this ect. 4.2. MS conptmioNs. Suppose we are dealing with a langusge with n distine- tive features, in which the longest morpheme has \ segments. Let U be the set of all FULLY seactFTED matrices with n rows, with entries “4? and ‘—? and with no more then 4 columns, Then U hes 2* members with one coloran, (2°) x (2) = 2% members with two columps..., and 2** members with » columns. Thus U is finite. In general, an MS coxprttoy is a finite statement of a property shared by some, but not all, matrices in U. Thus ‘hes two segments (columns), “Degins with a [4+Consonantall segment’, ‘has no two consecutive [—Voeatiel segments’, ete, are all examples of MS conditions, Itis easy to sce thet each MS condition divides the set U into two parts, one part consisting of alt matrices which possess the property stated in the MS condition, and snather part eo: ing of ail other matricos in U. We will sey that an MS condition C accepts a ‘matrix in U if M has the property stated in C, and that C reanors M otherwise. Instead of allowing any statement of properties of matzices lo serve us an MS condition, wwe will restrict ourselves in what follows to MS conditians expressed in rather restricted standard form." Specifically we will be considering three on ayotematte phone marisa, tionary matrices, end matsee fat ave the satve numer of rows. Ae aleaya, the particular ‘standard form’ chosen in intended to represent the most restricted form in which sil the relovant goneralizaioza cas be expresced.4126 LANGUAGE, VOLUME 2, NUMBER 2 (1967) kinds of MS conditions: ‘if then’ conditions, ‘positive? conditions, and ‘negative? conditions, and we will require that all MS conditions be expressible in one of ‘these three forms. 4.8. Tr-rmmw conortions. An if-then condition © is a pair of matrices T(C) and T(C), the “it” and the ‘then’ part of the condition respectively, where I(C) and T(C) sce each incompletely specified matrices which have n rows (one for cach distinctive feature} and entries ‘++", ‘—', or no entry (blank). Further, 1(©) and T(C) have the seme number of columns and are disjoint (for definition, see {§32). The if-then condition C has the following interpretation: for all matrices Min U Gee §4.2) such that 1(C) is a sub-matrix of M, © accuers M if TO) is also a sub-matrix of M,and C neseens M if T(C) is distinot from M; if 1(C) ia Gistinet frow: M, then C socepte M reganiless of what ‘T'(C) is, This interpretation is wall defined, sinee any incompletely specified matrix such as 1(C) or T(C) is either distinet from or a sub-matrix of any matrix M in U; the matrices in. U are fully specified, and thus it is not possible for either I(C) or T(C) to be in- ‘comparable with M. Intuitively, the if-then eondition G sags that if a matrix M in U meets condition 1(C), then M rust also meet condition T(C) if it is to be accepted by C; the condition C seys nothing chout matrices M which don’t meet, condition T(C) since it accepts all such matrices indiscriminately. The Fons. similarity between the definition of ‘f-then condition? and the definition of “MS rule in §3.2 should be apperent: in fact, evERY MS mute Ras A DIRECT TER: PRETATION AS AN IN-THEN coNDrTION C AND CONVERSELY, with SD(R) corre- sponding to 1(C) and SCR) corresponding to T(O). However, the Ponertowar, difference botweon an if-then condition and an MS rule should be kept clearly in mind, The former is 2 statement which defines a subset of the set. U of fully specified maizices, namely the subset consisting of those matrices in P that it accepts. The latter is an instruction for filling in blanks in matrices with +" and for “—", ‘We show, by example, how we will write an if-then eondition: 4-Consonantal 10) + E-Conseanta [+Comenans!] ‘ (33) —Voealie [—Continuant] —Grave TO | _ Compact +-Continuant, ‘The condition will reject all matrices in U which begin with a (+Consonantall segment followed by a true consonant and which do not also meet the eandition "P(C). This condition is valid in 2 language such as English, whore, in any initial cluster of true consonants, the first consonant is # and the second is a stop (spbere’ and ‘sphinx’ being exceptions). "Aa another example, an if-then condition of the formREDUNDANCY RULES IN PHONOLOGY a7 1G) [—Consonantal] J +Voiced (34) ™O) [ez Strident would reject all matrices in U containing [~Consonantal] segments which are [=Voiced], or (~Continuant}, or [++Strident]. (88) and (24) are soquence struc ture and segment structure conditions, respectively. 4.4, Positive conorrions. The second kind of MS condition ia s ‘positive’ condition. Bach positive condition consists simply of an incompletely specified matrix P(C). Its interpretation as a MS eondition isetraightforwatd: all matrices in U of which P(C) is a sub-matrix aro accepted, all other matrices im U are re- jected. Thus, or example, suppose all the morphemes in some language are of the form (35) cay (5) (where L represents a liquid). Then we would have the positive redundancy condition: 35) 10 + [Ayer (CBee) [vocaie ] + Cossonntarn The condition (36) actually stands for the four positive conditions obtainable from it by considering all possible combinations of optional elements. It accepts only those matrices in U which have the permitted form. Note that (36) is a sequence structure condition. A positive segment structure condition would never be used, since, of two features mentioned in such a condition, one would neces- sarily he completely dependent on the other; thus it would be non-distinetive and not mentioned in the grammar. 4.5. Negative connrrons. The third and final kind of MS condition is a ‘negative’ condition, Like positive conditions, each negative condition C consists aia singe incompletely specified mats: in thin cave me wil denote the mats by the symbol N(C). The interpretation of a negative condition C is that all matrices in U of which N(C) is a subset are agvecten, all other matrices are AccErren. As an example, consider a language which has systematic phonemes: 8 and f but no 2, This situation is described by the negative candition =Vocalie + Compact (37) Nic) ~] Grave +Continuant + Voiced Here the symbol (~) is the sign of negation. The condition (37) accepts just those ‘matrices in U which contain no voiced eontinuant palatal segments. Tt is = seg493, LANGUAGE, VOLUME 43, NUMBER. 2 (1967) ment structure condition. We will give an example below (§6.5) of » negative se- quence stmeture condition. 4.6, ‘Tae rrocnss oF arsrctton, We have shown the formal aature of if-then conditions, positive conditions, and negative conditions, and also hove they are to be interpreted. Specifically, we have shown how each of the three kinds of conditions acces cerisin kinds of matviees in P. A grammar of each natural language will have, in place of a set of MS rules, an unordered finite set M of MS conditions. This set will inciude, in general, conditions of each of the Cree types. The set of all matrices m in U, such that m is accepied by evezy MS condition in M, is well defined; wo call this set MCU). Since each MS condition in M represents a generalization about the morphemes of the language, it follows thatthe set M(U) represents all matrices which violate hone of these generalizations. Moreover, an evaluation measure will guarantee that, for every significant generalization that can he made shout the morphemes of the languoge, thote will be corresponding MS eandition.* Tn short, the set M(U) is exactly the sot of rosstatn morphernes of the language. The Segment structure conditions in M will guarantee that, M(U) eontains only thase matrices ‘whose calumns are systematie phonomes of the language; the sequence structure conditions ip M will guarantee that, no sequential constraints of the language are violated in matrices of M(U). The set M of MS conditions may thus be thought ‘of a8 Gltering out, from the set U of all matrices, those matrices which do not form possible morphemes of the language, leaving the set M(U). ‘We said in§4.1 that a system of redundaney conditions, as applied toa language LL, consists of two parts: first, a set M of MS conditions which defines » set M(U) of fully specified maizices, whore M(U) contains the systematic phonemic matrix of every possible morpheme of L; and second, a process of selection whereby the dictionary matrix D, of each morpheme m of L selects the completely specified systematic phonemic matrix Sq of m from the set M(U). Tn §§41~45 we have outlined the operation of the first part. We rust now proceed ta the second part, the description of the process of selection. ‘Suppose we are given a language I. and a set M of MS conditions written for the morphemes of L. ‘Then the set M(U) is well defined; in particular, it eoatains the systematic phonemic matrix Sq of every morphome m in L, We will show hhow the set M(U) takes the place of the MS rules of the previous theory in filing in blanks in dictionary matrices, Consider a partially specified matrix D with any number of columns and with the same nomber of rows as matrioes in MCU}; ‘cansider aleo any matrix X in MCU). Such matrix X is fully specifed. D is said to aptncr X if D is a sub-mairix of X.* Clesely, it is possible for D to select ‘more than one matrix from MCU), and, in general, the more blanks D haa, the ore matries it wil select. In the limiting case where D has no specifications at all {is completely blank, it will select every matrix in M(U), since in this ease ‘This evalustion messureissimilar to Use ous dicunced jp 14 for MS rales; yee also §52 Notice chat, if D selects X, then D and X must have the came uumber of calumny ef. {2 Since X sa fully specified, on equivalent definition of eeleetion would be that D selects XD and X are aot daREDUNDANCY RULBS IN PHONOLOGY 429 Dis a sub-tatrix of every matrix. On the other hand, if D is fully specifiod and is itself s member of M(U), then it will select just one matrix from MCU), namely the one identical to iteoll, D is fully specified and is not a membor of MCU), then it will select no members of M(U). ‘We are interested in those partially specified matrices D which select exactly one matrix from M(U}. Such matrices D will serve as dictionary matrices, and the process by which they sclect a fully specified matzix from M(U) will be the process, analogous to the application of the MS rules, by which the blanks in dictionary matrices are filled, ‘As an illustration of this process, consider a language which has the if-then condition ($4), For this language the only [—Consonantal] segments that appear in M(U) are also [+-Voieod], [+Continuant], snd {—Stridentl: all [—Conso- ‘oantal] segments that axe [—Voieed] or [—Continuant] or (+Strident] are re- jected by (4). Thus, in representing [—Consonantall segments in dictionary matrices, we enter a value for the feature Consonantal (snd for any other non- redundant features) but can leave blank the values of the festures Voieed, Con- tinuent, and Strident. To see in detail why this is possible, consider any diction- ary matrix D which has a [—Consonantal] segment as, say, its frst segment. Sapnose D selects the matrix 8 from M(U). By tho definition of seleotion, D snust bbe a sub-matrix of S, and thus the first segment of § is also {—Consonantal] Since S is in M(U), the condition (34) guarantees that this (—Consonsntal] eeg- ment is also [+Voieed, -}Continuant, —Strident}, and thus that these three features have been eorreetly selected for D. To illustrate further, consider a language which has the if-then eondition (33). Jn any morpheme of this language that begias with the feature configuration 1(C) of (83), we can leave blank the features of ‘T(C) in the dictionary matrix, since the condition (33) guarantees that such a dictionary matrix will select matrix from M(U) which has exactly the features of T¢C).. ‘Any language which has the positive condition (36) has morphemes which are ‘two, three, or four segments long. In the dictionary matrix of a two-segment morpheme, we can leave the features Consonantal and Voealic completely blank, since (36) guarantees that M(U) for this language has the proper value for these features in each of its two-segment matriees, In the dictionary matrix of a four- sogtnent morpheme, we need only to indicate the value of the feature Voealie in the last segment (to distinguish between » C and an L, both of which occur in this position); all remaining values of Consonantal and Vocalic may be loft blank. Sinee (86) allows three types of throe-sogment: morphemes, CLV, CVC, and CVL, we must. indieate the values of the features Vocalic and Conconantal in sufficient places to distinguish theso three type, all of which occur in M(U). There are several ways of doing this; for exaraple, either a [—Consonantal] in the thind segment or a [+Consonantal] in the second segment tells us that we are dealing with CLV morpheme, and for euch a morpheme no other indications for the features Consonantal or Vaealie need he made in the dictionary. Ins language which has the negative condition (37), wo need to indieate 5 for the feature Continuant but: not for the feature Voiced in dictionary matrices;430 LANGUAGE, VOLUME 48, NUMBER 2 (1987) (37) guarantees that each [4Continuant] palatal in M(U) is {—Voiced). Simi- latly, J must be indicated for the feature Voiced but not for the feature Continuant. Finally, note thet non-distinet dictionary matzices can select distinct matrioas from M(U). For example, the non-distinct matrices in (28) would select the distinet matzices (25) if we included, in the cet. M of MS conditions, if-then eondi- tions corresponding to the rules of (24). As pointed out in §8.5, the lack of a distinctness requirement is an advantage. ‘These examples, which illustrate the process of selection, complete the desarip- tion of the system of MS cousitions. We have seen how MS conditions replace MS rules as a deviee for predicting redundant, information in systematie pho- nemie matrices, In the next section we will discuss further details of the new system; we will also attempt to demonstrate its superiority over MS rules. PROPERTIES OF THE eySTEM 6.1. Tan ronsarime. We saw in §$ that one of the majorproblems in MS rule theory was to prevent the improper use of blanks; this problem was solved by imposing the true generalization condition on MS rules, Now that we have de- seribed MS condition theory, itis natural to ask whether or not improper use of blanks ean be obtained through MS conditions. It is easy to see that the answer to this question is no, by looking at the definition of the set M(U) by the MS conditions and at the process of selection of fully specified matrices from MCU) by dietionary matrices. Consider a set: M of MS conditions for some language. The function of Mis to define, from the cet U of all matrices, the eet M(U) of matrices which violate none of the constraints of the Ihnguage. However, since U consists of matrices which have no blanks in them at all, it is clear that this definition can invalve xno improper usé of blanks. Further i is olear that exch MS condition in M. must state 2 tre generalization about the morphemes af the language; violating this requirement would never lea to specious simplifications, only to an incorrect grammar. Thus, the true generslization condition of §33 is rtomatically met by ‘MS conditions. Consider now the process of selection from M((), Recall that, by the definition of dictionary matrix in a system of MS conditions, each dictionary matrix must select exactly one (fully specified} matrix from M(U). Since the set: M(W) eon- sists only of the possible morohemes of the language, it follows that each dietion- ary matrix must-contain precisely the araount of information needed ta identify fa morpheme GIVEN THAT TT 15 A POSSIRLE KOREMENS O THE LAXOUAGE. This, of course, is precisely what is desired. Tt means, on the one hand, that éictiongry rmattices need not contain extra specifications in order to meet. requirements such as distinetness or wellformedness, which we saw in §3 to be unnatural and un- motivated. On the other hand, it means that dictionary matrices cannot contain 30 few specifications that they are inadequately kept apart, for a matrix with too ow specifications would soleet. move than one mateix from M(U). 6.2, Tux vavuarion smasune. Each MS condition C in 2 tet M of MS con- ditions inn grammar must contribute to malting the set M(U) smaller; if C didREDUNDANCY RULES IN PHONOLOGY 431 not go contribute, it would he of no use in the grammar and would be disearded, We can soe how each MS condition decreases the size of M(U): the mote MS conditions there are, the fewer matrices are accepted by all conditions, and ‘M(U) is just the set of matrices accepted by every condition. Also, the syoaller the cet M(U) is, the emaller is the number of specifeations which have to be made in dictionary matrioes: when M(U) is small, then dictionary matrices have less to select from. Thus, for every MS condition C, we can talk about. how many feature specifications C allows us to save in any dictionary matrix D this is just the number of additional specifications that D would require to select a single ‘matrix from the set M’(U) where M’ is theset M with the condition C removed. Since we can talk shout the number of feature specifiestions exch MS eondition saves in the dictionary, it follows that we ean use the same kind of evaluation procedure as that used for MB rates, since this procedure is based on the number of feature specifications that each MS rule saves us in the dietionary.* For any subcet.M’ of the full set M of MS conditions in sny language, we ean talk shout the set M’(U). This set. contains the set M(U) snd also those matrices, rejected only by the MS conditions in M but not by those in M’.’This suggeets a ‘way of handling exceptions to the MS conditions. If a morpheme is an exception to the MS condition C., then we can simply mark this mompheme [—Condition, yl in the dictionary, and agree that a morpheme so marked will select not from, the set, M(U), but rather from M'(U), where M'is the set M minus the condition ©,, Selecting from the larger set. M'(U}) will require that a dictionary matrix be more fully specified, and this is exactly as desired: exceptional morphemes should require fuller specification, The trestiment of morphertes which are exceptions to more than one MS condition ean be handled in & perfeetly analogous fashion, 5.3. Tue, NEED FOR PostrivE coNDrTIONS. We give here an example of a, kind ‘of morpheme structure whieh is highly unlike that of any natural language, but which is deseribable in quite natural way by MS rules. Such an example is, surely devastating to a theory of MS rules. Thus, consider a ‘language’ whieh has the following four types of momphemes: aa LCvVG OVGL GLC GLOV where I= Liquid = [+Consonsatal, +Vocalic], C = Consonant = {-+Conso- nantal, ~Vocalic], V = Vowel = {—Consonantal, +Voealiel, G = Glide [—Consonantal, —Vocalie]. That is, L, unless final, is always followed immedi- ately by C; C, unless final, is always followed directly by V; ete, Such eyclical structure surely does not occur in natural Languages. However, observe bow naturally a set of MS rules can deseribe this structare: (89) [aConsonantal] lt : {-eConsonantal] + See Halle 1964b:938-40, Thia evaluation procedure is the meane by which we determine what the signideant genralizations of a language are, Of course to say, a8 this evalustion measure docs, that feature values in rules have theaame coe as feature values in dictionary ‘matrices ia strong claic, and ia by na means obviously correct. Perhaps aomee other ratio ‘of importance of feature values in rules to feature values in the dictionary fe needed,432 LANGUAGE, VOLUME 43, NUMBER 2 (1967) (40) [eVocalie) H L [-eVoealie) ‘The MS rues (39) and (40) completely charscterize the four structures in (38) snd only these. Indeed, the dictionary eniries for these four structures can be represented ag follows for the features Consonantal and Vooalic: + [Sse] [rere] Levooaiel [d+ cy 7 LeGeanee} rem] Evel E+ + [iceman] [rome] Coveeniol [ ] + + [ogee] [remem] Lovee [ ] + ‘The rules (29) and (40) would fil in all the other fenture values (for the two features considered here) in ench of the forms in (AL). “Thus we see that MS rules treat. the unnatural situation (38) in a simple way. On the other hand, no Postrive MS eondition eould possibly deseribe the struc- ture in (38) (other than by listing in a four-way disjunetion each of the possi bilities, a solution whieh must be disallowed sinee, using similar techniques, post tive conditions can describe ary situation). However, it is true that the theory of MS conditions could handle the structure in (88) using rp-rermw conditions equivalent to (32) and (40) (sinee, as we have seen, if-then eanditions end MS rrufes are always intereonveriible), and this seems to take the force out of our argument, Yet this can be avoided if, in the theory of M8 conditions, we Rn quire that positive eonditious be used in stating restrictions on syllable structure, that is, in stating restrictions involving the features Consonantal, Voealic and petliaps Obstruent. This requirement is motivated by the fact that these features fare typically interrelated in different ways than the other features, a fact: which implies that they should be treated formally in different ways." In susunary, then, we may accept the above argument: that positive MS conditions suecced ‘where M8 rules fail, This angument seems to be substantiated by the fact that the Syllable structure? of many languages ean only be described by a relatively complex set of MS rules whieh not only miss genoralizations but which also ean- not capture any restrictions on the length of morphemes which may exist. Po tive MB conditions on the other hand, seem ideally suited to deseribing such syllabic structare, 5.4. Tue sump vor weoartys conotztons. Ta this section we show that nega- tive MS canditions, but not MB rules, can provide a solution to a well-known problom in Proto-Indoeuropean (Lebmann 1952:17). Morphemes of the type ‘n'Thiadiference in the behavior of diferent features ig clearly tke kind of forms prop- arty! which, as we noted i §24, seo he necessary in axy non-afshoe characterintion of Hcrarchy among the features, Perhaps we could even Jook for a formal definition of eae {eatue is tors of the particular way % bebaves I morpheme strucire und inthe P rules, ‘Thie would be similar to the formal definitians of Nous and Veeb surgeated by Chenasky {1068:16-6), in terms of the diferent ways noting and vorbs behave with reget to sub categorizationREDUNDANCY RULES IN PHONOLOGY 433 {obstruont, vowel, obstruent] were severely restricted; using ¢ as a sample vowel, 77} dh as sample initial obstruents, and i d dh as sample final obstruent, those restrictions ean be summarized in the following table: Ypedh ped pet (42) bedh —tbed_ et bhedh hed = *bhet ‘That is, if the intial consonant is 2 voiceless non-aspirate, then the final eonso- nant cannot be sn aspirate, ete. Consider the negative MS condition wo ~+ [eine] [ ] [aise] + ‘This condition rules out all forms whorw the voicing of the initial segment dis- agrees with the aspiration of the final segment and where the voicing of the final segment disagrees with the aspiration of the initial segment; that is, it rules out all forms which show warmer an mnof the aspiration of the final seg ment to the voicing of the initial segment non an assimilation af the aspiration of the initial segment to the voicing of the final segment. And, in fact, the forms rujed out are the dasired ones, Le. the forms starred in (42). Thus the negative eondition (43) accounts exaelly for the facts summarized ia (42). Further, it is not hard to see why it is in principle impossible to write MS rules to account for these facts. This follows since, as noted above, the generalization involved in (42) is that ermmen one on another (or possibly but not necessarily both) of two assimilations must take place (that is, that all forms in which neither of the assimilations tales place are ruled out); yeb, even if we wrote an MS rule which corresponds to each of these two assimiations, we would have no way of stating that at least one but not necessarily both of them must apply. This of course necessary that many examples of nogative conditions be given. if their introduction into the theory is to be motivated. An important line of re- seareh would bo to discover to what extent negative conditions describe situations which oceur in natural languages. 6.6. Tux note oF nzanas, MS rules and MS conditions have one character istic in common; in each we have ehosea to formalize tho notion of ‘redundant? feature value by saying that a feature value is redundant in a eortain environ- ‘ment, just in ease the feature value, when left blank in this environment, can be determined by some finite procedure. In the case of MS rales this procedure con- sists of applying all the rules that fit this environment, until finally one of these rules inserts the value of the feature in question; in the case of MS conditions, ‘the procedure consists simply of looking at the set of matrices ‘nocepted’ by the MS condition, since every matrix in this set which has the environment in ques- tion aleo has the propor value of the feature in question. Thus, ‘redundant’ for as hhaa meant ‘can be left blank and lator predicted! Suppose we proceed differently and abandon the idea of having a dictionary matrix for each morpheme which is less fully specified than the systematic pho- 2 The condition (4) sao rules out forms such as pheth ut this all rght sinee voicelens sopirates met be ruled out anyway.a4 LANGUAGE, VOLUME 48, NUMBER 2 (1967) ‘nemic matrix of the morpheme. In fact, et us abandon (but for the moment only) the whole notion of dictionary matrix, and talk only about the fully specified sys- tematic phonemic matrices, regarding these both 2s the matrices to be listed in the dictionary and as the matrices which enter the P rales. Stil, let us retain in each gramamar the ect of MS conditions thet state (in the manner shown in the ‘previous section) the constraints that exist on the morphemes of the language. Now at first glance ié may seem that we cannot do this, since we could have no way of knowing what the right set of MS conditions for a given Inuguage is. As Jong as thore are dietionary matrices with blanks, then we have an evaluation measure which tells us what the best set of MS conditions is; it is, essentially, the shortest set of MS conditions that allows us to leave the greatest aumber of blanks in dictionary matrions (but cf. fn. 36). IF there are no dictionary entries ‘with blanke, then this measure means nothing. Looking deeper into the matter, however, we seo that even if we are to talle ‘only about sets of fully specified systematie phonemic matrices sad sets of MS conditions which state the constraints on these matrices, itis still possible to talle about the ‘right’ set of conditions for a given set of matrices; that is, there is still ‘4 meaningful evaluation procedure. Suppose we define the wricerr of MS condi- tion C, with respect to a systematic phonemic matrix 8, as the maximum number ‘of feature specifications thet could be removed from $ provided only that, from the resulting reduced mastrix R and the MS condition C, we be able ta recon- struct 8." Clearly, if an MS condition C docs not apply to @ systematic pho- nemie matrix §, then the weight of C with respect to $ is zero.# I? an if-then ‘MS condition C applies to a systematic phonemic matrix 8, then the weight of C with respect to 8 is the number of specifications in T(C), the ‘then’ part of C. a negative MS condition © applies to a systematic phonemic matrix 8, then the weight of C with respect to 8 is always the number ‘1’. If a positive MS con- dition C applies to a systematic phonemic matrix &, then the weight of C with respect to $ is the number of specifications in P(C).** The crucial idea is that the weight of an MS condition with respect to the various systematic phonemic matrices tells us how general a statement this condition makes about. these matrices. This notion of generality of MS conditions (and, previously, of MS rules) is really the important one, and with it we ean define an evaluation measure for MS conditions that does not count blanks in dictionary matrices. ‘This is done as follows, The meaning of ‘reeonsteuct” should be elear given the diseusion in $16. Tobe explicit, 8 should be the unique matrix aleeted by R from the ect C(U} (the eet ofall ratcices in avcepted by ©). ‘The nation of a MS condition applying (or not spying) to s aystematie phonessio mr Sig inuaitivey clear." be precise, we ean oxy: (1) an it-then condition C applies to 8 justin cane T(C) isa s-matrix of $; @) a to 8 just PC) is aonb-amatrx of 8; nogative condition C apple to'S joan case we ea the value of ows feature in NC) (any one feature) aud arrive at a matrix N‘(C) which is & sub-matrix of 8. ‘Actually, this a the number m; of apecication in P(C} minus the number ms of apesi- fications needed to keep P&C) diatinet rar all other positive conditions, if any, having tke fue number of colurors as P(C), sineedlearly we can eave at mest mins specifications i ‘amairin using P(C). See (96) and the later diceuasion of (8) in 46.REDUNDANCY RULKS IN PHONOLOGY 435 Suppose we have a language I. with the set SP of systematic phonemic matrices and with » proposed MS condition C, Define the Genesatiry Ixpex of C with respect to SP a8 the sum of the weights of C with respect to each member of SP. Also, define the smarzsceyr cost of C as the number of festure values needed to state C, Then we can agree to include C in the grammar just in case its state- ‘nent cost is Tess than its generality index (cf. fn. 86). This provides a meaningful evaluation measure. Further, itis not hard to see that this measure gives essen- tially the same results as the ususl evaluation measure, which depends on the existence of dictionary matrices in which the redundant feature values have been, Joft blank, ‘The proposed solution also removes a certain kind of arbitrariness whieh was inherent in the practice of leaving redundant feature values blank in the die- tionary. For example, if, in some environment F, the value [+] implies the value +g} and the value (+g) implies the value [+i], then it would be arbitrary which value we actuslly chooee to indicate in the dictionary; since the MS conditions ‘would rule out both of the feature combinations [+f, —g] and [—f, +2] in the environment E, we would need to indicate either [+f] or [+g] but not both ‘The arbitraciness stems from the decision to commit. ourselves as to whether f is predictable fram g or g from f, this decision being involved in any solution where redundant featare values ate left blank in dictionary matrices. Actually, however, this decision is not well motivated, since the correct statement may be simply that the values of f and g ato interrelated. In fact, we can profitahly draw a gen- eral conelusion regarding redundaney frora this discussion. That is, to say that a certain fully specified matrix is highly redundant in some language is actually to say that many of its feature values are interrelated in ways determined by the constraints of the Language, atd it is simply the statement of these constraints {in the MS conditions) which constitutes the most natural characterization of the redundancy of tho language. Once these. constraints have been stated, it is true that they may be utilized (as deseribed in §4.6) in givingdictionary representations their most economical form; but this is secondary fact, and these redundancy- free representations play no real role in a theary of redumdaney.$ # To make this evaluaiZon procedure meaningful, we must make the abvious requirement that eaak MS condition actually comtaibaces to making she see M(U) smaller, and does nat merely repeat the efect of other MS conditions. “20f course in MS mux theory there is this sume arbitrariness, but here this fs based on ‘more serious arbitrariness, namely che choice between (41> (ral / Band (al [11 / E'as the rule for formulating the constraint in question. + Lowe aapesial debt to Motra Halle, who gave the encouragersent and direction neces: sary for the woting of tie paper. Algo extended eotaments on nn earlier dratt by Barbar Hall Pariee, David Stampe, and John Rass have led ta considerable improvements. Finally, Lam grateful to Manfred Biersigch, Richard Carter, Noam Chomaty, Erica Gaseia, Kalon Kelley, Chin.W. Kim, Paul Kiparsky, George Lakoff, G. Hubert Matthews, P. Stanley Peters, and Pail Poatal for commenta and discussion, ‘This work wax supported in part hy the Joint Services Plectronies Prograta wader Con tract DAa#-000-AMC-320005} in pate hy the National Scienee Foundation (Grant GP- 12496), the National Tatituten of Health (Grant MH-O4737-0), the National Aeronattice and Space Administration (Grast 85C-490), and the US, Air Fores (OSD Conérsct AFIS (618-2887).436 LANGUAGE, VOLUME 43, NUMBER 2 (1967) REPERENCES Cannon, Jou B, 1961. The study of language. Cembridge, Masa, Harvurd University Press (Gaemny, Coty. 1961, On buman communication. Neve York, Wiley. Grrowster, Noak. 1957. Review of Jakobson and Halle, Fundamentals of language, LIAL 2.23442, 1964. Current iesues in linguistic theory, The Hague, Mactan, 1965. Aspeets of the theory of syntax. Cambrige, Mass., MIT Presa, ——, apd Gronae A. Mitixa. 1963. Tntpoduction ‘© dhe (armal analysic of natural languages. Handbook of mathematical peycholagy, ed. by BR. D. Luce, R. Bush, and E. Galanter, 2260-G21, New York, Wiley. Grsacon, H. A. Ja, 1961. An introduction ta deseriptive linguistios, New York, Holt, Hue, Mounts: 1959, The sound pattern of Russian. The Hague, Moavon. — lu6da, On the besee of phonology. The structure of language, ed. by J. Katz and J. Fodor, 824-33. Englewood Clie, N. J., Preztico-Hall — 19%tb. Phonology in generative grammar. The structure of language, ed. by J. Katz and J. Fodor, 384-62. Englewood Cifls, N. ., Prentice-Hall, ocrer, Cxaniies F. 1968, A course in modern linguistics. New York, Macnrillan, Iaxonsox, Rowan, Gowan Fut, and Morais Hate. 1951. Pretirainaries to speech analysis. Cambridge, Mass., MIT Press. and Monnis Hatt. 1954. Furdarontats of language. The Hague, Moutan. Iapuans, Woenen P. 1952. ProtoIndo-European phorology. Austin, University of ‘Texas Press and LEA. Lucuanen, TM. 1963. 4 rote on the formstion of phorological rales. Quarterly progress report, no, 68, 187-89, Research Laboratory of Eleetranice, MIT. ‘McCawony, Jams. 1964, The aecentual system of standard Japanese. MIT Ph.D. thesis, ‘Mrnure, Groner, and Noss Crowsky. 1963. Finitary mode's of language users, Hand Ibook of msthemstical peychology, ed. by R. D. Luce, R. Bush, arc E. Galanter, 2AIG-OL, New York, Wiley.