This book compares Distributed Morphology, a localist and serialist theory, with a global and parallel version of Optimality Theory in terms of their predictions and implications for phonologically conditioned allomorphy, phenomena that the two theories treat in very different ways. Distributed Morphology is a theory about morphosyntax, which makes very specific claims about possible interactions between phonology and morphosyntax. Optimality Theory has mostly been employed as a model for phonology, but has often focused on the interaction between phonological operations and exponence. Embick analyses very complex cases, taken from a wide variety of languages, and some of the issues he addresses are of very general interest, such as the division of labour between phonology and morphology and how these two modules interact. The book is in general very clear, and it is obvious that a major effort has been made to make all the claims and observations very explicit. In this review I concentrate mostly on those aspects that have a close relationship to phonology, and leave aside some of those that are more specifically related to morphosyntax.

The book is divided into an introductory chapter (Chapter 1), which gives a fairly detailed outline of the contents of the rest of the book, and two clearly differentiated parts. Part I (Chapters 2 and 3) presents the version of Distributed Morphology argued for in Embick & Marantz (Reference Embick and Marantz2008). Part II (Chapters 4 to 7) is devoted to a presentation of a number of arguments in favour of the localist theory and against a parallel and global version of Optimality Theory for dealing with phonologically conditioned allomorphy. Allomorphy implies competition between exponents, and the two theories discussed have very different conceptions of where and when competition can take place.

The essentials of the version of Distributed Morphology presented in Part I, which Embick calls the ₁-LIN Theory, are as follows. The objects that enter into syntactic relations are roots and functional categories. Roots are category-less; they become what are commonly called verbs, nouns or adjectives when they merge with functional categories like v, n or a. Even simple words like cat are complex objects, in this case consisting of a root, a category-defining head n and other relevant functional categories. In the syntax, roots include phonological information, in contrast to functional morphemes, which receive an underlying phonological form through the process of Vocabulary Insertion. This latter process targets only terminal nodes, and never complex objects (like words or phrases). Since allomorphy involves competition between exponents of a morpheme, Embick (following Embick & Marantz Reference Embick and Marantz2008) concludes that competition can be found only at the point of Vocabulary Insertion. Each time a category-defining head is merged, it triggers Spell-Out of all the material that is in its complement domain (cyclic Spell-Out), meaning that allomorphy is not only sensitive to linear adjacency (achieved in this model through a concatenation process), but also to whether or not a potential target and its context belong to the same Spell-Out domain (hence the term ₁-LIN for the theory). Since roots already appear in the syntax with their phonological form, differences in the root of forms like kept vs. keep must be obtained not through competition between different exponents of the root, but through readjustment rules, i.e. phonological rules restricted to specific morphosyntactic contexts, as already proposed in Chomsky & Halle (Reference Chomsky and Halle1968) (for some discussion of root allomorphy see Bonet & Harbour, forthcoming). For allomorphy sensitive to phonology, the phonological context is stated in the Vocabulary Item as an idiosyncratic lexical condition. For instance, the Korean genitive suffix has the exponent /-i/ in the environment /C__, and the exponent /-ka/ in the environment /V__. As a result, the competition between the two allomorphs is resolved at the point of Vocabulary Insertion, and not later. In Chapter 3, fairly complex examples of allomorphy are discussed in the context of the localist theory, and several kinds of PF operations are justified, like pruning of specific category-defining heads to obtain the necessary adjacency between two terminal nodes, local dislocation (a type of affixation) and readjustment rules. With respect to phonologically conditioned allomorphy, the theory does not exclude sensitivity to either the underlying or the surface form of the context. The choice would depend on the order between Spell-Out of an outer cycle and phonological operations in the inner cycle.

The view of the morphology–phonology interface just sketched is contrasted with a parallel and global version of Optimality Theory that argues for multiple inputs in phonologically conditioned allomorphy (Embick cites in this respect Kager Reference Kager and Kleinhenz1996 and Mascaró Reference Mascaró2007, among others). In these models, described in Chapter 4, the input for the Korean genitive suffix mentioned above contains the two allomorphs /-i/ and /-ka/. Gen provides candidates that contain one of these allomorphs, and ranked universal constraints like Onset and No.oda determine which of the two allomorphs is favoured in each context. In this case, competition is resolved by the grammar; it is not stipulated in the lexical entry.

As mentioned above, Part II of the book is devoted to the evaluation of the two models and the incompatibility between their views of phonologically conditioned allomorphy, an incompatibility that Embick calls the schism (p. 11). One criticism that can be made of the ₁-LIN Theory is its ‘putative loss of generalization’, to use Embick's own term (p. 21). In the case of the Korean genitive suffix, just stating the phonological context for each allomorph misses the observation that the choice in each case improves syllabification; it gives rise to a more unmarked syllable structure. One of the arguments against the multiple-input model, previously put forward by Paster (Reference Paster2006) and Bye (Reference Bye, Blaho, Bye and Krämer2007), and to which Chapter 5 is devoted, is that many cases of phonologically conditioned allomorphy are not so clearly optimising as the Korean genitive case suffix, and in some cases the choice of phonological context seems totally arbitrary. The existence of such cases would invalidate all analyses with multiple inputs. But what we pretheoretically call phonologically conditioned allomorphy is a set of observationally related phenomena that need not have a single explanation. One of the examples discussed in this chapter is the Haitian Creole definite article, a suffix with two exponents: -a, which appears after stems ending in a vowel, and -la, which appears after stems ending in a consonant. This looks like an anti-optimising distribution. Embick criticises the analysis of the case in Bonet et al. (Reference Bonet, Lloret and Mascaró2007), who propose that the two allomorphs are lexically ordered (-a>-la) and that a universal constraint called Priority penalises candidates with the non-preferred allomorph, in this particular case -la. Embick's main objection to this proposal is that positing lexical ordering between allomorphs is not less stipulative than Vocabulary Insertion. The difference is that in the ₁-LIN Theory all cases of phonologically conditioned allomorphy have a stipulated specific context, while in the alternative theory the preference for certain allomorphs over others derives from two factors, grammar (i.e. the set of ordered constraints) and a very restricted lexical property, the ordering of allomorphs. The issue raised with the argumentation in this chapter is actually more general and can be compared to a long-standing question in phonology: does the fact that many phonological alternations are exceptional mean that all phonological alternations must be attributed to lexical idiosyncrasies?

Much stronger arguments against the multiple-input parallel view are presented in Chapter 6. One of them is related to opacity: the multiple-input parallel approach predicts the right choice of allomorph only when the phonological context is surface-true, not when it is found only in earlier levels of representation. Embick illustrates the problem with the Turkish 3rd singular possessive morpheme, where the context of insertion for one of the allomorphs is rendered opaque by a process of velar deletion (more examples can be found in Nevins Reference Nevins, van Oostendorp, Ewen, Hume and Rice2011, for instance). Opacity is a major challenge for parallel versions of Optimality Theory, and serial versions of Optimality Theory can handle such cases more easily. Alternative analyses of the data must be proposed for parallel versions of Optimality Theory, which is desirable in some cases, because existing analyses are questionable. Another interesting argument is illustrated with cases from Latin and Saami, where the choice of allomorph depends on foot parsing. In Latin, for instance, the perfect has two allomorphs /u/ and /s/. According to Mester (Reference Mester1994), the occurrence of /s/ avoids having an unfooted syllable in the word. For the verb monēre ‘to warn’, the perfect [mo.nu]<ī>, with the /u/ allomorph, does not have any unfooted syllable (the last syllable, in angled brackets, is extrametrical; square brackets are used to delimit feet; the allomorph is underlined). However, the use of the same allomorph for a verb like augēre ‘to grow’ would have as a result an unfooted syllable, the only foot in the word consisting of a single heavy syllable: *[au]gu<ī>; the choice of the alternative allomorph /s/ avoids this problem: [aug]<sī>. The prediction for the multiple-input parallel approach is that this optimisation should be sought even if further suffixes are added; one would therefore expect to find alternation between the allomorphs depending on the foot structure of the whole word; this prediction is not borne out: for the Latin pluperfect one would expect the perfectly footed *[au][gu.e]<ram> instead of the existing [aug]se<ram>, with an unfooted syllable. For serial models like the ₁-LIN Theory, these cases should not be a problem, because the perfect suffix is inserted before outer suffixes, and only the weight of the root would be relevant at that point. Parallel theories must resort to output–output constraints to account for these facts. The issue raised here arises almost only in cases where the context for allomorphy has to do with foot formation. In most cases of phonologically conditioned allomorphy the context is so local (having to do with phenomena such as place or manner assimilation between adjacent segments, or being in onset or coda position) that the addition of further affixes will not alter the context of the allomorphs.

Chapter 7, the last chapter, includes some conclusions and some questions about future research, for instance whether the cyclic domains that are required for syntax, phonology and semantics are the same, which would be the best result. A far-reaching conclusion of the book is that the regularities found in phonologically conditioned allomorphy must find an explanation outside the grammar, as a product of language acquisition, historical change or phonetics; surface generalisations of this type are not to be accounted for in the grammar. Although this statement could end up being the right conclusion, the arguments presented in the book are not totally convincing, as pointed out in several places in this review.