Progress in any scientific field depends on its practitioners being in at least partial agreement about what the object of study is, what kinds of empirical observations fall within the scope of inquiry and by what criteria a theory or analysis may be judged to be valid.Footnote * In The phonological enterprise, Hale & Reiss stake out their territory in the diverse landscape of current theoretical phonology by presenting engagingly argued positions on these fundamental issues, based on the formal, rule-based, strongly rationalist approach well known from their previous work. This volume, a revised and expanded collection of their papers and presentations, provides a handbook of sorts on their phonological world view. It is organised into an introduction plus four parts, which treat various topics but are held loosely together by an overarching concern with the question: what should phonology look like?

The introduction sets out the main premise: there exists a valid object of study which is the phonological component of the human language faculty, shaped largely (and perhaps entirely) by the innate Language Acquisition Device (LAD), which determines ‘how the learner assigns analyses to data provided in the environment’ (p. 15). Phonological theory should characterise the LAD in a way that accounts for the range of cognitively possible or ‘acquirable’ languages rather than the set of attested languages per se (the latter being an incidental sample skewed by diachronic and physiological factors irrelevant to UG), and the correct phonological analysis of any particular phenomenon will be the one that is cognitively real for speakers. The book is primarily aimed at generative phonologists who agree with these statements. The problem, as Hale & Reiss see it, is that ‘such scholars often offer explanations and propose models and analyses which are largely inconsistent with … assumptions which they nevertheless continue to avow a belief in’ (p. 24; emphasis in original). They introduce two data sets, one from Georgian lateral allophony and one from Catalan manner alternations, to illustrate what they dub the ‘Linguist's Problem’: given that the language faculty is a mental object not amenable to direct observation, on what basis – if any – can we know which of several empirically adequate analyses is the one in the speaker's mind?

In the rest of the book, Hale & Reiss address this problem by aiming to establish which types of evidence can legitimately be considered – and based on what we know so far, what kinds of theoretical claims can logically be entertained – by anyone who (as presumably most generativists do) takes phonology to be the study of symbol-based mental phonological computation.Footnote ¹ Part I comprises three chapters examining ‘Phonological UG and acquisition’, a theme of obvious relevance to the Linguist's Problem: if we know how children analyse data, we will know what analyses they end up with. In Part II (‘Resisting substance abuse in phonology’), Hale & Reiss seek to delimit the set of observations for which phonology should be held responsible, and criticise work which they argue erroneously attributes extragrammatical phenomena (variation, articulatory and perceptual considerations, typological facts) to the grammar, building correspondingly misguided innovations into phonological theory.

Part III (‘Some aspects of Optimality Theory’) argues against the use of constraints in phonology, whether within a rule-based or optimality-theoretic framework. It also claims, based on a detailed examination of evidence for Output–Output Correspondence constraints, that Output–Output Correspondence is an unjustified mechanism and should be eliminated from phonological theory. These chapters are less closely tied in with the book's central themes and I will not discuss them extensively here, but see Krämer (Reference Krämer2010) for some comments. Finally, the very brief Part IV (‘Conclusions’) selects the analysis of Catalan manner alternations that the authors believe, on the basis of the principles developed, most closely matches cognitive reality. After many chapters of arguments against other work, they also provide a positive example of what might constitute theoretical progress: namely their proposal for a Unified Interpretive Procedure, a rigorously formulated and hypothetically universal algorithm for how phonological grammars interpret the structural descriptions of rules.

Perhaps the book's strongest point is its insistence on clarity of assumptions and goals. One would think that the field could only benefit if everyone was as explicit as Hale & Reiss are about what exactly we are trying to model, and why we think we are justified in the way we go about it. The main weakness is that some of the authors' assumptions are potentially controversial, but not always examined sufficiently before large claims are constructed on them. As I will argue below, these assumptions are not a priori logical truths or even ‘relatively mundane’, as suggested by Hale & Reiss on page 24; rather, they involve very real and still unresolved empirical issues. Arguably, Part I can be read as a series of meditations on the consequences for phonology of the ‘no negative evidence’ assumption, while Part II elaborates on what phonology would look like if Occam's Razor is taken not as a heuristic, but as a scientific principle permitting the discovery of actual facts (cf. Kapp Reference Kapp1958). It is to the authors' credit, though, that the argumentation is generally clear enough that the reader is often able to pinpoint exact loci of agreement or disagreement. In what follows, I will discuss some of their main claims and conclusions.

In the first three chapters on the theme of acquisition, Hale & Reiss invoke the familiar logic of the Subset Principle: since overgeneralisation errors can never be corrected through positive evidence, and learning from negative evidence is assumed to be impossible, the only path to successful acquisition is via minimal generalisation over input data. In Chapter 2 (‘The subset principle in phonology’) they argue that the appropriate level of generalisation is that of distinctive features, not of segments, since the former and not the latter are the representational primitives that the learner assigns to input. The consequence of this would be that children's phoneme inventories shrink rather than grow between earlier and later stages of acquisition: infants must initially store representations as maximally specified for phonological distinctive features, and they only gradually learn which distinctions – i.e. features – can be ignored in the ambient language. This idea is reminiscent of classic work such as Werker & Tees (Reference Werker and Tees1984) which shows reduced perceptual sensitivity to non-native phonetic contrasts in older as compared to younger infants.

Hale & Reiss argue compellingly for the logical necessity of innate categories for parsing the acoustic signal; in so doing, they shine a bright light on questions of what (and particularly, how fine-grained) these categories are, and how directly or indirectly they relate to contrast in any given phonological system. In their proposal, the innate parameters used by the child to assign cognitive representations to raw physical input are equated with phonological distinctive features of the kind used to capture phonemic contrasts and natural classes. While the authors acknowledge that these would not necessarily be the exact same features hypothesised by early 21st-century phonologists, they explicitly reject (p. 41) the idea that there is a parallel level of phonetic representation over which categories are formed and adjusted. This is why children's phoneme inventories would have to go from a full-specification stage to an optimised stage where allophonic variants have been collapsed: if children perceived each token as one of a small number of phonemes, and had no phonetic memory over which to create new categories, they would continue parsing each new token as an instance of a known phoneme and a new category could never be acquired (see also page 86). As Boersma (forthcoming) points out, though, it is unclear how the phonologically relevant features could be assigned immediately upon perception of speech (e.g. in cases such as voicing where phonetic correlates differ on a language-specific basis). Of course, it may well turn out that different phonetic correlates correspond to different features after all (see Petrova et al. Reference Petrova, Plapp, Ringen and Szentgyörgyi2006), but there must still be a mechanism for dealing with ambiguity in cases where the same correlate, such as a certain vowel quality or zero VOT, has multiple potential phonological representations; it could be interesting to attempt to extend Hale & Reiss's proposal with a compatible mechanism of this type.

A different argument made (p. 41) is that adults supposedly do not store phonetic representations (except only occasionally and non-linguistically), so this ability should not be posited for children either. This assertion seems to run counter to evidence that mature speakers consistently store and use fine-grained phonetic detail in linguistically systematic ways (e.g. Labov Reference Labov1994, Johnson Reference Johnson, Johnson and Mullenix1997, Pierrehumbert Reference Pierrehumbert, Bybee and Hopper2001). The authors might be better served here to draw on relevant psycholinguistic work such as Polka & Werker (Reference Polka and Werker1994), which suggests that infants distinguish fewer phonetic contrasts than adults while they are in the process of forming native phoneme categories, and reacquire phonetic sensitivity only later on. On the other hand, Fikkert & Levelt (Reference Fikkert, Levelt, Avery, Dresher and Rice2008) present evidence that children's phonological representations become gradually more rather than less specified; see Fikkert (Reference Fikkert2005) for an alternative along these lines to Hale & Reiss's scenario for the acquisition of a vowel system.

A drawback to the discussion is that, throughout, the status of the ‘no negative evidence’ assumption is somewhat ambiguous. On page 42 it seems to be treated as provisional, on the idea that a learning theory based only on positive evidence would also work if negative evidence was added, but a theory making crucial use of negative evidence would not work if it turned out that such evidence is not available to learners. Elsewhere however, Hale & Reiss wield this assumption to eliminate as logically untenable any learning path that would make use of negative evidence, so it ends up serving a slightly different function. This issue comes up again in the dominant assertion in Chapter 8 (‘Against constraints’) that it is impossible to acquire prohibitions, i.e. constraints, through induction over positive data. Ultimately, however, there is an empirical question – can people acquire constraints through induction over positive data? – that can probably be addressed by methods better than argument from poverty of imagination; see Marcus (Reference Marcus1993) and Clark & Lappin (Reference Clark and Lappin2011) for examples of work that takes this issue seriously.

In any case, the extensive discussion of the Subset Principle in Chapter 2 serves as a set-up for Chapter 4 (‘The Georgian problem revisited’). There, Hale & Reiss show how to choose between extensionally equivalent analyses of Georgian laterals by using the principle that learners formulate triggers and targets with the most specific set of distinctive features that encompass only and all known instances of application. Interestingly, generalisation beyond the data is still possible under this system, since patterns can be extended to novel segments having the specified featural description.

Elsewhere in the acquisition section, Chapter 3 (‘Competence and performance in phonological acquisition’) addresses the epistemological question of how we can know about the state of children's phonologies. The basic argument is that children's productions, even (and in fact particularly so) in systematic aspects such as chain shifts, quite possibly reflect performance factors rather than mental grammars and thus cannot be relied on as sources of evidence. In light of this discussion, the chapter includes a detailed critique of Smolensky's (Reference Smolensky1996) work on phonological acquisition within Optimality Theory, and an alternative proposal by the authors. Hale & Reiss's idea, reminiscent of Chapter 2's minimal generalisation arguments, is that all faithfulness constraints must outrank all markedness constraints in any plausible initial state of an OT grammar (i.e. if one does OT – not that the authors do – this is how it must work). Markedness constraints are promoted in the ranking hierarchy only as the child receives evidence for non-trivial mappings between underlying and surface forms. Hale & Reiss close the chapter by suggesting (p. 93) that in the current state of knowledge about performance issues it is difficult to empirically investigate children's linguistic competence, although it is important that methods be developed. Overall then, Hale & Reiss present in Part I results derived by deductive reasoning from a number of explicitly assumed cognitive science principles, while arguing that theories based on unreliable data (i.e. observations of child production) do not contribute to our knowledge of the LAD.

In Part II (‘Resisting substance abuse in phonology’), epistemological questions move from a supporting role onto centre stage. The focus is on the issue of which empirical facts do – or rather, in most cases, do not – constitute evidence for the nature of phonology. Chapter 5 (‘Isolability and idealization’) contains a meticulous and illuminating discussion of the need to isolate the effects of mental grammar from those of the separate systems that it interfaces with (such as the transduction from cognitive representation to physical output) in order to get at the phonology-internal principles that are the object of study as laid out in the introduction (especially Fig. 1.1 on page 4). Specifically, Hale & Reiss argue against claims that phonology must be able to compute over graded representations like frequency statistics (as in Hay Reference Hay2000) and produce variable, probabilistically generated outputs (as in Bresnan & Deo Reference Bresnan and Deo2001; see Bresnan et al. Reference Bresnan, Deo and Sharma2007 for a published version); both of these proposals would pose problems for traditional symbol-based computation. The authors' major concern is lack of clarity on whether stochastic models of variation are meant to be models of speakers' linguistic competence, cognitively irrelevant data-fitting exercises or explicit models of the interaction of grammatical and extragrammatical factors.

In the latter two cases, stochastic models of grammar would not fall within the scope of generative phonology, narrowly defined. As for the first possibility, Hale & Reiss assert that there is no need for grammars to generate variable output: they ‘can't imagine what evidence could not be accounted for’ (p. 139) by treating all sociolinguistic variation either as dialect mixing, with separate but parallel grammars producing different outputs, or as superficial dialect imitation via a ‘postprocessor’ that converts the phonological output of the native grammar. Here, in so quickly dismissing the possibility that a single grammar can have multiple outputs, they fail to engage with any of the literature on variation in phonological theory (see e.g. Hinskens et al. Reference Frans, van Hout and Wetzels1997, Anttila Reference Anttila, Chambers, Trudgill and Schilling-Estes2002). This is unfortunate, since Hale & Reiss have no theory of how a language learner is supposed to assign variable forms to separate grammars upon exposure to the patterns of variation that arguably characterise any single dialect of any language (by any definition of ‘single dialect’ that does not circularly ban variation); it presumably cannot be that a separate grammar is posited for each and every shade of social meaning. To be fair, the opposite issue is still largely unresolved in other work, namely how or why a child would decide to attribute variable outputs to the same grammar rather than to different ones. Still, a more interesting question might be what kinds of variation should or should not be accounted for by the phonological grammar at all, as opposed to extragrammatically (see Coetzee & Pater, forthcoming, for a survey of recent work, some of which postdates the book).

After eliminating variation from the empirical purview of phonology, Hale & Reiss go on to argue in Chapters 6 (‘Against articulatory grounding’) and 7 (‘Against typological grounding’) that phonetic substance does not influence how phonological representations are treated computationally within the grammar; in other words, that considerations of e.g. articulatory effort are not contained in the LAD. They observe that learners should be capable of acquiring phonological patterns without the aid of phonetically grounded constraints, by simply inducing generalisations over input data. They go on to criticise phonetically grounded constraints such as Steriade's (Reference Steriade1997) Fortition (‘Consonants must be realized with increased closure duration at the onset of stressed syllables’) as being incoherent insofar as they refer to properties of actual physical objects rather than the mental objects over which grammatical computation must by definition occur. While they mention that such constraints could instead refer to detailed aspects of mental representations in conjunction with the speaker's knowledge of their physical manifestations, they do not explore this idea further, as it would presumably go beyond the scope of the book to construct a detailed enough story on the origins of these constraints to enable an evaluation of cognitive plausibility (see page 153). Lastly, cross-linguistic tendencies for functionally motivated phonology are attributed to the process of language change rather than to the cognitive possibilities of synchronic grammar.

The contribution of Part II to the Linguist's Problem, then, is to eliminate potential analyses that make reference to phonetic substance and related notions such as markedness. However, it should be noted that the elimination of such analyses rests on Hale & Reiss's use of Occam's Razor not as a methodological principle to guide the formation of null hypotheses, which can then be subjected to empirical confirmation, but seemingly as an armchair method for deducing truths about the natural world. On page 169 they write: ‘Phonology is not and should not be grounded in phonetics … Duplication of the principles of acoustics and acquisition inside the grammar constitutes a violation of Occam's Razor and thus must be avoided’. Although elsewhere in the book (e.g. p. 63) they use the razor more appropriately as a heuristic, its use in this context seems to abandon the empirical question of whether humans actually do make any use of phonetic substance in phonological computation. Work by Ohala and others (e.g. Ohala Reference Ohala and MacNeilage1983, Blevins Reference Blevins2004) demonstrates the existence of solid diachronic explanations for many (if not all) sound patterns, putting the burden of proof rather squarely on theories that advocate phonetic grounding within synchronic phonology; but a null hypothesis, however aesthetically appealing or subjectively likely, cannot be taken as fact prior to actually being tested. As Flemming (Reference Flemming2001: 26) writes, ‘even if functional constraints originated [through an evolutionary process of adaptation through selection], it does not necessarily mean that these constraints are not also represented in mental grammars. … More definitive statements will have to await further empirical work in psycholinguistics’.

Overall, Hale & Reiss's often provocative claims and reasoning are constructed in such a way that the reader is impelled to turn the eye of scrutiny inwards as much as out; it is hard not to finish the book with a sharper picture of one's own philosophy and fundamental assumptions. For that reason alone, it can be recommended to all phonologists. While it is obviously impossible in a single book to explore all possible answers to the major questions raised, it is certain that many readers will be inspired to pick up this task where the authors left off.