The study of phonotactics (e.g., the ability of English speakers to distinguish possible words like blick from impossible words like *bnick) is a central topic in phonology. We propose a theory of phonotactic grammars and a learning algorithm that constructs such grammars from positive evidence. Our grammars consist of constraints that are assigned numerical weights according to the principle of maximum entropy. Possible words are assessed by these grammars based on the weighted sum of their constraint violations. The learning algorithm yields grammars that can capture both categorical and gradient phonotactic patterns. The algorithm is not provided with any constraints in advance, but uses its own resources to form constraints and weight them. A baseline model, in which Universal Grammar is reduced to a feature set and an SPE-style constraint format, suffices to learn many phonotactic phenomena. In order to learn nonlocal phenomena such as stress and vowel harmony, it is necessary to augment the model with autosegmental tiers and metrical grids. Our results thus offer novel, learning-theoretic support for such representations. We apply the model to English syllable onsets, Shona vowel harmony, quantity-insensitive stress typology, and the full phonotactics of Wargamay, showing that the learned grammars capture the distributional generalizations of these languages and accurately predict the findings of a phonotactic experiment.

Chomsky and Halle (1968) and many formal linguists rely on the notion of a universally available phonetic space defined in discrete time. This assumption plays a central role in phonological theory. Discreteness at the phonetic level guarantees the discreteness of all other levels of language. But decades of phonetics research demonstrate that there exists no universal inventory of phonetic objects. We discuss three kinds of evidence: first, phonologies differ incommensurably. Second, some phonetic characteristics of languages depend on intrinsically temporal patterns, and, third, some linguistic sound categories within a language are different from each other despite a high degree of overlap that precludes distinctness. Linguistics has mistakenly presumed that speech can always be spelled with letter-like tokens. A variety of implications of these conclusions for research in phonology are discussed.* The generative paradigm of language description (Chomsky 1964, 1965, Chomsky & Halle 1968) has dominated linguistic thinking in the United States for many years. Its specific claims about the phonetic basis of linguistic analysis still provide the cornerstone of most linguistic research. Many criticisms have been raised against the phonetic claims of the Sound pattern of English (Chomsky & Halle 1968), some from early on

Abstract not found

This paper reports a difference in English between the prosodic properties of contrastive FOCUS and presentational focus entities when they are immediately followed within the sentence by a presentational focus. In this context, the contrastive FOCUS shows not only a L+H * pitch accent, but also a following phonological phrase break, marked by both a L- phrase accent and temporal disjuncture. In the same context, presentational focus shows a pitch accent H * and no phonological phrase break (no L- and no disjuncture). Since these differences in prosody correlating with different Focus types are not plausibly construed as the realization of distinct tonal morphemes, I suggest that the contrastive/presentational contrast is present in the interface informational/syntactic structure itself, and makes itself felt in the phonological representation through the action of syntax-phonology interface constraints which distinguish the two Focus types.

Abstract. This paper investigates a gradual on-line learning algorithm for Harmonic Grammar. By adapting existing convergence proofs for perceptrons, we show that for any nonvarying target language, Harmonic-Grammar learners are guaranteed to converge to an appropriate grammar, if they receive complete information about the structure of the learning data. We also prove convergence when the learner incorporates evaluation noise, as in Stochastic Optimality Theory. Computational tests of the algorithm show that it converges quickly. When learners receive incomplete information (e.g. some structure remains hidden), tests indicate that the algorithm is more likely to converge than two comparable Optimality-Theoretic learning algorithms.

The markedness constraints of classic Optimality Theory assign violation-marks to output candidates without reference to the input or to other candidates. This article explores an alternative conception of markedness: markedness constraints compare the candidate under evaluation with another candidate, the most faithful one. Comparative constraints distinguish two situations: the candidate under evaluation contains an instance of a marked structure that is also present in the fully faithful candidate; or it contains an instance of a marked structure that is not present in the fully faithful candidate. Empirical consequences of comparative markedness are explored, including grandfather effects, derived environment effects, non-iterating processes, and counter-feeding opacity. Comparative markedness is found to have some advantages and some disadvantages in comparison with classic OT and alternatives like local conjunction, stratal OT, sympathy, and targeted constraints.

Abstract not found

Many languages respect the generalization that some or all unstressed vowels are deleted. This generalization proves elusive in classic Optimality Theory, however. The source of the problem is classic OT’s parallel evaluation, which requires that the effects of stress assignment and syncope be optimized together. This article argues for a version of OT called Harmonic Serialism, in which the effects of stress assignment and syncope can and must be evaluated sequentially. The results are potentially applicable to other domains where process interaction is best understood in derivational terms.

When a medial consonant cluster is simplified by deletion or place assimilation, the first consonant is affected, but never the second one: /patka / becomes [paka] and not *[pata]; /panpa / becomes [pampa] and not [panta]. This article accounts for that observation within a derivational version of Optimality Theory called Harmonic Serialism. In Harmonic Serialism, the final output is reached by a series of derivational steps that gradually improve harmony. If there is no gradual, harmonically improving path from a given underlying representation to a given surface representation, this mapping is impossible in Harmonic Serialism, even if it would be allowed in classic Optimality Theory. In cluster simplification, deletion or Place assimilation is the second step in a derivation that begins with deleting Place features, and deleting Place features improves harmony only in coda position.

Abstract not found