Sensitivity to string length and feature count subverts MaxEnt universals

HAL (Le Centre pour la Communication Scientifique Directe) · 2022-01-01

International audience

Tone and morphological level ordering in Dagaare

Phonology · 2022-08-01 · 1 citations

Abstract Dagaare is a language of northern Ghana and adjoining areas of Burkina Faso. There are two tones, H and L, and contrastive downstep H ! H that involves a non-automatic pitch drop between two H tones. The challenge is to explain the extensive morphological conditioning of tonal processes, including dissimilation, downstep and spreading. Our solution involves level ordering: tones are introduced at different morphological levels (stems, words and phrases) and later processes can make earlier processes opaque. Tonal differences between nouns (spreading) versus verbs (no spreading) and stems (dissimilation) versus words (downstep) arise from constraint ranking differences within and across levels. There are two kinds of downsteps: stem-level downsteps are underlying L tones affiliated with some morpheme; word-level downsteps are L tones inserted by a general process of word-final lowering. Only one downstep per word is allowed. If more would arise, the morphologically inner downstep blocks the morphologically outer downstep.

To Predict or to Memorize: Prominence in Inaugural Addresses

Proceedings of the Annual Meetings on Phonology · 2021

• Computer Science
• Natural Language Processing
• Artificial Intelligence

The assignment of phrasal prominence has been variously attributed to syntactic structure, part of speech, predictability, informativity, and speaker's intent. A recent account asserts that prominence is memorized on a by-word basis as Accent Ratio (AR), the likelihood that a word is accented (Nenkova et al. 2007). We examined whether AR outperforms the traditional predictors, in particular syntax and informativity, and if not, whether the traditional predictors shed light on the variance left unexplained by AR. We used a corpus of spoken American English consisting of the first inaugural addresses of six recent American presidents, hand-annotated for stress by two native English speakers. Regression models fitted to the data revealed that AR, syntax, and informativity all independently matter. Dividing the data into high-prominence and low-prominence tokens further revealed that AR and informativity are significant among low-prominence words, but only syntax is significant among high-prominence words. We conclude that although AR is a highly successful predictor, certain aspects of phrasal prominence require reference to syntax and informativity.

On the Phonology and Semantics of Deaccentuation

Proceedings of the Annual Meetings on Phonology · 2021 · 1 citations

• Computer Science
• Natural Language Processing
• Linguistics

The deaccentuation of given and/or repeated elements is familiar from many dialects of English. We propose that deaccentuation is essentially an optional postlexical phonological process of stress retraction triggered by two constraints: *Stress-Copy, which assigns a violation to a stress peak on a word with a segmentally identical copy in the left context, and Rightmost, which assigns a violation to every word between a stress peak and the right phrase edge. We quantify deaccentuation by defining it as being perceived with less stress than expected, where expected stress is calculated by an implementation of Liberman and Prince's (1977) phrasal stress algorithm. We provide empirical evidence for our analysis based on the first inaugurals of six former U.S. presidents.

Sentence stress in presidential speeches

De Gruyter eBooks · 2020 · 30 citations

• Linguistics
• Psychology
• Philosophy

Sentential prominence is not represented in writing, it is hard to measure phonetically, and it is highly variable, yet it undoubtedly exists. Here we report preliminary findings from our study of sentential prominence in the inaugural addresses of six U.S. presidents. We confirm the familiar hypothesis that sentential prominence has two sources (Jespersen 1920): it is partly mechanical and depends on syntax (Chomsky & Halle 1968, Liberman & Prince 1977, Cinque 1993) and partly meaningful in that it highlights informative material (Bolinger 1972). Both contribute independently to perceived prominence. Pursuing the view that sentential prominence is amatter of stress,we provide evidence for the linguistic reality of the Nuclear Stress Rule (Chomsky & Halle 1968) as well as the view that information coincides with stress peaks in good prose (Bolinger 1957).We also observe that part of speech matters to sentence stress: noun and adjective stresses are loud and mechanical; verb and function word stresses are soft and meaningful. We suggest that thismay explainwhy parts of speech differ inword phonology as well.

Equiprobable Mappings in Weighted Constraint Grammars

ScholarWorks@UMassAmherst (University of Massachusetts Amherst) · 2020-01-01 · 2 citations

We show that MaxEnt is so rich that it can distinguish between any two different input-output mappings: there always exists a nonnegative weight vector that assigns them different MaxEnt probabilities. Stochastic HG instead does admit equiprobable mappings, namely mappings that have the same probability for every weight vector, and we give a complete formal characterization of their violation profiles. Empirically, we show that the predictions of MaxEnt and Stochastic HG differ for a grammar of Finnish secondary stress. We test the predictions of both models against a large internet corpus and find preliminary support for Stochastic HG and against MaxEnt.

Equiprobable mappings in weighted constraint grammars

2019-01-01 · 1 citations

We show that MaxEnt is so rich that it can distinguish between any two different mappings: there always exists a nonnegative weight vector which assigns them different MaxEnt probabilities. Stochastic HG instead does admit equiprobable mappings and we give a complete formal characterization of them. We compare these different predictions of the two frameworks on a test case of Finnish stress.

Does MaxEnt Overgenerate? Implicational Universals in Maximum Entropy Grammar.

HAL (Le Centre pour la Communication Scientifique Directe) · 2018-01-01 · 3 citations

International audience