Jesse Zymet

Department of Linguistics
1203 Dwinelle Hall, UC Berkeley
Berkeley, California


Welcome! I’m a lecturer in phonology in the Department of Linguistics at UC Berkeley. My primary research develops and compares computational models of phonology, its interfaces (with the lexicon, morphosyntax, and phonetics), and phonological learning and its biases, and tests predictions made by current hypotheses in the field using experimental methods such as nonce probe and artificial language studies.

My research has brought fieldwork and corpus data from less-studied languages to bear on theories of phonology, its interfaces, and learning. I'm particularly interested in Malagasy (Austronesian) and Maragoli (Bantu), the latter of which I have conducted extensive consultant work on.

Recent updates:


Lexical propensities in phonology: corpus and experimental evidence, grammar and learning

Some theories of variation code individual morphemes with diacritics such as [+/- Rule X] to account for which of them trigger or undergo some variable process. Other theories predict that morphemes may not pattern merely on a binary scale, but rather display gradient lexical propensities to trigger or undergo a process (e.g., [0.7 Rule X]). This investigation conducts a series of corpus investigations into Slovenian palatalization and French liaison, obtaining that individual morphemes pattern on an entire propensity spectrum to trigger or undergo these processes. Moreover, a nonce-probe study into French speakers’ intuitions finds that learners internalize these propensities. These findings support theories that encode a morpheme’s participation on an entire spectrum rather than a binary scale.

Substantial evidence now suggests that language learners can frequency match to statistical generalizations across the lexicon, while learning the various lexical idiosyncrasies of their language. A popular MaxEnt-based approach proposes general constraints for learning a frequency matching grammar together with lexically-indexed constraints to capture idiosyncrasies. This investigation gives a series of learning simulations suggesting that this approach suffers an overfitting problem: late in the learning process the lexical constraints come to explain the entirety of the data, rendering the general constraints superfluous. I argue for a hierarchical theory of learning rooted in the mixed model, by showing that it can learn aggregate generalizations together with idiosyncrasies, with general constraints posed as fixed effects and lexical constraints posed as a random effect.

Malagasy OCP targets a single affix: implications for morphosyntactic generalization in learning

A growing family of research suggests that learners prefer for phonological restrictions to match across morphosyntactic domains. Recent corpus investigations quantified the degree of matching and found that a strong restriction in one domain is accompanied by a weaker, matching tendency in another (Martin 2011, Chong 2016). In this paper, a corpus analysis of Malagasy finds that backness dissimilation applies exclusively and consistently to the passive imperative suffix, but is entirely uncorroborated by phonotactics, which displays a modest but highly significant harmony tendency. Though recent learning models (Martin 2011) enforce a degree of matching, these data suggest that such models must accommodate morphologically specific restrictions that lack matching tendencies elsewhere.

Irreducible parallelism in phonology: evidence for lookahead from Mohawk, Maragoli, Sino-Japanese, and Lithuanian

My colleague Jeff Adler and I found lookahead effects in the languages we were studying, coming to roughly the same conclusions (1, 2, 3). We collaborated on a paper, pooling together cases of lookahead across a diverse breadth of systems displaying stress assignment, reduplication, assimilation, and more. Optimality Theory, which has full lookahead, captures these cases naturally, while Harmonic Serialism, which has no lookahead, is challenged by them. We give a unified explanation for why this is: in each case, entire procedures -- entire sequences of processes applying to the input -- must be compared to best satisfy constraints in the language. Our findings suggest that grammatical models must have lookahead capability.

I'm currently exploring the implications of these systems for computational complexity in phonology.

My fieldwork on Maragoli focused on its rich possessive agreement system, which displays lookahead: the order in which reduplication and hiatus repair apply is conditioned by the well-formedness of the fully formed reduplicant. The system is capturable in OT, but challenging for HS.

Distance-based decay in long-distance phonological processes: probabilistic models for Malagasy, Latin, English, and Hungarian

Long-distance phonological processes can exhibit distance-based decay, whereby application rate decreases as distance between the trigger and target increases. This paper gives a comprehensive treatment of the effect within Maximum Entropy Harmonic Grammar, fitting a single model to corpus data on four languages. The decay effect is established as general phenomenon, arising in harmony and dissimilation processes regulating both vowel pairs and consonants pairs. Distance is shown to be best measured in syllables rather than in segments or moras. A crosslinguistically fixed scale adjusting a single constraint weight fits the decay effect across all surveyed languages; by comparison, a constraint pair regulating local and nonlocal application consistently underpredicts gradience, whereas a family of distance-specific constraints is found to be unmotivated in light of scaling. The results support a theory in which the learner utilizes a restrictive scale for a single constraint.

Substantive bias and final (de)voicing: an artificial language study
Glewwe, Zymet (co-first authors), Adams, Jacobson, Yates, Zeng, Daland

While final devoicing facilitates articulatory ease and is typologically common, final voicing is unattested. The substantive bias hypothesis (i.e., phonetic naturalness bias hypothesis) predicts that final devoicing is easier to learn, relative to final voicing. We conducted an artificial grammar learning experiment that tested for substantive bias in the case of final devoicing/voicing. There were three training patterns, Devoicing, Voicing, and Change (a more complex final voicing/devoicing exchange rule). In test, a two-alternative forced choice task tested participants’ learning of their training pattern. Voicing was learned better than Devoicing and Change, which did not differ. This result is inconsistent with the substantive bias hypothesis, which predicts better learning of final devoicing. Thus we found no evidence supporting substantive bias, in line with Moreton & Pater (2012). In particular, our results suggest that articulatory ease does not bias phonological learning.


Lexical propensities in variable phonology: corpus and experimental evidence from Slovenian and French

Substantive bias and artificial (de)voicing: an artificial language learning study

Malagasy OCP targets a single affix: implications for morphosyntactic generalization in learning

Irreducible parallelism in phonology

On the relationship between sonority and glottal vibration (Risdal, Aly, Chong, Keating, Zymet)

What factors contribute to the sonority of a segment? We analyzed recordings of voiced segment types (obstruents, nasals, liquids, glides, vowels), obtaining the following primary results: strength and degree of voicing differs depending on segment type, with weaker and breathier voicing being associated with segments having a tighter restriction (as in obstruents); moreover, measures of voicing strength (strength of excitation) and breathiness (closed quotient) correlate with the sonority hierarchy (obstruents < approximants < vowels).

A case for parallelism: reduplication and repair in Maragoli

Distance-based decay in long-distance phonological processes

An investigation into apparent sublexical coordination in English (with Clara Sherley-Appel, UCSC & Google)

We present data on coordinated affix constructions (CACs): constructions such as pre- and post-operative. We argue that CACs are not merely coordinations of affixes, but rather coordinations of phrases that underwent subsequent right-node raising ([pre-operative] and [post-operative] → [[pre-t] and [post-t]] operative).

Updated May 2019.