Breiss, Katsuda & Kawahara (2026): Token frequency modulates optional paradigm uniformity in Japanese voiced velar nasalisation #

@cite{breiss-katsuda-kawahara-2026} @cite{mccarthy-2005} @cite{steriade-2000} @cite{ito-mester-1996} @cite{ito-mester-2003} @cite{hibiya-1995} @cite{coetzee-pater-2008} @cite{paster-2019}

The Japanese velar /g/ → [ŋ] alternation in N1+N2 nominal compounds is optional: speakers vacillate between [g] and [ŋ] for many compounds, and the rate of nasalisation varies across compounds, items, and speakers. The paper's central architectural claim is that this optionality is modulated by token frequency through two opposite-sign channels:

High token frequency of N2 as a free wordform decreases the rate of nasalisation (negative regression coefficient on N2 token frequency). The free-form [g] is a more accessible paradigm exemplar; paradigm-uniformity pressure preserves it, suppressing [ŋ].
High token frequency of the compound itself increases the rate of nasalisation (positive regression coefficient on compound token frequency). More-attested compounds drift further from their constituent forms.

Both effects only apply when N2 is morphologically free. When N2 is bound (occurs only inside compounds — no surface [g] paradigm exemplar to anchor to), nasalisation is categorically obligatory. The two-channel frequency story collapses to a single-channel (markedness- only) story in the bound case.

Examples from the paper #

Taken verbatim from @cite{breiss-katsuda-kawahara-2026}:

(1a) /haigan/ ~ /haiŋan/ "lung cancer" — N2 'cancer' (癌) is free.
(1b) /noogeka/ ~ /nooŋeka/ "brain surgery" — N2 'surgery' (外科) is free.
(2a) /dokuga/ ~ /dokuŋa/ "poison moth" — N2 'moth' (蛾) is free.
(2b) /dokuŋa/ *[dokuga] "poison fang" — N2 'fang' (牙) is bound.
(3) /gaʒoo/ "main castle" — initial-position [g] never nasalises.

The minimal pair (2a)/(2b) is the paper's central piece of evidence: two compounds with identical surface form /dokuga/ but different free/bound status of the segmentally-identical N2 yield categorically different nasalisation behaviour.

Connection to Paradigm Uniformity #

The architecture is paradigm uniformity (PU) + frequency-conditioned strength. The compound and its free N2 stand in a paradigm relation; PU prefers their shared segments to be alike. The PU pressure is modulated — not just on/off — by the token frequency of the N2. This puts the paper at the intersection of:

@cite{mccarthy-2005} (PU as the symmetric pairwise lift over members; see ParadigmUniformity/OptimalParadigms.lean).
@cite{steriade-2000} Lexical Conservatism (PU pressure is anchored on attested wordforms; see ParadigmUniformity/LexicalConservatism.lean).
@cite{coetzee-pater-2008} Frequency-scaled weights (the modulation channel — token-frequency drives a continuous weight; see ItemSpecificity/ScaledWeights.lean).

The previous constraint-based account of @cite{ito-mester-1996} / @cite{ito-mester-2003} treats nasalisation as the result of a high-ranked markedness constraint; @cite{hibiya-1995}'s sociolinguistic study established the variable, lexically-modulated character of the alternation. BKK 2026's contribution is the sign of the two frequency channels and the architectural commitment that the two-direction story collapses to one direction in the bound case.

Connection to ItemSpecificity theories #

The companion modelling paper (Breiss, Katsuda & Kawahara, lingbuzz/009508) fits a MaxEnt grammar with @cite{steriade-2000}'s Lexical Conservatism. We do not formalise the fitting routine here. The discrimination this study makes against the four siblings in Theories/Phonology/ItemSpecificity/:

ScaledWeights (@cite{coetzee-pater-2008}): consistent with the data, with separate slopes per channel (positive on cpd freq, negative on N2 freq).
RepresentationStrength (@cite{moore-cantwell-2021}): consistent — high N2 activation preserves the boundary segment.
UseListed (@cite{zuraw-2000}): ruled out by Experiment 2 (novel compounds show the same N2-frequency gradient as familiar ones) — see novel_compounds_show_n2_gradient below.
Indexed constraints (@cite{pater-2010}): in principle a multi-stratum approximation could fit, but parsimony favours the continuous accounts.

@cite{paster-2019}'s critique of "counting" patterns in phonology is relevant to BKK Experiment 2's finding that N2 length (not total compound length) matters — undermining a mora-counting analysis and favouring a paradigm-anchored account.

Boundary #

We formalise the qualitative direction-of-effect predictions, not numerical fits, sample sizes, or specific corpus statistics.
"Optional" is taken at face value as variable surface realisation; we do not commit to a stochastic OT vs. MaxEnt vs. mixed-effects encoding of the variation. The relevant fact for downstream theory is only that nasalisation rate is monotonic in the relevant log-frequency, with the appropriate sign per channel.
The wug-test methodological contract lives in Paradigms/WugTest.lean; this file consumes that paradigm via the novel_compounds_show_n2_gradient discriminator.