📄 An Optimality Theoretic Account of Nasal Assimilation in English

Summary

The paper offers a new treatment of English nasal place assimilation within Optimality Theory (OT). It focuses on how the velar nasal [ŋ] arises from /n/ before velars, and how /n/ becomes [m] before labials. Earlier rule-based accounts fall short. Halle & Mohanan (1985) tie blocking of assimilation to a following stressed syllable, but pairs like linguist/linguistic contradict this. Borowsky (1986) needs two distinct rules, one spreading and one copying, operating at different lexical levels, and her account still mishandles prefixed forms like impossible. Kang replaces both with a small ranked set of violable constraints:

• a ban on n+velar inside a foot or monomorpheme, outranking an anti-spreading constraint;
• a constraint against nasal+g within a syllable;
• an undominated ban on n+labial.

The analysis also needs the output to keep the content of the input. Here the paper argues that structural (parse-based) faithfulness picks the wrong winner for words like angma. Correspondence (McCarthy & Prince 1994), extended to individual features as proposed by Lamontagne & Rice (1995), selects the right form. This supports feature-level correspondence as a general tool.

Key points

• The data contrasts stress-sensitive prefix+root pairs (co[n]gress vs co[n]gressional) with stress-insensitive monomorphemes (li[ŋ]guist/li[ŋ]guistic, Hu[ŋ]gary/Hu[ŋ]garian). These examples are taken from Borowsky (1986).
• The proposed hierarchy is:
  • top: *[…n{k,g}…]foot/monomorpheme, PARSE (featural correspondence) and *[+nasal]g]σ;
  • below them: SPREAD and *Multiple Correspondence (*MC).
• *n[+labial] is undominated. It explains why labial assimilation applies regardless of stress (compose/composition, symbol/symbolic, impossible).
• For angma, long and bring, the winning output has a single [ŋ] that corresponds to both input /n/ and /g/. This is a merger that violates only the low-ranked *MC, not PARSE.
• Under Prince & Smolensky's (1993) structural Parse, the wrong candidate an<g>.ma would win. Footnote 6 adds that the segment-only correspondence version fails as well.

Interesting ideas

• I1. Strict ranking with candidate tableaux makes every decision traceable: a candidate loses at the first constraint where a rival does better, and ties fall through to lower constraints.
• I2. Letting one output element correspond to several input elements is allowed, but it carries a low-ranked cost, which models merger without counting it as deletion.
• I3. The paper refutes a surface cue (stress) with minimal pairs and replaces it with a structural domain (foot or morpheme boundary). The lesson is to test a conditioning factor against pairs where the cue changes but the outcome does not.
• I4. The argument for the new framework rests on showing that the rival evaluation picks a wrong winner on one concrete form (angma). One decisive failing case can be worth more than broad coverage claims.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): Ranked, violable preferences with a printed tableau-style trace could decide between competing role or frame readings, so each choice comes with an explicit "lost at constraint X" explanation.
• I2 (back): The text → graph → text compressor could record many-to-one links (e.g., contractions, merged mentions) so that fused material counts as preserved rather than lost when checking round-trip fidelity.
• I3 (learn): Self-induced repair rules could be checked against automatically gathered minimal-pair negative controls, rejecting rules whose trigger is only correlated with the real domain.
• I4 (learn): Evaluation suites could keep a small set of "decisive" items where competing parsers or rule sets diverge, and use them to adjudicate between witness parsers.