📄 Toward Treating English Nominals Correctly

Summary

Sproat and Liberman address how a text-to-speech system should place stress in English complex nominals, meaning a head noun that follows at least one open-class modifier (e.g. dog catcher, Madison Avenue, City Hall parking lot). Such nominals are frequent and often phrase-final, so misplacing their main accent noticeably harms synthetic speech. The authors argue that stress depends on syntactic level (N⁰ compounds versus higher X-bar attachments), on meaning relations between the words, and on special constructions such as names. Their program has two parts. For binary nominals it first checks a phrasal lexicon of exceptionally left-stressed items. Otherwise it collects weighted heuristic patterns (measure phrases, location/time/substance modifiers, deverbal heads, self, place names, a weak noun-noun default) and lets them vote. For n-ary nominals it recursively combines binary results on a metrical grid and applies the Rhythm Rule. On 200 binary nominals the program reached 80% accuracy, compared with 45% for uniform right stress. The same lexical knowledge also helps bracket complex nominals.

Key points

• The Brown Corpus contains over 73,000 nominals longer than two words. Focus or contrast shifts the main stress in only a fraction of a percent of phrase-final cases.
• The lexicon lists roughly 7000 left-stressed nominals. The authors stress that being lexicalized neither guarantees nor is required for left stress (red herring is listed yet right-stressed; dog annihilator is unlisted yet left-stressed).
• In deverbal heads (-ing, -er, other affixes), a left member read as the verb's object tends to get left stress (dog catcher). A left member read as its subject tends to get right stress (woman swimmer).
• Evaluation on 200 [N N] items from Bell Labs News gave these accuracies:
  • uniform right stress: 45%
  • left stress whenever the item is noun+noun: 66%
  • the full program: 80%
• Of about 40 errors, roughly 30 came from missing knowledge (e.g. chemical terms like gallium arsenide) and only about 10 from misanalysis. The authors judge 90–95% to be a reachable target.

Interesting ideas

• I1. Each heuristic casts a vote weighted by how reliable it is, so a near-certain pattern like [self N] dominates while the vague noun-noun default barely counts; this cleanly combines weak and strong evidence.
• I2. The paper asks what role the modifier plays for a deverbal head (object versus subject of the underlying verb) and treats this as a real semantic decision rather than a surface fact.
• I3. The authors split failures into "insufficient information" and "misanalysis," which shows whether to invest in more knowledge or in better rules.
• I4. The program is compared against two trivial baselines (always right, and left only for noun+noun), which makes the 80% figure meaningful.
• I5. Modifier classes such as measure unit, location, time, and substance determine how the compound behaves, and semantic word classes double as rule triggers.

For Mova

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

• I1 (tree): Bracketing of multi-noun compounds could use such weighted, inspectable votes, with each vote logged as an explanation for the chosen attachment.
• I2 (frames): Nominalizations like baby sitting or student dancing could be mapped to verb frames, with the left noun assigned a Patient or Agent role.
• I3 (learn): Evaluation reports could tag every error with this split, directing effort toward lexicon or ontology growth versus repairs to the rules.
• I4 (learn): Each component evaluation could include such cheap default baselines as negative controls, so that gains are measured against what a naive rule already achieves.
• I5 (world): WordNet/Wikidata classes for measures, places, and materials could label compound relations (pint jug = capacity, wood floor = made-of), and these labels could feed both the world model and prosody in generation.