Phrasal stress

Sentence-level prominence from dependency parses (MetricalTree port)

How prosodic computes phrasal (sentence-level) prominence, where the algorithm comes from, and how our implementation differs from its ancestors. Code: prosodic/texts/phrasal_stress.py. Enabled with TextModel(txt, syntax=True).

Linguistic background

Lexical stress tells you which syllable of within is strong; it says nothing about whether within is prominent in its sentence. That second, relational kind of prominence is the subject of metrical phonology’s founding work:

  • Chomsky & Halle (1968, SPE) formulated the Nuclear Stress Rule (NSR): in a phrase, primary stress falls on the rightmost stressable constituent — recursively, so sentence prominence cascades down through phrases. The Compound Stress Rule (CSR) is its mirror inside compound nouns: stress the left member (TIME machine, not time MACHINE).
  • Liberman & Prince (1977) recast these rules relationally: constituents are strong or weak only relative to their sisters, represented in binary metrical trees, with the metrical grid as the derived rhythmic representation. A word’s prominence is a function of how many weak nodes dominate it.
  • Hayes (1983, 1995) developed the grid-based side of the theory — prosodic’s grid_str() display (marks stacked over syllables by prominence level) is a Hayes-style grid.
  • Anttila and colleagues (e.g. Anttila, Dozat, Galbraith & Shapiro
    1. operationalized these theories at corpus scale to study English phrasal stress, rhythm in prose, and its interaction with meter — using Dozat’s MetricalTree tool, the direct ancestor of this module.

What Dozat’s MetricalTree did

MetricalTree (Timothy Dozat, 2015) computes Liberman–Prince prominence over Stanford CoreNLP parses:

  1. Tree: a constituency parse, with dependency labels decorating preterminals.
  2. Lexical stress classes: each word is scored 0 (stressed content word), −1 (unstressed function word), or −0.5 (ambiguous) using word lists (it; this/that/these/those), PTB tag lists (CC, PRP\(, TO, UH, DT unstressed; MD, IN, PRP, WP\), PDT, WDT, WP, WRB ambiguous), and dependency-label lists (det, expl, cc, mark unstressed; cop, neg, aux, auxpass ambiguous). Punctuation and contracted clitics are NaN.
  3. Disambiguation ensemble: ambiguous words are resolved three deterministic ways — all stressed; unstressed only if monosyllabic; all unstressed — and results are averaged. This is what makes the output gradient rather than categorical: an ambiguous monosyllable ends up with fractional prominence.
  4. NSR pass (set_pstress): bottom-up through the tree; in each phrase, the rightmost child with pstress 0 stays strong and every other child is demoted to −1. Inside NN-sequences (compounds) the scan is inverted per the CSR: the rightmost noun is provisionally weak and the noun to its left is strong.
  5. Total stress (set_stress): each word’s cumulative prominence = its lexical stress + its preterminal pstress + the sum of pstress values of every phrase dominating it — a direct arithmetization of Liberman & Prince’s “count the weak nodes above you.”
  6. Normalization: per sentence, min–max normalize to [0, 1], where 1.0 is the sentence’s nuclear stress. Two outputs per word: pstress (preterminal phrasal stress: strong/weak within the local phrase) and tstress (total cumulative stress: global sentence prominence).

cadence (archived 2026-07) carried this implementation forward over Stanza constituency parses, with one simplification: by default it substituted binary dictionary lexical stress for the ambiguity ensemble.

What prosodic does differently

Prosodic v3 deliberately has no constituency parser (no Stanza, no torch model downloads); its syntax=True pipeline is spaCy dependency-only. The port therefore runs Dozat’s algorithm over head-projection trees derived from the dependency parse: every word projects a phrase whose ordered children are the projections of its dependents plus its own preterminal.

Naive projection is not equivalent to constituency, and the differences matter exactly where the NSR looks. Three topology rules (found by differential testing, below) restore congruence:

  1. NP-internal prenominal modifiers join bare. A dependent with no dependents of its own whose relation is in {det, amod, compound, nummod, predet, neg, aux, auxpass} sits in its head’s phrase as a bare preterminal — matching the flat Penn NP (NP (DT the) (JJ quick) (NN fox)), so NSR demotion reaches the adjective directly.
  2. Arguments always project, even as single words — constituency wraps them (NP(Jack), WHADVP(When)), and that wrapper shields the preterminal: when a subject phrase is demoted, the demotion lands on the phrase node, and the word’s own pstress (where the ambiguity ensemble lives) survives. Notably poss is not in the bare set: a possessor is an inner NP (NP(NP(Beauty 's) rose)), and bare-joining it would wrongly trigger the compound rule (BEAUTY’S rose).
  3. Noun heads with right-side complements get an inner core. For the house that Jack built, constituency gives NP(NP(the house) SBAR): the relative clause wins the NSR, but the demotion hits the inner NP node and house keeps its preterminal strength. The projection replicates this by wrapping left material + head in an inner node whenever an NN-headed phrase has right dependents.

The compound rule needs no special handling beyond rule 1: compound dependents are typically leaves, so they sit as bare NN preterminals adjacent to the head noun — exactly the NN-sequence Dozat’s scan expects.

Everything above the tree skeleton is ported faithfully: the lexical stress classes, the 3-variant ensemble, the NSR/CSR scan (including its break/skip idiosyncrasies), cumulative total stress, and per-sentence min–max normalization. Output columns in _syll_df: pstress, tstress (Float64, ∈ [0,1], NaN for punctuation and for sentences with no variation), broadcast from words to their syllables.

Validation

The port was cross-validated against cadence (Dozat’s algorithm over Stanza constituency, ambiguity ensemble replicated) on a 9-sentence differential covering NSR baselines, compounds, ambiguous function words, relative clauses, coordination, possessives, ditransitives, and PP attachment (2026-07-05). Result: identical nuclear-stress placement on 9/9; identical orderings throughout; several sentences identical to two decimals on every word. The differential caught two real bugs (the shielding asymmetry behind topology rules 1–3, and poss in the bare set) — the corrected, cross-validated values are pinned as unit tests in tests/test_phrasal_gradient.py.

One deliberate divergence: in coordination (dogs and cats), Dozat’s flat-NP scan demotes non-final conjuncts to pstress −1; prosodic keeps each conjunct’s strength (both conjuncts carry accent in natural speech), while tstress still ranks the final conjunct higher, matching the reference ordering.

Known limitations: (a) attachment-height differences between spaCy dependencies and constituency (e.g. fronted When sits higher as a WHADVP than as an advmod edge) shift tstress magnitudes on a few words; (b) whole-word possessive tokens (Beauty’s) misparse in both spaCy and Stanza when pre-tokenized — an input-level limitation; (c) the lexical class lists are English-only.

The two value systems, and what uses them

discrete phrasal_stress gradient pstress / tstress
lineage v3-native depth heuristic Dozat MetricalTree port
values 0 (root) … −6, Int [0, 1] normalized, Float
computation vectorized dep-tree depth + NSR/CSR sibling promotion ensemble of NSR passes over projection trees
constraints w_prom, s_demoted w_stress_p, s_unstress_p, w_stress_t, s_unstress_t (cadence thresholds: prominent > 0, weak == 0)
grid — phrasal rows: word’s primary syllable +1 level at tstress ≥ 0.5, +1 at nuclear

All phrasal constraints are inert without syntax=True.

Empirical caveat (from the v3 MaxEnt experiments): on Shakespeare’s sonnets with a fixed wswswswsws target, phrasal constraints added no accuracy over lexical stress. Phrasal prominence matters for prose rhythm and naturalness ranking — cadence’s research agenda — not for fixed-template scansion.

References

  • Chomsky, N. & M. Halle (1968). The Sound Pattern of English.
  • Liberman, M. & A. Prince (1977). On stress and linguistic rhythm. LI 8.
  • Hayes, B. (1983). A grid-based theory of English meter. LI 14.
  • Hayes, B. (1995). Metrical Stress Theory.
  • Dozat, T. (2015). MetricalTree. github.com/tdozat/Metrics
  • Anttila, A., T. Dozat, D. Galbraith & N. Shapiro (2020). Sentence stress in presidential speeches. In Prosody in Syntactic Encoding.
  • Heuser, R. (2021–23). cadence. github.com/quadrismegistus/cadence