Two backends for phrasal stress: dependency vs constituency

A faithful Liberman & Prince implementation, and what validating it taught us about the target

Prosodic computes phrasal (sentence-level) prominence two ways, from two parsers, for two purposes. This page documents the second, experimental backend — a maximally faithful realization of Liberman & Prince (1977) over real constituency parses — how it was validated, and the sharper understanding of what phrasal stress is for that the validation produced.

Companion code: prosodic/analysis/metrical_lp.py. Background on the shipping (dependency) backend: Phrasal stress.

Two backends, two purposes

spaCy (shipping, default) Stanza constituency (experimental)
parse dependency projection (n-ary) real constituency (binary metrical tree)
output gradient tstress in [0,1] categorical grid + binary s/w tree, and a gradient
lineage Dozat MetricalTree (§ Phrasal stress) Liberman & Prince (1977), implemented literally
cost ~0.7 ms/line ~11 ms/line (16× slower; see below)
best at a fast prominence signal the theory’s structure — grid, DTE, s/w

They are not competing implementations of one thing. The dependency backend gives a continuous prominence value cheaply. The constituency backend gives the categorical objects of the theory — the binary metrical tree, the designated terminal element (DTE), and the metrical grid as the tree’s RPPR projection — which a dependency projection cannot represent (it is n-ary, so the strong/weak relation, defined only on pairs of sisters, has no place to live).

The faithful implementation

The constituency backend builds L&P’s theory bottom to top:

  • Within-word trees. Each word’s syllables are grouped into L&P feet from dictionary stress; the primary foot is the word’s DTE.
  • Phrasal trees. A Stanza constituency parse is binarized per the NSR (phrases are right-strong) and CSR (common-noun compounds are left-strong); proper-noun modifiers stay phrasal, so Montana cowboy is [Montana(w) cowboy(s)], not a compound.
  • Lexical tier. POS and dependency labels classify each word as content / ambiguous / unstressed — the dependency labels matter because copulas and auxiliaries (is, has, was) wear content-verb POS tags but are prosodically weak (cop, aux). L&P’s provision (114) then floors content words at grid level 2.
  • The grid is the minimal metrical grid satisfying the Relative Prominence Projection Rule (their eq. 104): the DTE of the strong subconstituent outranks the DTE of the weak one.

Token handling: prosodic keeps clitics merged (summer's, won't), while Stanza’s model expects Penn tokenization with clitics split. The backend lets Stanza tokenize freely, then collapses its tokens back to prosodic words by character offset — so 's/n't never appear as tree or grid nodes, and possessive structure is preserved.

Validation 1 — against L&P’s own figures (structure)

L&P print exact trees and grids for their examples. Encoding them as a gold set (using their input scansions, since the Rhythm Rule is not implemented) gives two numbers:

  • Grid engine: 7/7. Hand-encoded trees for thirteen men, Tennessee air, Montana cowboy, execute, knowledgeable, poly vinyl chloride, John’s three red shirts reproduce L&P’s printed grid heights exactly.
  • Automatic pipeline: 10/11 nuclear. From raw text, the full Stanza pipeline recovers L&P’s published nuclear placement on every phrase but one — poly vinyl chloride, which L&P themselves flag as structurally ambiguous in (119).

A hard-case set (compounds, possessives, proper-noun modifiers, PP attachment, relative clauses) scored the two backends head to head on nuclear placement against hand-derived gold. On the defensible cases the constituency backend was 28/28, the dependency projection 27/28 — its one miss an adjective phrase.

ImportantCorrection: the possessive gap was a tokenization artifact

An earlier version of this comparison had the dependency projection at 25/28, missing all three possessives (John's car → John), and concluded constituency was “measurably better on possessives.” That was wrong — and instructively so. The spaCy backend had been fed prosodic’s merged tokens (beauty's as one token), which spaCy mis-tags NNP/compound, handing prominence to the possessor. Splitting the clitic through spaCy’s own tokenizer (beauty's → beauty + 's → poss) fixes it: spaCy then scores 3/3 on possessives and the gap closes to a single adjective phrase. On 118 sonnet lines the two backends, both properly tokenized, agree on nuclear placement 91% of the time (per-word tstress Spearman ≈ 0.70) — and spaCy even gets thy beauty’s legacy → legacy right where the constituency pipeline picks the possessor. Lesson: tokenization, not the dependency-vs-constituency choice, drove the possessive difference. The fix now ships in the default spaCy backend (spacy_free=True).

On L&P’s own figures the split was 10/11 vs 7/11 (the dependency projection mis-stresses Tennessee air, Montana cowboy, law degree requirement changes). So constituency retains a marginal structural edge — it has the sisterhood the NSR/CSR are defined over — but, tokenization held equal, it is much smaller than the confounded comparison implied.

Validation 2 — against real speech (and why it’s the wrong target)

To ask whether the gradient is better, we compared both backends against human word-level prominence judgments — the Crowdsourced LibriTTS Prominence Annotations (Morrison et al. 2024; 977 usable utterances). The result was clarifying, and mostly negative:

  • Mean per-utterance Spearman ρ: spaCy 0.440, Stanza 0.386 (Wilcoxon p ≈ 10⁻²¹). The dependency backend is significantly better here.
  • But a lexical-class-only baseline (“content words are prominent,” no tree at all) scores 0.581 — beating both models. The tree-based gradient degrades a trivial baseline.
  • Among content words only — where focus and the real prominence variation live — both models score ≈ 0.05 (spaCy 0.055, Stanza 0.041): no signal. The mature dependency backend has no more than the rough constituency one.
  • Yet content-word prominence is moderately predictable from perception: the human inter-annotator ceiling on content words is 0.258. The text models capture almost none of it.

The reading: neither backend captures focus, because focus is discourse-driven (given/new, contrast) and lives outside a single sentence’s syntax. The whole correlation either model achieves is the function/content distinction — which a one-line lexical rule does better. The spaCy “win” is not more signal; it is less corruption of the lexical baseline. On this target, the entire syntactic gradient is net-negative.

But this is not prosodic’s target. Prosodic analyzes written verse and prose. For metrical analysis the relevant prominence is the linguistic default (broad-focus) prominence that L&P and the NSR actually describe — not the performed acoustic focus a narrator adds. LibriTTS is expressive audiobook narration, the hardest case for default prominence and the furthest from a metrical grid. So the models did not fail to predict prominence; they correctly predicted default prominence, which expressive speech overwrites with focus. For text and meter, default prominence is the right object — and there constituency is the more faithful computer of it.

Decision

  • spaCy stays the default for syntax=True: it is fast (16× faster — ~1.5 s vs ~25 s on the sonnets), and its gradient is at the human ceiling for what a sentence model can predict.

  • The Stanza constituency backend is available as a non-default option, because for metrical and theoretical work — where accuracy to the linguistic structure is the point — it is measurably more faithful (the grid, the binary tree, the s/w relation, and better nuclear placement on the theory’s own examples). It is opt-in:

    TextModel(text, syntax=True, syntax_model="stanza")

    is a one-parameter swap producing the same _syll_df columns as the default spaCy path — phrasal_stress, pstress, tstress, gstress, pstrength — grouped by the same prosodic sentences. The two engines differ only in the parse. Prosodic’s own tokenization is the reference, so clitics (summer's, won't) collapse to their host and punctuation is dropped.

Two prominence columns: tree stress and grid stress

L&P describe prominence two ways, and both are exposed, for both engines, each in [0,1] with nuclear = 1:

  • tstress — tree stress: cumulative domination (their eq-12 “stress numbers” / Dozat’s total). Fine-grained. spaCy computes it natively; Stanza reads stress_numbers() off the LP tree.
  • gstress — grid stress: the RPPR grid height — L&P’s preferred, coarser representation. Stanza reads grid_heights() off the LP tree; spaCy binarizes its n-ary dependency projection (NSR/CSR) into an L&P binary tree and runs the same grid_heights.

So the grid is computed identically across engines — only the tree differs. On the sonnets the two engines’ grids agree closely (coarse, structural) while the tree stresses diverge more (fine, cumulative), which is exactly L&P’s tree-vs-grid distinction made visible. The grid display (grid_str()/grid_plot()) renders gstress; the constraint families w_stress_t/s_unstress_t weight against the tree, w_stress_g/ s_unstress_g against the grid.

Performance: caching

Stanza constituency parsing is the cost (~3.5 s/sonnet, ~16× spaCy). _stanza_parse() caches serialized stanza.Document objects in a HashStash (~/prosodic_data/data/cache/stanza_constituency) keyed by (lang, config-version, text): cold ~3.5 s → warm ~0.01 s (267×), byte-identical. It caches the raw parse, so improving the L&P tree/grid logic never invalidates the cache — only a pipeline-config change does. - The faithful grid + tree already ship as the Line View toggle (see the web app), rendering L&P’s figure with real s/w edge labels.

Does the gradient earn its place? (answered)

An earlier version of this page left open whether the gradient tstress earns its place for any empirical text task, or is purely theoretical. A follow-up MaxEnt study — Phrasal stress & scansion — answers it: yes, tree stress does, for recovering canonical meter. Scored against strict iambic pentameter on Shakespeare it lifts accuracy +8 points (both engines, real learned weights), and it aids trochaic scansion +14; it only fails to help when matching human-marked deviations or on unparseable archaic verse (Milton). That study also settles the tree-vs-grid question — the grid does not scan better than the tree (grid helps only in spaCy, tree in both), so keep tstress as the primary gradient; the grid’s value stays categorical (the figure, teaching, matching L&P).

References

  • Liberman, M. & A. Prince (1977). On stress and linguistic rhythm. LI 8.
  • Morrison, M., P. Pawar, N. Pruyne, J. Cole & B. Pardo (2024). Crowdsourced and automatic speech prominence estimation. ICASSP. Data: Zenodo 10402793 (CC BY 4.0).
  • Dozat, T. (2015). MetricalTree. See Phrasal stress.