Opinionated, Hesitant and Stressed: Three Studies of How Politicians Speak in Four Slavic Parliaments
Ivan Porupski, Nikola Ljubešić
cs.CL
Quick verdict
Three 6,000h studies tie sentiment to acoustics, model filled pauses, and map Croatian stress coherence; credible but confound control is thin.
Summary
The paper reports three large-scale empirical studies of parliamentary speech in Croatian, Czech, Polish, and Serbian, all built on the ParlaSpeech 3.0 corpus (over 6,000 hours aligned with ParlaMint transcripts). Study 1 relates utterance-level sentiment (XLM-R-ParlaSent predictions) to Praat-extracted pitch, intensity, and speech rate in over one million utterances: negative speech is produced with higher pitch, greater intensity, and faster rate, with a J-shaped upturn at the most positive extreme (inflection near sentiment 3.5) and arousal modulation strongest in Polish. Study 2 models filled-pause frequency with negative binomial GEE over 1,001,787 utterances from 1,561 speakers: speech rate (-36% per SD), age (-16% per decade), and sentiment (+6% per unit) are robust predictors; most notably, male speakers produce far fewer filled pauses than women in South Slavic parliaments (HR -47%, RS -60%) while no gender difference appears in Czech and Polish — a reversal invisible to single-language designs. Study 3 shows Croatian speakers' early-vs-late stress preferences cohere across verbs, adjectives, and nouns (r up to 0.88) but decouple for adverbs and proper nouns (935k tokens, 491 speakers). Main limitations: sentiment is model-predicted rather than human-labeled, multiple-comparison handling is unstated, confounds such as scripted-versus-extemporaneous delivery remain unmodeled, and much of the paper is a forward-looking research agenda; core results are cross-referenced to companion papers.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Negative utterances are produced with higher pitch, greater intensity, and faster speech rate across all four Slavic parliaments, with a cross-linguistically stable J-shaped upturn at the most positive sentiment extreme (inflection at 3.5).
- Filled-pause frequency is robustly predicted by speech rate (-36% per SD), age (-16% per decade), and sentiment (+6% per unit) across parliaments.
- Gender effects on filled pauses reverse between South Slavic parliaments (men produce 47–60% fewer in HR/RS) and West Slavic parliaments (no significant difference), a pattern only cross-linguistic designs can expose.
- Croatian speakers' early-vs-late stress preferences cohere strongly across verbs, adjectives, and nouns (r=0.84–0.88) but decouple for adverbs (r=0.23–0.27) and are only partial for proper nouns (r=0.63–0.70).
Strengths
Unmatched scale and comparability
6,000+ hours, over 1M utterances, 1,561 speakers, and one identical annotation pipeline (same sentiment model, FP detector, stress tagger) across four languages makes the cross-parliament contrasts in Sections 2–4 methodologically meaningful rather than an artifact of toolchain differences.
The gender-reversal finding justifies the cross-lingual design on its own terms
Section 3 shows male speakers produce 47% (HR) and 60% (RS) fewer filled pauses than women while CZ/PL show no difference; the paper correctly argues any single-language study would have merely confirmed or contradicted received wisdom rather than exposing the reversal.
Honest limitation handling
Stress study operationalizes a descriptive debate
Research agenda grounded in existing data layers
Weaknesses
Predicted sentiment treated as ground truth
Thin statistical reporting
Confound structure acknowledged but unmodeled
Headline gender claim lacks a formal interaction test
Results distributed across companion papers
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This short paper packages three empirical studies on the ParlaSpeech 3.0 collection — 6,000+ hours of parliamentary speech in Croatian, Czech, Polish, and Serbian aligned to ParlaMint transcripts (Erjavec et al., 2025) — plus a substantial research agenda (Section 5 occupies a large share of the text). The claims are deliberately scoped as current state: Study 1 (Section 2) claims that utterance sentiment, predicted by XLM-R-ParlaSent, leaves a consistent acoustic trace — negative speech is higher, louder, faster — with a secondary J-curve upturn at the most positive extreme (inflection at sentiment 3.5, the neutral-positive boundary; 11/12 extreme-difference and 12/12 trend tests significant; arousal curvature in 42% of tests, 63% for speech rate). Study 2 (Section 3) claims speech rate (-36% per SD), age (-16% per decade), and sentiment (+6% per unit) robustly predict filled-pause (FP) frequency over 1,001,787 utterances from 1,561 speakers, and that gender effects reverse between South Slavic parliaments (HR -47%, RS -60% fewer FPs for men) and West Slavic parliaments (no difference). Study 3 (Section 4) claims Croatian early-vs-late stress preference is a structured bundle: coherent across verbs, adjectives, and nouns (r=0.84–0.88), partial for proper nouns (r=0.63–0.70), near-independent for adverbs (r=0.23–0.27; adverbs vs. proper nouns r=0.04, n.s.).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ParlaSpeech 3.0?
A public multilingual collection of spoken parliamentary corpora for Croatian, Czech, Polish, and Serbian — over 6,000 hours aligned with official ParlaMint transcripts — enriched with annotation layers including Universal Dependencies, utterance-level sentiment predictions, transformer-based filled-pause detection, and (for Croatian and Serbian) word- and grapheme-level primary stress markers. It is released via CLARIN (handle 11356/1833).
Do men or women hesitate more in parliament?
According to this paper, it depends on the language group. In the South Slavic parliaments (Croatian, Serbian), male politicians produce substantially fewer filled pauses than women (about 47–60% fewer), while in the West Slavic parliaments (Czech, Polish) no significant gender difference was found — reversing the 'men hesitate more' pattern often reported for conversational corpora.
How does negative emotion change how politicians sound?
The paper finds that negative utterances are spoken with higher pitch, greater intensity, and faster speech rate in all four parliaments studied, and that the most positive utterances show a secondary upturn, producing a J-shaped curve whose minimum sits between neutral and positive sentiment (around 3.5 on the sentiment scale).
Is each Croatian speaker's word stress style consistent across the vocabulary?
Not entirely. The study shows a speaker's preference for early vs. late stress is highly coherent across verbs, adjectives, and nouns (correlations up to 0.88), only partially aligned for proper nouns (0.63–0.70), and largely independent for adverbs (0.23–0.27) — so a speaker's stress system behaves as a structured bundle rather than a single setting.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 appear only as captions in the extracted text; the plots themselves (exact CI widths in Figure 2, MDS coordinates in Figure 3) cannot be inspected.
- Per the paper's own references, full model tables live in companion papers (one under review, one accepted); only statistics printed in this text could be checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.