ictchenbo/public-discourse-corpus

Public Discourse Corpus (PDC)

The Public Discourse Corpus (PDC) is the first dataset of public-figure interview speech jointly annotated for affective valence (positive / neutral / negative) and epistemic modality (emphatic / neutral / hedged). The corpus spans 998 videos from 100 speakers across seven professional domains, comprising 186,642 sentences (3.1 million words).

The PDC is the primary contribution of the accompanying paper. Two additional contributions support the corpus: a Target Speaker Participation (TSP) taxonomy for verifying target-speaker presence in automatically retrieved videos (κ = 0.616), and an audio-first diarization pipeline combining local Whisper ASR with pyannote speaker separation.

Note on file naming: the paper refers to the annotation file as anno.jsonl; the released file is named main.jsonl.

Dataset Structure

The release consists of five files: three core files forming the corpus, and two cross-provider validation files.

main.jsonl (186,642 records)

The primary annotation file. Each record has six fields:

Field	Type	Description
video_id	string	YouTube video identifier, joining to videos.jsonl
speaker	string	Speaker slug (e.g., ray_dalio), joining to speakers.jsonl
sentence_index	int	Zero-based position within the video transcript
sentence	string	Cleaned, diarized target-speaker text
valence	string	Affective label: positive / neutral / negative
modality	string	Epistemic modality label: emphatic / neutral / hedged

speakers.jsonl (100 records)

Speaker metadata, keyed by speaker:

Field	Type	Description
speaker	string	Slug (primary key)
name	string	Display name
domain	string	Professional domain (see below)
n_videos	int	Number of videos contributed
video_ids	list	List of video IDs for this speaker

videos.jsonl (998 records)

Video metadata, keyed by video_id:

Field	Type	Description
video_id	string	YouTube identifier (primary key)
title	string	Video title
channel	string	Channel name
upload_date	string	Upload date (YYYY-MM)
duration	string	Video duration (MM:SS or HH:MM:SS)
description	string	Truncated video description

Cross-provider validation files

Two independently annotated 1% validation samples (1,842 records each) produced by GPT-5.5 on the same stratified sample, differing only in reasoning effort:

File	Reasoning effort
gpt5_none_w6-80.0.01.jsonl	none
gpt5_xhight_w6-80.0.01.jsonl	extra_high

Each record contains video_id, sentence_index, text, valence, and modality. These files enable users to bound label uncertainty in their own analyses (see Cross-provider validation below).

Professional Domains

Domain	Slug	Speakers	Videos	Sentences	Examples
Academia/Economics	academia	15	124	26,151	Rogoff, Acemoglu, Krugman, Stiglitz
Central Banking/Policy	central_banking	9	84	10,202	Powell, Yellen, Bernanke, Lagarde
Finance/Investing	finance	21	215	41,319	Dalio, Wood, Ackman, Dimon
Geopolitics/Strategy	geopolitics	14	147	32,385	Zeihan, Mearsheimer, Bremmer, Kissinger
Media/Commentary	media	16	216	47,291	Shapiro, Carlson, Peterson, Klein
Politics/Government	politics	22	183	25,348	Trump, Biden, Clinton, Sanders, Macron
Technology/Business	technology	3	29	3,946	Musk, Altman, Andreessen

Corpus Statistics

Statistic	Value
Total sentences	186,642
Total words	3,112,812
Mean words/sentence	16.7
Speakers	100
Videos	998
Professional domains	7
Temporal range	2016-04 to 2026-06

Sentences average 16.7 words (σ = 10.4), with 72.4% in the 6–20 range. Speaker contributions follow a long-tail distribution (median 1,300, IQR 539–2,865), characteristic of real-world public discourse.

Label Distribution

Valence:

Label	Count	%
Neutral	75,372	40.4%
Negative	64,792	34.7%
Positive	46,478	24.9%

Modality:

Label	Count	%
Neutral	108,220	58.0%
Hedged	44,499	23.8%
Emphatic	33,923	18.0%

Joint distribution (valence × modality):

Valence \ Modality	Emphatic	Neutral	Hedged	Total
Positive	14,396 (31.0%)	21,415 (46.1%)	10,667 (22.9%)	46,478 (24.9%)
Negative	16,562 (25.6%)	32,789 (50.6%)	15,441 (23.8%)	64,792 (34.7%)
Neutral	2,965 (3.9%)	54,016 (71.7%)	18,391 (24.4%)	75,372 (40.4%)
Total	33,923 (18.0%)	108,220 (58.0%)	44,499 (23.8%)	186,642

Construction Methodology

The PDC is constructed through a four-stage pipeline:

1. Data Collection: YouTube videos retrieved via yt-dlp search for 100 public figures across seven domains (2,479 candidate videos from 124 candidate speakers).
2. Target Speaker Participation (TSP) Annotation: A five-category manual annotation taxonomy verifies that each retained video contains analyzable speech by (not merely about) the target speaker. Videos are classified as TSP-0 (target absent), TSP-1 (referenced subject), TSP-2 (multi-party discussion), TSP-3 (sole-respondent interview), or TSP-4 (solo presentation). Only TSP-2/3/4 are retained. Inter-annotator agreement is substantial (Cohen's κ = 0.616 for five-way classification; κ = 0.690 for binary inclusion).
3. Audio-Based Speaker Diarization: A local pipeline combining faster-whisper (large-v3) for ASR and pyannote/speaker-diarization-3.1 for speaker separation extracts target-speaker utterances. 998 of 1,088 diarized transcripts (91.7%) are rated GOOD or FAIR by a multi-dimensional quality assessment; the remaining 90 are discarded.
4. LLM-Based Annotation: DeepSeek-V4-Flash performs zero-shot dual-dimension classification (valence + modality) on all 186,642 sentences.

Cross-provider validation

A stratified 1% sample (1,842 sentences) was independently re-annotated by GPT-5.5 to assess label stability across providers:

Metric	Valence	Modality
Exact agreement	75.9%	72.5%
Cohen's κ (unweighted)	0.63	0.53

Per-domain agreement (DeepSeek-V4-Flash vs. GPT-5.5, reasoning = none):

Domain	N	Valence Agr.	Valence κ	Modality Agr.	Modality κ
Media/Commentary	472	77.8%	0.66	71.4%	0.52
Finance/Investing	413	75.5%	0.62	71.9%	0.50
Geopolitics/Strategy	323	74.0%	0.59	74.0%	0.57
Academia/Economics	261	75.5%	0.60	70.1%	0.46
Politics/Government	253	73.9%	0.60	73.1%	0.55
Central Banking/Policy	102	79.4%	0.66	77.5%	0.60
Technology/Business	39	79.5%	0.67	76.9%	0.64

The two released validation files (reasoning none vs. xhigh) further allow users to measure sensitivity to provider reasoning settings.

Usage

import json

# Load annotations
records = []
with open("main.jsonl") as f:
    for line in f:
        records.append(json.loads(line))

# Load speaker metadata and index by domain
speakers = {}
with open("speakers.jsonl") as f:
    for line in f:
        sp = json.loads(line)
        speakers[sp["speaker"]] = sp

# Get all sentences from finance speakers
finance_speakers = {s for s, sp in speakers.items() if sp["domain"] == "finance"}
finance_sentences = [r for r in records if r["speaker"] in finance_speakers]

# Cross-tabulation of valence × modality
from collections import Counter
cross = Counter((r["valence"], r["modality"]) for r in records)
for (v, m), c in cross.most_common():
    print(f"{v:>10} + {m:>10}: {c:>6}")

# Load a validation file and compare against main annotations
val = {}
with open("gpt5_none_w6-80.0.01.jsonl") as f:
    for line in f:
        r = json.loads(line)
        val[(r["video_id"], r["sentence_index"])] = r

agree = mismatched = 0
for r in records:
    key = (r["video_id"], r["sentence_index"])
    if key in val:
        if r["valence"] == val[key]["valence"] and r["modality"] == val[key]["modality"]:
            agree += 1
        else:
            mismatched += 1
print(f"Full-label agreement: {agree}/{agree + mismatched}")

Researchers can filter by domain, TSP category, or speaker to construct analysis subsets; compute per-speaker valence and modality profiles; and analyze cross-dimension relationships at multiple levels of aggregation.

Illustrative Analysis: Valence–Modality Correlation

To demonstrate one research question the PDC enables, the paper reports cross-dimension phi correlations (mean per-speaker φ):

Domain	φ(pos,emp)	φ(neg,emp)	φ(pos,hed)	φ(neg,hed)
Overall (pooled)	0.191	0.140	−0.012	−0.000
Academia/Economics	0.153	0.162	−0.002	−0.023
Central Banking/Policy	0.225	0.110	0.009	0.028
Finance/Investing	0.230	0.101	0.009	0.014
Geopolitics/Strategy	0.113	0.168	0.030	−0.007
Media/Commentary	0.171	0.163	−0.019	−0.024
Politics/Government	0.211	0.125	−0.020	−0.033
Technology/Business	0.519	0.056	−0.054	−0.047

A descriptive asymmetry is visible: across most domains, positive statements are more likely to be emphatic than negative ones (pooled φ(pos,emp) = 0.191 vs. φ(neg,emp) = 0.140). Correlations involving hedged modality are near zero across all domains. These are presented as descriptive patterns that may motivate confirmatory studies with controlled designs.

Limitations

• Speaker attribution: The local diarization pipeline has been validated for transcript consistency against AssemblyAI (Jaccard 0.706, 84.1% word overlap on 212 files), but direct human evaluation of per-utterance speaker-attribution accuracy has not yet been conducted. The existing ground truth (60 utterances across 3 Ray Dalio files, 91.7% accuracy) applies to the AssemblyAI pipeline, not the local pipeline.
• LLM annotations: All annotations are produced by DeepSeek-V4-Flash. Cross-provider validation spans two providers on one sample. Human gold-standard annotation is in progress.
• Coverage: English-language public-figure interviews sourced from YouTube only. Video selection involves researcher judgment. Findings may not generalize to languages with different modality-marking systems.
• Computational requirements: The local diarization pipeline requires a consumer GPU, though it eliminates per-file API costs entirely.

Ethics Statement

The PDC consists of publicly available YouTube interviews featuring public figures. All content was already in the public domain at the time of collection. We do not redistribute video or audio; only cleaned, diarized text transcripts and their annotations are released. Speaker names are preserved because the corpus's value lies in its speaker-attributed structure, but we encourage users to consider ethical implications of individual-level analysis, particularly for non-public-figure applications of the methodology.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Citation

If you use the PDC in your research, please cite:

@article{chen2026pdc,
  title   = {The Public Discourse Corpus (PDC): A Speaker-Attributed Dataset
             for Valence and Epistemic Modality with Target Speaker Participation},
  author  = {Bo Chen},
  journal = {arXiv preprint},
  year    = {2026},
  url     = {https://huggingface.co/datasets/ictchenbo/public-discourse-corpus}
}

Links

• Dataset: https://huggingface.co/datasets/ictchenbo/public-discourse-corpus
• Pipeline: https://github.com/ictchenbo/pdc-construction-pipeline
• Annotation tool: https://github.com/ictchenbo/pdc-annotator

Contact

For questions, corrections, or collaboration inquiries, please open an issue on the dataset repository or the pipeline repository.