Datasetpublic

bpalacios/chilean-political-signed-network

Chilean Political Signed Network (2014–2026) A signed directed network of Chilean political actors, extracted at scale from ~480k news articles spanning three governments (Bachelet II, Piñera II, Boric). Each edge is a polarized political act between two actors: actor_u → act_type → actor_v with a sign (+1 ally / −1 antagonist / 0 neutral). 480,002 articles → 239,684 connected actor nodes → 2,539,835 signed edges Window: 2014–2026 · 3 administrations · two constitutional… See the full description on the dataset page: https://huggingface.co/datasets/bpalacios/chilean-political-signed-network.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes45downloads

Dataset Card

Chilean Political Signed Network (2014–2026)

A signed directed network of Chilean political actors, extracted at scale from ~480k news articles spanning three governments (Bachelet II, Piñera II, Boric). Each edge is a polarized political act between two actors: actor_u → act_type → actor_v with a sign (+1 ally / −1 antagonist / 0 neutral).

—480,002 articles → 239,684 connected actor nodes → 2,539,835 signed edges
—Window: 2014–2026 · 3 administrations · two constitutional processes (2020, 2023)

⚠️ No news text is included. The source articles are copyrighted (CC-BY-NC), so this release contains only the extracted graph + metadata + `article_id`s — no article bodies, titles, or evidence quotes. article_id (md5(body)) lets holders of the licensed corpus join back to the source.

Files

File	Rows	Description
`nodes.parquet`	239,684	actors: `node_id`, `canonical`, `node_type`, `degree`, `pos_degree`, `neg_degree`, `n_mentions`, `curated`
`edges.parquet`	2,539,835	signed edges: `from_node_id`, `to_node_id`, `article_id`, `act_type`, `polarity`, `sign` (+1/−1/0), `issue`, `publish_date`, `period`
`articles_meta.parquet`	480,002	`article_id`, `source`, `publish_date`, `year`, `period` (no title/body)

Build the graph by joining edges.from_node_id / edges.to_node_id to nodes.node_id.

Statistics


Articles	480,002 (2014–2026)
Nodes (connected)	239,684
Signed edges	2,539,835
Polarity	42% negative · 30% positive · 28% neutral
Node types	person 140,603 · org 46,439 · institution 39,311 · party 6,390 · coalition 4,426 · movement 2,515

Top actors by signed degree: Gobierno de Chile, Gabriel Boric Font, Sebastián Piñera, Senado, José Antonio Kast, Michelle Bachelet, Donald Trump, …

How it was built (the text2SG pipeline)

A three-pass extraction, precision-first (a false edge pollutes the graph; a missed one only omits):

1.NER (entities). GLiNER (urchade/gliner_multi-v2.1, zero-shot, local, fp16) tags typed mentions: person / party / institution / coalition / movement / org. ~$0, deterministic.
2.Entity resolution (nodes). Surface forms are collapsed into canonical actors via token-blocked fuzzy clustering with cross-type and surname guards, then curated: a deterministic layer (exact-name + acronym dictionary, e.g. UDI ↔ Unión Demócrata Independiente) followed by a fan-out of LLM judges (6 parallel Sonnet adjudicators) that resolve the semantic grey zone (Ejecutivo / La Moneda → Gobierno de Chile; Fiscalía → Ministerio Público), strictly precision-first (never merges distinct people, never crosses types).
3.Relation extraction (signed edges). A tuned LLM extractor (gemini-2.5-flash-lite, given the resolved actors, abstaining without evidence) emits (actor_u → act_type → actor_v, polarity) grounded in an evidence quote. Validated at f0.5 ≈ 0.89 against a synthetic gold standard. The edge `sign` is the extracted polarity (positive=+1, negative=−1, neutral=0).

How it was validated (community detection + structural balance)

The network is validated not edge-by-edge but on its aggregate structure — the test that decides whether it is usable for analysis:

—Community detection (Louvain on the positive/ally subgraph) reproduces the real Chilean political coalitions and their evolution across 12 years: Nueva Mayoría (Bachelet era) → Apruebo Dignidad (Boric era); the Communist Party bloc; and the split of the traditional right (Chile Vamos) from the republican right (Kast / Kaiser). Foreign actors cluster by ideology (Lula / Evo / Maduro with the left; Trump / Bolsonaro / Milei / Macri with the republican right).
—Structural balance. *92% of the negative edge weight falls between communities* (enemies are separated), consistent with balance theory — strong evidence the signed structure is reliable.

⚠️ Temporal role nodes. Actors like Gobierno de Chile, Oposición, Oficialismo, Presidente change referent at each change of government (e.g. Piñera → Boric, March 2022). Analyze their dyads per `period`/`year`, not aggregated over the whole window. Stable actors (people, parties) aggregate fine across all years.

Lineage

Built with the text2SG pipeline, evolved against an open synthetic benchmark:

1.[text2signed-graph-gold](https://huggingface.co/datasets/bpalacios/text2signed-graph-gold) — the synthetic gold standard (287 articles, 914 signed relations).
2.[text2graph-evolve](https://github.com/bpalas/text2graph-evolve) — evolutionary engine that optimized the extractor against that gold (precision-first, f0.5 fitness).
3.[chilean-political-dataset-signed-networks](https://github.com/bpalas/chilean-political-dataset-signed-networks) — this dataset: the champion applied at scale to 2014–2026.

Citation

By Benjamín Palacios. Part of ongoing work on signed political cleavage networks in Chile.