Datasets:

bpalacios
/

text2signed-graph-gold

license: cc-by-4.0
task_categories:
  - text-classification
  - token-classification
language:
  - es
tags:
  - signed-graphs
  - relation-extraction
  - political-networks
  - text2graph
  - benchmark
pretty_name: text2signed-graph Gold Standard (Synthetic v2)
size_categories:
  - n<1K
configs:
  - config_name: relations_full
    data_files: relations_full.parquet
    default: true
  - config_name: articles
    data_files: articles.parquet
  - config_name: nodes
    data_files: nodes.parquet
  - config_name: edges
    data_files: gold.parquet

text2signed-graph — Gold Standard (Synthetic v2)

A synthetic benchmark for extracting signed political relation graphs from news articles. Each relation is a directed, signed triple between political actors: (actor_u → act_type → actor_v, polarity), grounded in an evidence quote.

Built to evaluate text2SG extraction pipelines (NER → entity resolution → relation extraction) in a precision-first setting, with planted distractors to penalize over-extraction.

File	Rows	Description
`relations_full.parquet` ⭐	914	the signed graph, denormalized & human-readable — `actor_u → act_type → actor_v, polarity` with actor names + article metadata. Start here.
`articles.parquet`	282	input — synthetic news articles (incl. distractors)
`nodes.parquet`	1,579	graph nodes (actors per article) — normalized form
`gold.parquet`	914	graph edges (by `union_id`) — normalized form
`split.json`	—	`train` (207) / `test` (75) article split + seed
`unions.json`	—	raw nested form of the nodes

`relations_full.parquet` schema (start here — viewer-friendly)

article_id, actor_u, act_type, actor_v, polarity (the signed edge), u_type, v_type, is_reactive, issue, evidence_quote, source, dispute_type, title, dominio, dureza, es_distractor. Each row is one gold edge actor_u → act_type → actor_v (polarity) with everything joined in.

The signed graph (the core)

Each article has a gold signed graph: nodes.parquet (actors) + gold.parquet (signed edges between them). Join them on (article_id, union_id ↔ u_from/u_to).

`nodes.parquet` schema (the actors)

article_id, union_id (e.g. U7), type (roster_actor = known political figure / institutional_actor / non_roster_actor), canonical_name, surfaces (alias list). 1,579 nodes total.

`gold.parquet` schema (the signed edges)

article_id, u_from, u_to (→ nodes.union_id), act_type (e.g. endorses, attacks, allies_with, questions…), polarity (positive / negative / neutral — the edge sign), is_reactive, issue, evidence_quote, source (planted / …), n_inclusion_votes, dispute_type.

`articles.parquet` schema (the input)

article_id, title, body, dominio (topic), dureza (relation difficulty: explicita / mixta / oblicua), es_distractor (distractor flag).

Why synthetic

Real Chilean political news carry copyright (CC-BY-NC). This synthetic gold reproduces the register, difficulty, and structure of real coverage — including distractor articles and near-miss relations — so it can be shared openly as a reproducible benchmark.

Evaluation

Score a predicted relation set against gold.parquet on the test split. The reference extractor (id15 / gemini_champion_v2) reaches f0.5 ≈ 0.89 on detection (undirected), precision-first. Distractors and dispute_type test robustness to over-extraction.

Lineage

This gold standard was built and used to develop the text2SG extraction pipeline:

text2graph-evolve — evolutionary engine that optimizes extraction prompts/genomes against this gold (fitness = f0.5, precision-first). The champion extractors (id15, gemini_champion_v2) were evolved here.
chilean-political-dataset-signed-networks — applies those champions at scale to build a signed political network of Chile (2014–2026): ~480k articles → 442k actor nodes, 2.5M signed edges.

So the flow is: this gold → evolve the extractor → run it on real news → signed graph.

Citation

Part of the text2SG line of work on signed political networks. Built by Benjamín Palacios.

text2signed-graph — Gold Standard (Synthetic v2)

Contents

relations_full.parquet schema (start here — viewer-friendly)