# News Metadata Dataset (7K): Media Bias, Narrative Signals, and Readability Features This **news metadata dataset** is designed for reproducible analysis of **media bias**, **narrative signals**, readability, and other interpretable **NLP features** in a **tabular CSV** format. It supports research and **machine learning** feature engineering on attribution, uncertainty, framing, and content-structure indicators without redistributing full article text. This is an **open metadata-only dataset** intended for transparent benchmarking and repeatable downstream experiments. ## Dataset At A Glance - Artifact: `nnai_news_metadata_narrative_signals_7k_v1.csv` - Rows: `7,000` - Columns: `22` - Language: English (`en`) - Format: CSV - Public sample policy: metadata only, no full article body text ## Quick Start ```python import pandas as pd df = pd.read_csv("nnai_news_metadata_narrative_signals_7k_v1.csv") print(df.shape) print(df.head(3)) ``` Non-Python option: open `nnai_news_metadata_narrative_signals_7k_v1.csv` directly in a spreadsheet tool (for example, Excel, Numbers, or Google Sheets). ## What's Included - Identity: `source_id`, `title`, `url`, `source`, `source_type` - Labels: `topic`, `political_orientation`, `bias_strength`, `credibility_level`, `clickbait_level`, `subjectivity_level` - Structural: `sentence_count`, `word_count`, `claim_density_per_1k_words`, `quote_density`, `passive_voice_ratio` - Narrative Alpha: `sentiment_extremism_density_per_1k_words`, `hedging_rate`, `primary_source_ratio`, `headline_body_alignment_score` - Readability: `readability_reading_level`, `readability_flesch_grade` ## Column Notes | Column | What it means | Possible values | How to interpret | |---|---|---|---| | `source_id` | Unique ID for one article record. | String/integer ID (non-empty). | Primary key for joins, deduping, and traceability. | | `title` | Article headline from the source page. | Free text. | Core summary of the article framing. | | `url` | Source URL for the article. | Valid HTTP/HTTPS URL. | Use to fetch/review source content (if rights allow). | | `source` | Publisher/outlet name. | Free text (for example `BBC News`, `CNN`). | Human-readable brand/source reference. | | `source_type` | Normalized source class from registry mapping. | `outlet`, `wire`, `primary_doc`, `fact_checker`. | `wire` and `primary_doc` are often useful for provenance-style segmentation. | | `topic` | Article topic/category label. | Current sample: `Politics`, `Business`, `Science`, `War`, `Technology`. | High-level subject grouping for filtering and benchmarking. | | `political_orientation` | Coarse political leaning label. | Current sample: `left`, `center`, `right`. | Comparative signal; not a definitive ideological truth label. | | `bias_strength` | Estimated strength of bias in writing style/content framing. | Current sample: `low`, `medium`, or empty (`null`). | Empty means unavailable; higher category implies stronger bias signal. | | `credibility_level` | Estimated reliability/credibility category. | Current sample: `high`, `medium`, `low`, or empty (`null`). | Empty means unavailable; use with other fields, not standalone. | | `clickbait_level` | Estimated clickbait intensity. | Current sample: `low`, `medium`. | Higher level suggests more attention-grabbing headline style. | | `subjectivity_level` | Estimated subjective vs factual writing style. | Current sample: `low`, `medium`. | Higher level indicates more opinionated/interpretive tone. | | `sentence_count` | Number of sentences in analyzed text. | Integer (`>= 0`). In sample: `6` to `1857`. | Length/structure feature; useful normalization denominator. | | `word_count` | Number of words in analyzed text. | Integer (`>= 0`). In sample: `146` to `51474`. | Core length feature; used to normalize density metrics. | | `claim_density_per_1k_words` | Frequency of extracted factual claims per 1,000 words. | Numeric (`>= 0`). In sample: `0.5828` to `101.3514`. | Higher values indicate denser explicit claim language. | | `quote_density` | Quote-span intensity relative to text structure. | Numeric (`>= 0`). In sample: `0.0` to `22.2381`. | Higher values indicate heavier use of quoted material. | | `passive_voice_ratio` | Share of sentences detected as passive voice. | Numeric ratio in `[0,1]`. | Higher values can indicate less direct attribution/accountability language. | | `sentiment_extremism_density_per_1k_words` | Frequency of extreme/sensational sentiment terms per 1,000 words. | Numeric (`>= 0`). In sample: `0.0` to `34.9650`. | Higher values suggest more emotionally intense framing. | | `hedging_rate` | Share of sentences containing hedging/speculative language. | Numeric ratio in `[0,1]`. | Higher values suggest more uncertainty/speculation cues. | | `primary_source_ratio` | Proportion of attributed quote words vs article words. | Numeric ...