Situational Context: MetaBias (Newsvoy.com)

Polynym: a superordinate function with subordinate, measurable cues.

  • Coverage = {Selection, Omission, Placement}
  • Framing = {Angle, Lexicon, Contrast}
  • Sourcing = {Authority, Balance, Disclosure}

PD: (Positive Displacement): selection pressure from article toward bias label.

This polynym-based system differs fundamentally from most media bias ranking sites in its strategic goal, level of granularity, actionability, and direct connection to naming as a strategic act (Nymology).

Here is a breakdown of the key differences:

Feature Polynym/PD (Nymology) Approach Standard Media Bias Ranking Sites (e.g., Ad Fontes, AllSides)
1. Strategic Goal Strategic Analysis & Counter-Action. The goal is to gain insight into how a specific bias operates (the strategic push) and to generate actionable fixes (counter-actions). It’s about strategy. Labeling & Positioning. The goal is to give a source a static, overall rating (e.g., “Left-Center,” “Mixed Reliability”) to help consumers choose content. It’s about consumption.
2. Unit of Analysis & Granularity Article or Quote-Level. Measures bias per-article, per-quote, or per-story element. It provides three distinct scores ($PD_{\text{Coverage}}, PD_{\text{Framing}}, PD_{\text{Sourcing}}$) for why the article pushes a bias label. Outlet-Level (Primary). Measures the average, overall bias of a news outlet over time. While some track individual articles, the primary output is the static position of the source.
3. Measurable Focus The Nymological Strategy (Selection Pressure). Focuses on measurable, low-level features grouped by the strategic intent of the editor/writer: what did they select, how did they frame it, and who did they source. The score $PD$ is a quantified selection pressure. Political Orientation & Reliability. Focuses on the political direction (left/right) and factual reliability/veracity (high/low). They measure where the source sits ideologically.
4. Normalization & Context Dynamic & Contextual. Cues are normalized against both Topic and Outlet baselines. This prevents “house style” (an outlet’s normal tone) from being rated as “bias” and allows the system to focus on the deviation (the strategic push) in the moment. Static/Panel-Driven. Ratings are often based on large, weighted averages of human analysis (e.g., panels of left/center/right reviewers) or citation patterns, making them less sensitive to topic-specific, real-time strategic shifts within a single article.
5. Actionability (The “So What?”) Directly Actionable. The system directly outputs Counter-Actions tied to the top drivers. (“Driver: High Assertive Lexicon ($D$) $\rightarrow$ Action: Introduce Nuance/Alternative Framing”). This is useful for editors or agents seeking to adjust the strategic naming. Informational Only. The output is a label or a coordinate on a chart. It tells the reader or advertiser what the bias is, but not how to change it or how the bias was constructed at the editorial level.

In short, the Polynym/PD approach moves from simply labeling what a source is (the political position) to analyzing what an article is doing (the strategic naming and selection pressure), providing a toolkit (the counter-actions) to respond to that strategy.

A faceted navigation system

Each polynym aggregates its cues into a single PD score and identifies the top drivers and counter-actions. Let’s determine the best way to provide programs for these categories. Meta-bias analysis can use these categories and apply them to articles from different outlets—such as left-leaning and right-leaning sources. The idea is to categorize the lexicon used by each outlet. This allows us to evaluate how words are framed, as well as which words are selected or omitted. The goal is to gain insight into how a particular bias operates within the article or quote.

Reuters (moderate): From President Donald Trump * At a White House press conference on November 4, 2020, as votes were still being counted, Donald Trump said, “This is a fraud on the American public. This is an embarrassment to our country. We were getting ready to win this election. Frankly, we did win this election”. * Trump later tweeted on November 7, 2020, as news organizations called the race for Joe Biden, “I WON THIS ELECTION, BY A LOT!”. * In the years after the election, Trump continued to allege widespread fraud, referring to it as “the scam of the century” and “the crime of the century”.

What is the approach?

Treat “bias” as measurable selection pressure (PD) produced by article features, organized under three Polynyms:

  • Coverage = {Selection, Omission, Placement}
  • Framing = {Angle, Lexicon, Contrast}
  • Sourcing = {Authority, Balance, Disclosure}

Each Polynym aggregates its facet-level cues into a single PD ∈ [0,1] for that category. PD answers: “How strongly do the observable features of this article push toward a bias label?”

How it works (simple pipeline)
  1. Extract cues from the article (text + media) → e.g., charged headline verbs, early/late counter-quotes, primary-source share.
  2. Normalize each cue against topic/outlet baselines (z-scores or percentiles) so “house style” and subject matter are accounted for.
  3. Aggregate per Polynym with a bounded model (e.g., logistic over a weighted sum):PDPolynym=σ ⁣(∑wf⋅zf+b)\text{PD}_{\text{Polynym}} = \sigma\!\Big(\sum w_f \cdot z_f + b\Big)Keep feature contributions wf⋅zfw_f \cdot z_f to show top drivers.
  4. Propose a label when PD is high (e.g., framing_bias if Framing PD crosses a threshold).
  5. Suggest counter-actions tied to drivers (e.g., “move counter-quote to ¶3”, “link primary document in lede”).
Why this is practical
  • Operational & auditable. Every PD comes from concrete, inspectable cues (with snippets and media placements), not vibes.
  • Outlet-aware. Baselines by topic and outlet prevent false positives caused by house style or beat conventions.
  • Compact & comparable. Three PDs (Coverage/Framing/Sourcing) let analysts compare articles and track outlets over time.
  • Actionable. Driver-specific fixes translate directly into newsroom edits (front-loading counters, adding denominators, linking sources).
  • Modular. You can add/retire cues inside a Polynym without breaking the overall PD design.
Tiny example
  • Framing cues: “crack down” in headline, high sentiment intensity in lede, strong headline–body mismatch → PD_framing ≈ high
  • Counter-action: Replace charged verb with policy noun; add lede link to the underlying ordinance.

That’s it: Polynyms organize measurable cues; PD quantifies their push toward a label; drivers explain why; counter-actions show how to fix.:

Core definition
  • PD (Positive Displacement): selection pressure from article features (propositional, via Polynyms) toward a bias label.
  • Computation: normalize cues vs topic/outlet baselines, then aggregate with a bounded model (e.g., logistic). Keep per-cue contributions for explainability.
Coverage SM (Selection · Omission · Placement)
Facet Feature hook (name) Measurement (how) Direction Typical drivers Counter-actions (examples)
Selection selection_ratio present_entities / expected_entities lower → PD↑ thin_entity_set Add missing primary entities early
denominator_balance % use of rates/ratios vs raw counts in early paras lower → PD↑ no_denominators Convert counts→rates in ¶1–3
Omission counter_evidence_presence share of paras with counter-views lower → PD↑ late_counter_view Move counter-quote to ≤ ¶3
expected_but_missing weighted list of absent key facts higher → PD↑ key_fact_omitted Insert required fact + link
Placement coverage_placement_curve first position of neutral/balancing info (scaled 0–1) lower → PD↑ late_key_fact Front-load neutral timeline
placement_entropy spread of key facts across first N paras lower → PD↑ front_skew Disperse facts earlier

Aggregation (Coverage): PDC=σ(∑wfzf+b)\text{PD}_C = \sigma(\sum w_f z_f + b) • Baselines: {topic, outlet, global} • Outputs: pd, drivers[], driver_evidence_ids[].

Framing SM (Angle · Lexicon · Contrast · Visuals)
Facet Feature hook (name) Measurement (how) Direction Typical drivers Counter-actions (examples)
Angle angle_bias narrative template match score higher → PD↑ one_note_angle Add policy/process angle in lede
Lexicon lexicon_polarity charged vs neutral score (outlet-aware) higher → PD↑ charged_headline_lexicon Swap verb to neutral/policy noun
sentiment.intensity per-span intensity (headline/lede) higher → PD↑ absolutist_claims Add qualifiers; soften modals
Contrast contrast_markers_early “but/however/yet” presence early lower → PD↑ contrast_after_denouement Move contrast to ¶2–3
headline_body_mismatch embedding/register mismatch higher → PD↑ mismatch_headline_body Align headline with body evidence
Visuals visual_placement_pressure hero image salience above fold higher → PD↑ conflict_image Replace with neutral chart/graphic

Aggregation (Framing): same logistic combo; weights modulated by OutletLexicon. Outputs as above.

Sourcing SM (Authority · Balance · Disclosure)
Facet Feature hook (name) Measurement (how) Direction Typical drivers Counter-actions (examples)
Authority authority_mix_primary_pct primary docs / all sources lower → PD↑ low_primary_pct Link & quote primary doc in lede
Balance ideological_spread calibrated diversity across sources lower → PD↑ narrow_spread Add ideologically distant source ≤ ¶3
balance_ordering opposition placement & depth lower → PD↑ bury_opposition Promote opposition quote earlier
Disclosure disclosure_presence COI/funding/affiliation fields present lower → PD↑ missing_disclosure Add disclosure under byline

Aggregation (Sourcing): logistic with emphasis on spread + primary share. Outputs as above.

Shared mechanics (all Polynyms)
  • Normalization: per feature zf=(x−μtopic,outlet)/σz_f = (x – \mu_{\text{topic,outlet}})/\sigma (or percentile). Store baseline IDs in the card.
  • PD score: bounded [0,1][0,1]; keep feature_effects = w_f · z_f for explainability.
  • Evidence linking: every feature cites evidence_id spans (headline/lede/quote/paragraph/media).
  • Counter-actions: map drivers → concrete edits with expected_delta (PD↓) and cost_est.

This SM is ready to wire into extraction → normalization → PD aggregation → proposals (labels) with clear drivers and fixes.