Glossary
Key terms and definitions for understanding NarrativePrism data. This glossary covers the concepts that power our narrative intelligence platform.
Content & Narratives
- Narrative
- A distinct story or framing that groups related claims across multiple sources. Narratives represent how a topic is being discussed and debated in the media. Example: "AI Will Replace Most Jobs" is a narrative that aggregates claims from various outlets about automation and employment.
- Claim
- An atomic, verifiable assertion extracted from content. Claims are the building blocks of narratives. Each claim is normalized (reworded to be voice-neutral) and deduplicated across sources. Types include FACTUAL (verifiable), PREDICTIVE (future-oriented), EVALUATIVE (judgment), and PRESCRIPTIVE (should/ought statements).
- Frame Type
- The tone or perspective of a narrative. Helps categorize how a story is being told. Types: ALARMING (warns of danger/crisis), OPTIMISTIC (highlights progress/success), SKEPTICAL (questions claims), NEUTRAL (balanced analysis), CRITICAL (opposes/critiques), SUPPORTIVE (advocates/endorses).
- Novelty
- How established a narrative is. NEW narratives have just emerged, EVOLVING narratives are actively developing with new claims, and ESTABLISHED narratives have stabilized and appear consistently over time.
- Coverage
- Supporting quotes and articles that back up a claim. Coverage links claims to their source material. Types: DIRECT_QUOTE (exact text), PARAPHRASE (close rewording), INFERENCE (implied), AGGREGATED (synthesized from multiple sources).
Voices & Attribution
- Voice
- An entity that produces or promotes content. Voices can be INDIVIDUAL (specific person like a journalist or pundit), OUTLET (publication like NYT or Fox News), or PROGRAM (show or podcast). Voices are distinct from authors—an author writes an article, but the voice represents who is pushing a perspective.
- Voice Attribution
- The process of identifying which voice is behind a piece of content. For articles, this is typically the author. For unsigned editorials, it's the outlet. For YouTube videos, it's the channel. Voice attribution helps track who is saying what across the media landscape.
- Political Leaning
- Where a voice falls on the political spectrum. We use a scale from far-left to far-right: LEFT (progressive), CENTER-LEFT (liberal), CENTER (neutral/balanced), CENTER-RIGHT (conservative-leaning), RIGHT (conservative). Based on AllSides ratings and AI analysis of content patterns.
- Influence Tier
- A measure of a voice's reach and impact. TIER_1 voices are major national outlets and household-name commentators. TIER_2 voices have significant but more niche audiences. TIER_3 voices are local, specialized, or emerging. Tier affects how claims are weighted in analysis.
- Stance
- A voice's position on a specific claim. SUPPORTS means the voice agrees with or promotes the claim. OPPOSES means they argue against it. NEUTRAL means they report without taking a side. Tracking stances reveals where voices agree and disagree.
Topics & Classification
- Topic
- A hierarchical categorization of what content is about. Topics use path notation like "technology/ai/regulation" or "politics/immigration/border". Claims are tagged with topics to enable filtering and analysis by subject area. A claim can have multiple topics.
- Topic Depth
- How specific a topic classification is. A depth of 1 is broad (e.g., "technology"), depth 2 is more specific (e.g., "technology/ai"), and depth 3+ is highly specific (e.g., "technology/ai/regulation"). Deeper topics enable more granular analysis.
- Domain
- A broad category that defines the type of app and its specialized sources, voices, and analysis. Domains include Politics, Technology, Business, Sports, etc. Each domain has curated sources and voice lists appropriate to that beat.
Analysis & Metrics
- Confidence Score
- How certain the AI is about an extraction or classification (0-1 scale). High confidence (0.8+) means the AI is very sure. Medium (0.5-0.8) may need human review. Low (<0.5) indicates uncertainty. Used for claims, voice detection, and stance analysis.
- Comparison
- A head-to-head comparison of how two voices cover the same topic. Comparisons reveal where voices agree, disagree, and what claims each emphasizes. Useful for understanding different perspectives on controversial topics.
- Landscape
- The overall view of active narratives and how they relate. The landscape shows what stories are being told, how prominent they are, and how they connect to each other. Think of it as a map of the current discourse on a topic.
- Signal
- An alert about significant changes in the narrative landscape. Signals notify you when new narratives emerge, existing narratives spike in coverage, or voices shift their stances. Helps you stay on top of rapidly evolving stories.
Data Quality
- Voice Attribution Rate
- The percentage of content that has been attributed to a specific voice. Higher is better—our target is 70%+. Low attribution means we can't track who said what, reducing analysis quality.
- Quote Linking Rate
- The percentage of claims that have supporting quotes. Target is 80%+. Claims without quotes are harder to verify and less useful for analysis.
- Frame Diversity
- How many different frame types are represented in narratives. We aim for 4+ distinct types to ensure we're capturing the full range of perspectives, not just one dominant framing.
- Topic Coverage
- The percentage of claims that have been assigned topic classifications. Target is 70%+. Unclassified claims can't be filtered or analyzed by subject.
Quick Reference: Key Metrics
| Metric | Target | What It Measures |
|---|---|---|
| Voice Attribution | ≥70% | Content with identified voices |
| Quote Linking | ≥80% | Claims with supporting quotes |
| High Confidence Claims | ≥60% | Claims with confidence ≥0.8 |
| Frame Diversity | ≥4 types | Distinct narrative frame types |
| Topic Coverage | ≥70% | Claims with topic assignments |
Political Leaning Scale
LeftCenterRight
Far LeftLeftCenter-LeftCenterCenter-RightRightFar Right
Political leaning is determined by analyzing outlet bias ratings (AllSides, Ad Fontes Media) and content patterns. Individual voices may differ from their outlet's overall lean based on their specific coverage.