Purpose

Identify and track recurring topics, themes, and frames across the speech corpus to understand:

1. What the speaker talks about (topic distribution)
2. How the speaker frames each topic (frame analysis)
3. When topics emerge and decline (longitudinal coverage)
4. For whom topics are emphasized (audience-stratified analysis)
5. Why specific framings recur (frame consistency, escalation, drift)

Topic Modeling

Approach Selection

• BERTopic (recommended for modern Trump corpus): Transformer-embedding-based topic modeling; handles short and long texts; produces coherent topic clusters
• LDA (Latent Dirichlet Allocation): Classical topic modeling; well-understood; baseline comparison
• Top2Vec: Embedding-based; alternative to BERTopic
• Hierarchical topic modeling: For nested topic structure (e.g., “immigration” → “border wall”, “ICE”, “asylum policy”)

Topic Extraction Workflow

1. Preprocessing: Lemmatize, remove stopwords; preserve named entities
2. Embedding: Use sentence-transformers (e.g., all-mpnet-base-v2) for semantic embeddings
3. Clustering: HDBSCAN or k-means; tune for desired topic granularity
4. Topic representation: c-TF-IDF (BERTopic default) or LLM-generated topic names
5. Topic stability assessment: Run topic model on bootstrap samples; assess topic-stability score

Output Schema

{
  "topic_id": 0,
  "label": "Border / Immigration",
  "top_terms": ["border", "wall", "illegal", "immigrants", "mexico", "ice", "deportation", "caravan"],
  "representative_speeches": ["...id...", "...id..."],
  "speech_count": 234,
  "first_observed": "2015-06-16",
  "last_observed": "2026-04-30",
  "trend": "stable_high"
}

Frame Analysis

Frames are how a topic is presented — the perspective, attribution, and evaluative loading. For each speech-topic instance, code:

Frame Dimensions

• Problem definition: What is the problem? Who is responsible?
• Causal interpretation: What caused the problem?
• Moral evaluation: How is the situation morally framed?
• Treatment recommendation: What is the proposed solution?

(Following Entman’s frame-analysis framework; widely used in political-communication research.)

Common Trump-Era Frames (illustrative, not exhaustive)

• Crisis-and-rescue frame: A specific group/situation is destroying America; only the speaker can fix it
• Reverse-victim frame: Trump or in-group is the real victim, despite holding power
• Conspiracy-of-elites frame: A small group of elites conspires against the in-group
• Witch-hunt frame: Investigation or oversight reframed as politically-motivated persecution
• Lawlessness frame: Out-group activities reframed as criminal even when lawful (asylum-seekers as “invaders”)
• Patriot frame: In-group activities reframed as patriotic even when illegal (J6 as “patriots”)
• Strength-vs-weakness frame: Policy positions framed as strong/weak rather than effective/ineffective
• Loyalty-vs-betrayal frame: Officials, family, and allies framed by loyalty rather than competence
• America-First-zero-sum frame: Other countries’ gains framed as American losses
• Fake-news frame: Critical reporting reframed as fabricated
• Deep-state frame: Career civil servants reframed as adversarial conspiracy

Frame Consistency and Drift

For each frame, track:

• Frame consistency: Does the same frame recur across speeches?
• Frame escalation: Does the frame become more extreme over time?
• Audience-targeted framing: Does the frame vary by audience type?
• Frame substitution: When does Frame A get replaced by Frame B?
• Co-occurrence patterns: Which frames co-occur in the same speech?

Recurring Phrases / Slogans / Memes

Track specific recurring linguistic units:

• Slogans: “Make America Great Again”, “Drain the Swamp”, “Lock Her Up”, “Build the Wall”, “Stop the Steal”, “America First”, “Witch Hunt”
• Epithets: “Sleepy Joe”, “Crooked Hillary”, “Pocahontas”, “Lyin’ Ted”, “Little Marco”, “Crazy Nancy”
• Catchphrases: “Believe me”, “Many people are saying”, “We’ll see what happens”, “Nobody knows X better than me”, “I have the best X”
• Conspiracy keywords: “Deep state”, “Fake news”, “Witch hunt”, “Hoax”, “Rigged”
• Apocalyptic phrasing: “Destroying our country”, “Last election”, “End of America”, “Country we won’t recognize”

For each recurring item:

• Track first appearance, frequency by year, audience-stratification
• Note any morphological variants (e.g., “fake news” → “fake news media” → “fake news enemy of the people”)
• Document evolution: how did the phrase change over time?

Methodological Stack

• BERTopic for topic modeling
• spaCy for NER and lemmatization
• scikit-learn for n-gram extraction
• NLTK / spaCy phrase matchers for slogan tracking
• Manual coding for frame analysis (with inter-rater reliability)
• Streamlit / Dash for interactive frame-and-topic exploration

Output Products

The framework produces:

1. Topic dashboard: Topic distribution over time, by audience, by event
2. Phrase-frequency tracker: Daily/weekly/monthly phrase frequency
3. Frame database: Coded frame instances with context and evidence
4. Comparative reports: Trump vs. baseline (other speakers; Trump 2016 vs Trump 2025)
5. Alert pipeline: New phrases / new frames flagged for editorial review

Limitations

• Topic-modeling stability: Topic boundaries depend on hyperparameters; document the model version
• Frame-coding subjectivity: Frame analysis involves judgment; document coding criteria and inter-rater reliability
• Out-of-vocabulary terms: New events introduce new vocabulary; topic models drift
• Audience effects: Topics shift by audience; stratify when comparing
• Concept-vs-term ambiguity: Same concept may be expressed in different terms; same term may refer to different concepts

See Also

• sc-overview.md
• Analytical Framework — Rhetorical Analysis (overlapping marker taxonomy)
• Analytical Framework — Linguistic Features
• sc-framework-claim-tracking.md
• Narcissstic Staff, Narcissistic President (Patriot KB)