State of the Nation Addresses (SONAs) are far more than just political theater. They are dense, multi-topic policy artifacts—part agenda-setting, part accountability ledger, and part national mood-board. If you want to understand what a government is prioritizing (and how it wants you to feel about it), the SONA is your richest available dataset.
But how do we move beyond “vibes” and into real policy analytics? The answer lies in a balanced approach to AI: The “Large-Small” Model strategy.
Why “Large-Small” Models?
In the current AI hype cycle, the instinct is to throw a Large Language Model (LLM) at every problem. But for public-sector discourse analysis, LLMs aren’t always the right hammer. A more productive, auditable approach uses a two-tier system:
- Small Models / Classical NLP for Measurement: Use these for token stats, n-grams, topic models, and sentiment trajectories. They provide a stable, repeatable signal.
- Large Models for Interpretation: Use LLMs to summarize measured changes, generate narrative explanations, and explore “what if” scenarios.
This “Large-Small” pattern ensures auditability. You don’t just get a summary; you get a summary backed by hard, reproducible data.
The Rule of Thumb: Small tools measure reality; large tools help humans understand it.
The Project: SONA 2025 vs. 2026
I’ve developed a reproducible pipeline that performs a comparative semantic analysis between the 2025 and 2026 addresses. This isn’t just a toy demo; it’s a framework for data understanding.
🔗 Project Repo: sona-semantic-comparative-plus
What the Pipeline Generates:
With a single command (make report), the system produces a suite of analytical lenses:
| Feature | Analysis Type | Insight Provided |
| Top Bigrams | Frequency | What specific phrases (e.g., “economic growth”) dominate? |
| Topic Drift (LDA) | Modeling | How does the emphasis shift from the beginning to the end of the speech? |
| Concept Mind Maps | Network | Which concepts (e.g., “energy” + “investment”) appear together? |
| Sentiment Trajectory | Emotional Flow | How does the emotional tone ebb and flow across paragraphs? |
| Policy Classifier | Categorization | How much “airtime” do Water, Crime, or Health actually get? |
4 Key Findings: What Changed in 2026?
By running SONA 2025 and 2026 through this pipeline, several fascinating shifts emerge:
- Increased Provincial Granularity: In 2026, references to specific South African provinces increased. This suggests a shift toward domestic “delivery” language—focusing on where things are happening, not just what the policy is.
- Inward Focus: There was a marked decrease in mentions of foreign countries and organizations. In a shifting global landscape, the 2026 address appears to tilt toward internal consolidation and domestic priorities.
- From “Work” to “National”: While “work” remains a pillar, the term “National” saw a frequency spike in 2026. This points to a rhetorical shift toward national unity, sovereignty, and “whole-of-society” framing.
- The Persistent Constants: Despite the shifts, four themes remain the “bedrock” of the discourse: Service Delivery, Jobs, Water, and National Unity. These are the core latent topics that persist year-over-year.
Beyond Analysis: Can AI Write a SONA?
Yes, LLMs can generate a SONA-style address. But the difference between a generic speech and a credible one is grounding.
A sensible workflow for civic tech isn’t just “prompt and pray.” It’s:
- Analyze historical patterns (using the small models).
- Correlate with socio-economic data (inflation, unemployment, energy stability).
- Generate with constraints based on those measurements.
- Audit the output using the same measurement framework to ensure it hasn’t hallucinated.
Conclusion: Data Understanding First
If you are a student, analyst, or builder in the civic tech space, this repository offers a baseline that is reproducible, explainable, and extensible.
Semantic analysis isn’t an optional extra—it’s the first step toward building useful, responsible AI in political contexts. Before we ask AI to be “intelligent,” we must first use it to understand the data we already have.