Sven Geboers
d2310edfc4
feat(right-wing): LLM-based policy extremity scoring
...
Implements U4: extremity_scorer.py uses ai_provider.chat_completion_json_parallel
with a JSON schema enforcing integer 1-5 + Dutch explanation.
Design:
- Batch size 10, max_workers 5 for parallel API calls
- Prompt asks for concrete policy + radicalism score in Dutch
- Stores results in table (motion_id, score, explanation, error)
- Updates with yearly averages
- Default sample=50 for validation; --sample -1 scores all motions
Sample validation (50 motions): scores distributed 1→2, 2→34, 3→7, 4→7,
yearly averages ~2.0-2.5 (mild-to-moderate radicalism).
2 months ago
Sven Geboers
1bc83c4384
feat(right-wing): temporal aggregation of right-wing motion trends
...
Implements U3: temporal_analysis.py computes yearly_summary from the
right_wing_motions table (U2 output).
Metrics per year:
- total_right_wing, pct_of_total, total_motions
- avg_right_support, avg_left_opposition, centrist_support
- avg_right_keyword_matches, extremity_index (U4 placeholder)
- yoy_right_wing_delta, yoy_pct_delta
Key finding: right-wing motions grew from ~4% (2018) to ~12% (2024-2025)
of all motions, with rising centrist support over time.
2 months ago
Sven Geboers
d3dfb0ce2f
feat(right-wing): hybrid motion classifier using keywords + votes
...
Implements U2: classify_motions.py loads keywords from U1 and classifies
motions as right-wing when:
- right_support >= 60% (CANONICAL_RIGHT parties voting 'voor')
- left_opposition >= 40% (CANONICAL_LEFT parties voting 'tegen')
- AND at least 1 right-wing keyword match in title/body_text
Outputs DuckDB table with:
- motion_id, year, title, right_support, left_opposition, centrist_support
- right_keyword_matches, left_keyword_matches, classified flag
Classified 2986 of 28331 motions (10.5%) as right-wing.
2 months ago
Sven Geboers
c6f8540671
feat(right-wing): derive right-wing keywords via differential TF-IDF
...
Implements U1: derive_keywords.py uses party voting patterns to classify
motions as right-wing vs left-wing, then computes differential TF-IDF on
cleaned motion titles to surface policy terms distinctive to right-wing
motions.
Key design choices:
- Vote threshold: 60% of parties in group must vote 'voor'
- Text cleaning strips motion prefixes aggressively (handles multi-word
surnames, plural 'leden', t.v.v. parentheticals)
- Expanded Dutch stopword list filters procedural and generic noise
- Results written to analysis/right_wing/right_wing_keywords.json
Produces ~50 filtered terms including: asielzoekers, defensie, kernenergie,
boeren, vreemdelingenbeleid, stikstof, asielstop, strafrecht.
2 months ago
Sven Geboers