17,123
Monuments
Total scraped
10,211
AAT-Annotated
209 concepts
563,738
Mentions
Term matches
8,207
Persons
Historic actors
5,713
Interventions
Conservation records
89,250
Images
Photographic docs
Construction Activity by Century
Peak: 18th century (3,020 mentions). Shift from religious to residential dominance in 19th–20th c.
Description Length → Concepts
r = 0.791 correlation. Texts >5k chars yield 63.4 concepts on average vs 3.3 for <200 chars.
Conservation Interventions
Conservation activity peaked in the 1980s–1990s under EU structural funding.
Key Findings
Zipf exponent = −1.78 (steeper than natural language → highly concentrated vocabulary)
8 distinct monument clusters by AAT concept profile
Goa has highest concept richness (mean 47.9 concepts/monument)
Sacred/secular vocabulary split: 98× ratio for baptismal fonts
Overseas heritage uses 5× more galleries, 4.6× more cloisters
Pointed arches: 22% → 5% (decline from 14th to 20th century)
Aluminum emergence: 0% → 5% (industrialization signal)
Chancel–nave–bell tower PMI cluster = latent church grammar
Strongest Positive Associations (PMI)
The top cluster reveals a latent church grammar: chancels, naves, bell towers, and axial plans form a tight co-occurrence group encoding the typological blueprint of Portuguese religious architecture.
Negative Associations
"Stories" anti-correlates with all religious elements. Multi-story buildings (residential) and single-height churches (naves, bell towers) occupy different architectural domains. This validates the pipeline: co-occurrence structure reflects real architectural logic, not noise.
CRM implication: PMI scores can weight
P46 is composed of relationships. A chancel–nave pair (PMI 1.86) should be modeled as a stronger compositional relationship than an accidental co-occurrence.
Emerging Concepts (% of datable monuments)
Balconies surge in the 18th–19th century (Pombaline & liberal era). Concrete appears only in 20th century. These curves encode industrialization and stylistic evolution.
Declining Concepts (% of datable monuments)
Pointed arches peak at 22% in the 14th century (Gothic) then halve. Ramparts vanish after gunpowder warfare. The curves trace technological and stylistic obsolescence.
A Dateable Vocabulary
AAT concepts are not just labels — they are chronological markers. If a monument has pointed arches, gargoyles, and buttresses, the concept profile alone dates it to the 13th–15th century with high confidence. If it has balconies, window frames, and stucco, it's 18th–19th century. This is a form of computational typological dating — a vocabulary that encodes time.
Sacred-Exclusive Concepts
Secular-Exclusive Concepts
Two distinct architectural vocabularies emerge from the same NLP pipeline. Baptismal fonts appear 98× more often in sacred buildings — effectively a binary classifier. This split can constrain CIDOC-CRM typing: if a monument has naves + chancels + altars, it's
E22 → P2 has type → "religious architecture" without human intervention.Most Similar Districts (Jaccard)
Northern Portugal forms a tight cluster (Guarda–Porto–Braga–Viana). The granite belt — these districts share nearly identical concept vocabularies. Lisboa–Évora form a southern cluster.
Most Dissimilar Districts
Moroccan heritage (Marrakech) shares only ~10% of concepts with any Portuguese district — a completely distinct architectural vocabulary documented by the same inventory system.
Regional Concept Richness
More Prevalent Overseas
Galleries (5×), cloisters (4.6×), and parapets (3.6×) — colonial architecture favored covered circulation spaces and defensive elements adapted to tropical climates.
Less Prevalent Overseas
Axial plans (0.08× ratio) — the longitudinal church layout barely transfers. Eaves (0.41×) and plinths (0.43×) also drop, suggesting adaptation to different building traditions and climates.
Colonial Adaptation Factor
The overseas comparison reveals a measurable "colonial adaptation factor" — the same inventory system documenting different architectural responses to climate, materials, and cultural context. Quicklime (3.8×) and brick (2.4×) replace granite in tropical settings. This is a contribution to colonial architecture studies: quantified vocabulary transfer across the Portuguese empire, derived entirely from NLP on inventory text.
Architecturally Anomalous Monuments
Monuments scored by how unusual their concept profile is for their tipologia. High scores = concepts appearing in <5% of buildings of the same type. These are buildings that defy their category.
The "Edifício na Rua de São Julião" in Lisbon is classified as residential multifamily but contains sacristies, choir lofts, and chapels — likely a former convent repurposed as housing. Anomaly detection can flag misclassifications in the inventory and identify buildings with complex reuse histories.
Data Science Techniques Applied
10 techniques applied to 10,211 monuments × 209 AAT concepts × 563,738 mentions. Each reveals patterns invisible in raw text.
Technique
Purpose
Key Finding
Application