Prompt-based knowledge acquisition for living sacred heritage

Taxonomy, extraction prompt, derived outputs, and analysis code accompanying:

Pant, P., Newpane, S., Sitaula, H., Sakhakarmy, M. Prompt-based knowledge acquisition for living sacred heritage: a domain-adaptive entity taxonomy and semantic network from visitor reviews. Submitted to Data & Knowledge Engineering.

What this is

A domain-adaptive knowledge-acquisition pipeline for a domain with no annotated corpus and no reference ontology. A 12-type entity taxonomy is specified declaratively in an extraction prompt — no labelled training data, no fine-tuning — and applied with GPT-4o to visitor reviews of Pashupatinath Temple, Kathmandu, a living UNESCO sacred heritage site. The output is a typed, queryable semantic network of visitor experience.

Eight of the twelve entity types are new to sacred heritage and account for 64.8% of extracted entities. The taxonomy introduces a dual_valence class for experiences described with simultaneous grief and transcendence at cremation-active sites.

Contents

taxonomy/
  taxonomy_specification.md   12 entity types, constraints, relation schema, aspect list
  extraction_prompt.txt       verbatim prompt (Appendix A of the paper)
data/
  entity_mentions.csv         11,286 extracted entities
  relations.csv               5,590 relation records -> 5,110 nodes / 5,308 typed edges
  aspect_sentiments.csv       6,232 aspect-level sentiment records
  README.md                   column dictionary and provenance notes
notebooks/
  nlp_vs_llm_comparison.ipynb traditional NLP baseline vs. LLM-SC comparison
src/
  nlp_baseline.py             traditional NLP baseline: NLTK ne_chunk NER, VADER, co-occurrence graph (§3.2)
  network_construction.py     graph construction, edge weighting (α=0.7, β=0.3), degree centrality, Louvain communities (§3.4)
  extraction_pipeline.py      OpenAI Batch API extraction — documented skeleton (§3.3; see note below)

src/nlp_baseline.py and src/network_construction.py are the analysis code used for the paper. src/extraction_pipeline.py documents the Section 3.3 extraction procedure and parameters but is a reconstructed skeleton, not a runnable reproduction: the original OpenAI Batch API driver was run outside version control and is not available. The authoritative extraction prompt is taxonomy/extraction_prompt.txt. The recovered network code computes degree centrality and weighted degree; betweenness centrality is not included because it was not part of the recovered analysis modules.

Data provenance and what is deliberately excluded

Raw review text is not redistributed, in line with the source platform's terms of use. This repository contains only derived extraction outputs: entity mentions, typed relations, and aspect-level sentiment records, each with a short supporting span (typically under 20 words) rather than full review text.

Review identifiers are pseudonymised. Platform-assigned numeric review IDs were used during processing for de-duplication only. They have been replaced here with stable surrogate keys (R00001 …) that are consistent across all three files, so within-review joins still work, but no record can be traced back to a source review or reviewer profile.

No usernames, profile information, or other personal identifiers were ever collected. No individual reviewer is identifiable from any file in this repository.

Corpus summary

Full corpus	3,939 TripAdvisor reviews, 14 Oct 2008 – 16 Apr 2026, English
Analysed sample	2,500 randomly drawn reviews
Valid entity extraction	2,485 reviews
Valid aspect-sentiment extraction	2,487 reviews
Contributing relation records	2,433 reviews
Review length	mean 68.3 words (SD 60.0), median 49, max 612
Ratings	5★ 58.4% · 4★ 27.8% · 3★ 9.2% · 2★ 2.8% · 1★ 1.8%
Trip types	Friends 26.0% · Solo 21.0% · Couples 20.7% · Family 16.0% · Business 5.6% · Unspecified 10.7%

Reproducing

pip install -r requirements.txt
jupyter notebook notebooks/nlp_vs_llm_comparison.ipynb

The comparison notebook runs the traditional NLP baseline (NLTK ne_chunk + VADER) live and compares it against the pre-computed LLM outputs in data/. Point DATA_DIR at data/.

Note that the notebook's NLP baseline requires raw review text, which is not redistributed here. To re-run that half of the comparison you will need to collect reviews yourself from the source listing.

Extraction used gpt-4o-2026-05-28 via the OpenAI Batch API, temperature 0.1, max_tokens 1200, in chunks of 45 requests with a 24-hour completion window.

Citation

See CITATION.cff, or cite the paper above.

Licence

Code is released under the MIT Licence (LICENSE). The derived data files in data/ and the taxonomy specification are released under CC BY 4.0 — reuse freely with attribution to the paper.