Knowledge Graph from Urban Policy Text · Method Comparison

Six relation-extraction pipelines, compared on the same corpora, as a step toward building a concept knowledge graph from urban-resilience policy documents (OneNYC 2050 and a Shanghai policy corpus).

Question: when the goal is a usable concept graph rather than a benchmark score, which extraction method gives the most coherent structure per unit of effort?

---

Methods

#	Notebook family	Approach	Strength	Weakness
1	Method1_frequency*	Term co-occurrence counts over a fixed 10-node vocabulary	Fast, fully interpretable	Blind to direction and context
2	Method2_TF-IDF*	TF-IDF weighting + fixed keyword nodes	Downweights corpus-wide filler	Still needs a seed vocabulary
3	Method3_LDA_Cooccurance*	LDA topics, then co-occurrence within topics	Finds unseeded themes	Topic count is a judgement call
4	Method4_direct gpt*	Direct LLM prompting for triples	Handles paraphrase and direction	Cost, variance, needs verification
5	Method5_BerTopic*	BERTopic with embedding clusters	Strong topic separation	Cluster labels need human naming
6	Method6_Embedded_model_*	Sentence-embedding similarity (bge / e5 / Qwen / Gemini / Cohere)	Multilingual, model-swappable	Similarity ≠ relation

Every method has a _SH (Shanghai) counterpart, so the two corpora can be compared within the same pipeline.

Supporting notebooks: Bert.ipynb (fine-tuning setup), M5/Method5_BerTopic_fine_tuning*.ipynb, b_SOM_kmeans_process.ipynb (clustering borrowed from the urban-heat study — not part of the KG comparison).

Data — not included

The corpora are not committed (*.csv, *.json, *.npy, *.pkl are ignored). Notebooks read them from an absolute local path such as C:\Users\owner\Desktop\KG\OneNYC_2050_Strategic_Plan_final.txt. To reproduce, supply the policy text and update that path in each notebook.

Environment

Python 3 + Jupyter: nltk, gensim, scikit-learn, bertopic, sentence-transformers, transformers, torch, pandas, matplotlib. No requirements.txt is included; pin versions from the notebook imports.

Notes

• The default branch is cleanup.
• bert-base-*-finetuned-imdb*/ directories are Hugging Face Trainer output folders; only TensorBoard logs are committed and the imdb in the name is a leftover of the training script's default, not the dataset used.
• Files named 0.26.0' and 2.7.0 are accidental pip output, not data.
• There is no ground-truth triple set, so this repository compares methods qualitatively; treat the tables as exploratory.

License

Not declared.