WaterKG — wastewater-treatment literature knowledge graph (v1.0)

WaterKG links 118,131 canonical entities of six types to 672,380 research papers on water and wastewater treatment. It has two layers:

• Entity layer — which entities each paper mentions, with every mention's surface form and character offsets, and a two-level (L3/L2) taxonomy of the entities.
• Relation layer — removes, has and removal_rate relations extracted by WaterBERT-RE, with per-paper evidence and graph-level aggregates.

Entities were extracted from titles and abstracts with WaterBERT-NER, normalised and merged into canonical entities, and classified into 57 L3 / 124 L2 categories.

Download

The data files (Parquet, CSV.gz, Neo4j import files and paper embeddings; 1.8 GB) are in the Hugging Face dataset Mudi12137/WaterKG:

from huggingface_hub import snapshot_download

snapshot_download("Mudi12137/WaterKG", repo_type="dataset", local_dir="WaterKG/graph")

The GitHub repository Mudi12138/WaterBERT holds this documentation, the hybrid retrieval package, the Neo4j scripts and a 20-row sample of every table (graph/sample/).

Contents

graph/
├── parquet/          one file per table (typed, recommended)
├── csv/              the same tables as gzip-compressed CSV
├── neo4j/            header files, import_neo4j.sh, constraints.cypher
├── vectors/          BGE-large paper embeddings for the retrieval system
├── summary.json      counts and build statistics
└── checksums.sha256

Table	Rows	One row is
entities	118,131	a canonical entity
entity_aliases	146,905	a surface form of an entity from the curated entity list
categories	181	an L3 (57) or L2 (124) taxonomy category
papers	672,380	a paper
edges_entity_category	94,622	entity → L2 category
edges_category_parent	124	L2 → L3 category
edges_entity_paper	4,079,774	entity mentioned in a paper
mentions	8,800,205	one entity mention in a paper
relation_evidence	922,511	one extracted relation in one paper
relations	204,374	a relation between two entities, aggregated over papers

Identifiers

Prefix	Example	Object
E	E00001	canonical entity
P	P0000001	paper (anonymous release id)
<TYPE>-L3-nn / <TYPE>-L2-nnn	POL-L2-003	taxonomy category; TYPE is POL, WTP, RCT, TRP, MIC or DOS
M	M00000001	mention
V	V00000001	relation evidence
R	R0000001	aggregated relation

Papers carry DOI, publication year and journal only. 614,235 papers (91.4%) have a DOI; 378 rows share a DOI with another row because the source database holds two records for them (doi_shared = true).

Tables

entities

Column	Description
entity_id	canonical entity id
name	canonical name (lower-case normalised form)
entity_type	Pollutant, Wastewater_Treatment_Process, Reactor, Treatment_Parameter, Microorganism, Dosed_Material
l3_category_id, l2_category_id	taxonomy categories (empty when unclassified)
l3_en, l2_en / l3_zh, l2_zh	category names in English / Chinese
classification_status	classified (94,622) or classification_unknown (23,509)
domain_flag	in_domain, out_of_domain (category "Out of domain") or unclassified
n_surface_forms	number of distinct surface forms merged into the entity
mention_count	mentions across the corpus
paper_count	papers mentioning the entity

entity_aliases

entity_id, alias — the curated surface forms of each entity (canonical name included). Every other spelling observed in the corpus is in mentions.surface_text.

categories

category_id, level (L3/L2), entity_type, name_en, name_zh, parent_id (the L3 of an L2).

papers

paper_id, doi (lower-case), year, journal, doi_shared.

edges_entity_paper

entity_id, paper_id, mentions (number of mentions of the entity in the paper).

mentions

Column	Description
mention_id	mention id
paper_id, entity_id, entity_type	where and what; entity_id is empty for the 1,726,204 mentions (19.6%) whose surface form was not merged into a canonical entity
surface_text	the span as written in the paper
char_start, char_end	character offsets into the abstract text of the source record
ner_confidence	WaterBERT-NER confidence

Offsets refer to the abstract as exported from the bibliographic database; abstracts are not redistributed, and abstracts obtained elsewhere may differ slightly.

relation_evidence

Column	Description
evidence_id, paper_id	evidence id and paper
relation	removes, has or removal_rate
re_confidence	WaterBERT-RE probability of the predicted label
head_*, tail_*	for each endpoint: entity_id, type, text, start, end, ner_confidence, link
value_percent	for removal_rate: the value as a number when it is a single percentage in 0–100

For removal_rate the tail is a value (tail_type = Value, tail_entity_id empty). A removal efficiency can be attached to a removes relation by joining on paper_id and the pollutant's entity_id.

*_link records how the endpoint was linked to the canonical entity: span_exact, span_overlap, alias_exact, alias_loose or value.

relations

Column	Description
relation_id	aggregated relation id
head_entity_id, relation_type, tail_entity_id	REMOVES (process/reactor → pollutant) or HAS (process/reactor → parameter)
n_evidence, n_papers	supporting evidence rows and distinct papers
mean_re_confidence, max_re_confidence	over the evidence rows

Loading

import pandas as pd

ents = pd.read_parquet("WaterKG/graph/parquet/entities.parquet")
rel = pd.read_parquet("WaterKG/graph/parquet/relations.parquet")
name = ents.set_index("entity_id")["name"]

# processes that remove a pollutant, ranked by number of papers
tc = ents.loc[(ents.name == "tetracycline") & (ents.entity_type == "Pollutant"), "entity_id"].iloc[0]
rm = rel[(rel.relation_type == "REMOVES") & (rel.tail_entity_id == tc)]
print(rm.assign(process=rm.head_entity_id.map(name)).nlargest(10, "n_papers")[["process", "n_papers"]])

Neo4j

bash WaterKG/graph/neo4j/import_neo4j.sh WaterKG/graph/csv waterkg     # Neo4j 5, offline import into a new database
cypher-shell -d waterkg -f WaterKG/graph/neo4j/constraints.cypher

This loads (:Entity), (:Paper) and (:Category) nodes and the BELONGS_TO, SUBCLASS_OF, MENTIONED_IN, REMOVES and HAS relationships. mentions and relation_evidence stay as tables.

MATCH (p:Entity)-[r:REMOVES]->(t:Entity {name: 'tetracycline'})
RETURN p.name, r.n_papers ORDER BY r.n_papers DESC LIMIT 10;

License

The graph data are released under CC BY 4.0. Bibliographic identifiers (DOI, year, journal) are factual metadata; no titles, abstracts or authors are included.