WaterBERT-NER

Named-entity recognition for wastewater- and water-treatment literature, fine-tuned from WaterBERT. It tags six entity types with a BIO scheme (13 labels):

Entity type	Covers (examples)
Pollutant	contaminants and water-quality targets — ammonium, sulfamethoxazole, COD, microplastics
Wastewater_Treatment_Process	treatment processes and unit operations — adsorption, Fenton process, anaerobic digestion
Reactor	reactors, units and devices — sequencing batch reactor, membrane bioreactor, constructed wetland
Treatment_Parameter	operating, reaction and material parameters — pH, hydraulic retention time, specific surface area
Microorganism	microorganisms and microbial groups — Nitrosomonas, anammox bacteria, microalgae
Dosed_Material	added materials, catalysts, adsorbents and reagents — biochar, persulfate, PAC

This is the model used to build the WaterKG entity graph.

Usage

from transformers import pipeline

ner = pipeline("token-classification", model="Mudi12137/WaterBERT-NER", aggregation_strategy="simple")
ner("A sequencing batch reactor inoculated with Nitrosomonas removed 95% of ammonium and "
    "sulfamethoxazole at pH 7.5 after dosing biochar.")
# Reactor: sequencing batch reactor | Microorganism: nitrosomonas | Pollutant: ammonium,
# sulfamethoxazole | Treatment_Parameter: ph | Dosed_Material: biochar

Inputs longer than 512 tokens are truncated; split long documents into sentences or passages. The tokenizer lower-cases input, so returned word strings are lower-case — use the start/end character offsets to recover the original spelling.

Labels

O, then B-/I- for Pollutant, Wastewater_Treatment_Process, Reactor, Treatment_Parameter, Microorganism, Dosed_Material (see label_map.json).

Training

Fine-tuned from WaterBERT on manually annotated water-treatment abstracts; the annotation guideline and data are described in the accompanying paper.

Limitations

• Numeric values and removal efficiencies are not entity types of this model. The relation model WaterBERT-RE expects value spans from a separate value detector.
• Boundaries of long compound names (e.g. "rotating disc electrocoagulation system") and the type of cross-category words (e.g. adsorption as a process vs. a parameter) are the most common sources of error.

License

Apache-2.0.