Training Our Own Ecological Language Model

We’ve built a working NLP pipeline that tags organizations, locations, and ecological concepts from multilingual reports. We’ve even used Doccano to correct and improve those tags.

Now it’s time to teach the system to do better — by training our own custom NER model using the annotated data.

This post explains how we take human-labeled examples and turn them into a smarter, more accurate ecological language model.

Why Train a Custom Model?

spaCy’s built-in models are trained on general-purpose data. They work well, but they don’t understand:

• Domain-specific terms like “climate resilience”, “eco-spirituality”, or “integral development”
• New languages or spelling variants
• Cross-domain relationships common in integral ecology (science + ethics + economics)

Training your own model lets you:

• Add new entity types (e.g. ECO_CONCEPT)
• Improve accuracy for your documents
• Adapt the system to your language and context

Tools We Use

We use spaCy’s training framework. It supports:

• Training from scratch or fine-tuning
• Evaluation metrics (precision, recall, F1)
• Easy use of exported Doccano data

From Doccano to Training Data

After annotating in Doccano:

1. Export your project as .jsonl
2. Use a converter to turn that into spaCy format:

spacy convert yourfile.jsonl ./train_data --lang en --ner

This gives you .spacy binary files for training.

Or use a Python script to convert to train.json, dev.json split manually.

⸻

Training the Model

1. Create a config file:

python -m spacy init config ./config.cfg --lang en --pipeline ner

2. Edit config.cfgto match your needs (entity labels, GPU, batch size, etc.)
3. Prepare training assets:

python -m spacy init fill-config config.cfg config_final.cfg

4. Train the model:

python -m spacy train config_final.cfg --output ./output --paths.train ./train.spacy --paths.dev ./dev.spacy

When done, your trained model will be in:

./output/model-best

You can load it like this:

import spacy
nlp = spacy.load("output/model-best")

Measuring Progress

During training, spaCy tracks:

• Precision — how many predicted entities were correct
• Recall — how many true entities it found
• F1 Score — a balance of the two

This helps you compare:

• Pre-trained model (baseline)
• Your fine-tuned model (specialized)

⸻

Try It Yourself

Once you have annotations from Doccano:

make train-model

Or step through each stage with:

python -m spacy init config ./config.cfg --lang en --pipeline ner
python -m spacy train config.cfg --output ./output --paths.train ./train.spacy --paths.dev ./dev.spacy

Test the result:

nlp = spacy.load("output/model-best")
doc = nlp("WWF works on climate resilience in the Amazon rainforest.")
print([(ent.text, ent.label_) for ent in doc.ents])

A Living Model

As more reports are added and more annotations made, we can:

• Retrain the model periodically
• Add multilingual support (e.g. fr, es, zh)
• Extend to more entity types (like SPECIES, DOCUMENT, THEOLOGICAL_TERM)

This is how the Digital Library of Integral Ecology gets smarter over time — powered by human insight and collaborative learning.

⸻

What’s Next?

In the final post, we’ll evaluate the full system and compare how well the pipeline performs before and after training, across languages and document types.