train-ner-model-with-spacy

Custom training NER model with spacy library and annotaded dataset in JSON

The goal of this project is to create model that can annotate custom entities in text like various cryptocurrency names and prices. It can be trained to recognize also other abstractions like people names, organizations and many others.

To train and use model prepare dataset like one attached in data.json file. You can use https://github.com/doccano/doccano browser tool for this purpose.

Part I: Label data

Dataset was prepared based on https://www.kaggle.com/datasets/kaushiksuresh147/bitcoin-tweets To prepare tagged dataset run steps below:

1. Download and install browser program for labeling data "doccano":

go to link below and run described steps: https://github.com/doccano/doccano

# install doccano
pip install doccano
# Initialize database.
doccano init
# Create a super user.
doccano createuser --username admin --password pass
# Start a web server.
doccano webserver --port 8000
# Start the task queue to handle file upload/download (separate console session).
doccano task

Open doccano in browser using http://127.0.0.1:8000/.

2. Create new project:

Mark "Sequence Labeling" Add some project name and description Check options: "Allow single label" and "Randomize document order"

3. Import dataset:

From left menu choose option "Dataset" and Actions/Import Dataset. In "File format" select "TextLine" Upload file bitcoin_tweets_text_lines.txt Click "Import" and wait until file is imported

4. Add labels:

From left menu choose option "Label". Add labels as below, by clicking Actions/Create Label:

• Cryptocurrency name
• Cryptocurrency rate
• Organisation
• URL
• Emoticon
• Date

5. Annotate data:

On left sidebar click "Start Annotation" and tag all texts by marking each meaningful phrase that belongs to one of created categories.

6. Export training set:

From left menu choose option "Dataset" Actions/Export Dataset. In "File format" select "JSONL" Unzip created file and copy .jsonl file into project dir.

Part II: Train model

To use training set from previous part (Label data) you need to change jsonl file into json. You can achieve it by adding comma in all lines except last one and wrap everything in square braces, see samples.json as example.
See bitcoin_tweets_annotated.jsonl

To train model open doccano-spacy.ipynb in jupyter lab and run all cells. Replace samples.json with bitcoin_tweets_text_lines.json to train model on prepared dataset. Replace also labels with ones used for annotation.

To run isolated environment with packages install conda:
https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html

To setup environment run commands:

conda create -n ner_model_training_spacy python=3.10
conda activate ner_model_training_spacy
pip install -r requirements.txt

Start jupyter notebook:
jupyter notebook

Now you can code with traing NER model in notebook doccano-spacy.ipynb

To deactivate conda environment run:
conda deactivate

Evaluation

During training evaluation metrics like loss and F1 score are captured to mlflow