Release 1.0.0 contains the code for the article "Detecting Toxic Spans with spaCy". Release 2.0.0 contains the code for the article "Custom evaluation of spans in spaCy".
An expression is toxic if it uses rude, disrespectful, or unreasonable language that is likely to make someone leave a discussion. Toxic language can be short like "idiot" or longer like "your 'posts' are as usual ignorant of reality".
We will use the SpanCategorizer from spaCy to detect toxic spans.
For illustration we will use a well-researched dataset.
Code uses python 3.8.
Requirement is only spaCy >= 3.1.0
If not installed yet, (create your favourite local environment for python) and do
pip install -r requirements.txt
Also, it uses a small language model for English:
python -m spacy download "en_core_web_sm"
The code is run through a spaCy "project" defined in project.yml.
Get the data:
python -m spacy project assets
Convert data, train the model, and evaluate the best model:
python -m spacy project run all
For separate steps:
python -m spacy project run corpus python -m spacy project run train python -m spacy project run evaluate
To use the resulting best model:
python run.py "Enter your toxic sentence here, my apologies for the example, you idiot"
(Don't put interpunction at the end!)
Enjoy!