Toxic Comment Classifier Server
Models to detect hateful comments served with flask trained on the well-known Kaggle's dataset.
Dataset
Kaggle's Toxic Comment Classification is a multi-label classification task containing a set of 6 possible labels:
- toxic
- severe_toxic
- obscene
- threat
- insult
- identity_hate
Thus, any sample can have between 0 and 6 labels at a time.
Data files (train.csv, test.csv and test_labels.csv) from Kaggle should be put in data.
Analysis
This part correspond to this script:
python analysis.py
The dataset is highly imbalanced with near 90% of the 159,571 samples not labeled. The distribution of labels is:
- toxic : 15,294 comments
- obscene: 8449 comments
- insult: 7,877 comments
- severe_toxic: 1,595 comments
- identity_hate: 1,405 comments
- threat: 478 comments
The length of comments are skewed distributions:
- characters are on average (394 +/- 590).
- average length of ( 80 +/- 120 ) tokens.
we also noted a lot of noise and language contractions in text since they are comments. We cleaned them in the NLP pipeline.
Models
We have two approaches: the ML pipeline and the transformer-based one. The first contain baseline classification models with TF-IDF features in one vs rest setup. The second consist of current SOTA model which have revolutionized NLP even if they are harder to put into production.
We do not focus on hyperparameter optimization in this project. We will take default know configuration and use our multiple models to make a final choice.
We evaluate our models with the Macro F1 metric which the average of individual F1 scores. We use F1 to optimize both precise detection and finding most toxic comments (recall).
Classical ML NLP pipeline
The pipeline is linear and done with sklearn:
- cleaning text
- processing vocabulary and TFIDF
- classifier in one vs rest setting
All pipelines are saved with pickle as dictionary with trained pipeline as values. This way dict comprehension is useful to infer labels iteratively.
To train all baseline models run:
python baselines.py
Linear SVC
Model based on SVM architecture for classification known in NLP to be SOTA before the deep learning era.
F1 scores for svc:
- toxic: 0.768866
- severe_toxic: 0.404000
- obscene: 0.787289
- threat: 0.314815
- insult: 0.670479
- identity_hate: 0.378378
Macro-F1: 0.553971
Logistic Regression
Model well-known in ML with good interpretability. It is equivalent to one artificial neuron but often too simple for high-level task.
F1 scores for logistic:
- toxic: 0.721715
- severe_toxic: 0.376068
- obscene: 0.733263
- threat: 0.193548
- insult: 0.622429
- identity_hate: 0.238636
Macro-F1: 0.480943
XGBoost
An ensemble classifier using boosting that is considered very effective in many tasks by the ML community.
F1 scores for xgboost:
- toxic: 0.710846
- severe_toxic: 0.301969
- obscene: 0.786611
- threat: 0.350877
- insult: 0.648352
- identity_hate: 0.387255
Macro-F1: 0.530985
Pre-trained Transformer
In the last couple of years, pre-trained transformer models have completely change the NLP field and SOTA. They are very huge models pre-trained on massive corpora. Furthermore, they have established performances often above human levels (e.g.: SQuaD) since they are very effective for high-level task like detecting toxic comments.
To fine-tune from pre-trained model run:
python transformer_finetune.py
To pre-download the fine-tuned model, you can run the example in:
python distilroberta.py
It will take some time.
The final fine-tune model was pushed on my huggingface.co's account (a S3 bucket) as:
jpcorb20/toxic-detector-distilroberta
It is directly downloaded and used by the distilroberta module.
DistilRoBERTa
We chose the DistilRoBERTa model for two reasons: RoBERTa is the most optimized and robust available model and the distillation process makes it smaller by 30-40% while maintaining most performances which is great for production.
F1 scores for distilroberta:
- toxic: 0.72
- severe_toxic: 0.38
- obscene: 0.72
- threat: 0.52
- insult: 0.69
- identity_hate: 0.60
Macro-F1: 0.61
Thus, this model is the most performing one with great improvements on most labels with best improvements on threat and identity_hate.
Install dependencies
This project uses the pip command to install dependencies without issues.
pip install -r requirements.txt
For Windows user, the torch dependency require this specific installation:
pip install torch==1.5.1+cu101 torchvision==0.6.1+cu101 -f https://download.pytorch.org/whl/torch_stable.html
Tests
To run the unit tests, it is this command:
pytest --cov=. --cov-report=term-missing tests\
All tests should be completed before building and releasing the container.
Production-ready Docker version
Build the image for better reproducibility and get a production-ready image (might take some time):
docker build -t toxic-comment-server
Run a container with environment variables from .env file:
docker run -p 8080:8080 --env DEBUG=0 --env [...OTHER ENV...] toxic-comment-server
gunicorn is used for production.
API
Auth is basic one :
username: toxic-filtering-agent
password: nc5wbW8nxgY6Vf48
GET : /toxic_comment?
-
text: your text between single quotes.
-
model: to access transformer model send 'distilroberta' (optional, default SVC model).
http://localhost:8080/toxic_comment?text='YOUR TEXT'&model='model'
Futher works
- Finish all docstring.
- 100% coverage of code with unit tests (mostly transformer code).
- Enhance security with POST request instead of GET in main.py.
- Do hyperparameter optimization to enhance model performances.
- Ensemble models to push results.
- Generate a TinyBert version of transformer for very optimized production-ready model.
- Add a CI for better upgrades and smooth deployments.
- Deploy container on cloud.
References
Since it is a very known task in scientific and gray literature, we found interesting previous work on Toward Data Science: