futurice/spice-hate_speech_detection

Name: spice-hate_speech_detection

Owner: Futurice

Description: A SPICE-program funded project where the goal is to detect hate speech in social media.

Created: 2017-03-14 08:07:30.0

Updated: 2018-05-21 23:35:04.0

Pushed: 2017-10-29 20:00:47.0

Homepage: null

Size: 33

Language: Python

GitHub Committers

User	Most Recent Commit	# Commits

Other Committers

User	Email	Most Recent Commit	# Commits

README

Automatic hate speech detection

Setup

1. Install requirements
2. python3
3. python packages: pandas, sklearn, fasttext, sqlalchemy, …
4. Configure collector
5. Edit hiit_collector.py.example and save it as hiit_collector.py
6. Configure PostgreSQL
7. Edit postgre_keys.py.example and save it as postgre_keys.py
8. Get the data
9. FastText model for Finnish trained by Facebook using Finnish Wikipedia: Facebook's trained models

Usage:

Collect new data

usage:

`collector.py [-h] [–user USER] [–password PASSWORD]

                [--hostname HOSTNAME] [--outdir OUTDIR]
                [--startdate STARTDATE] [--enddate ENDDATE]

optional arguments: -h, –help show this help message and exit –user USER Username –password PASSWORD Password –hostname HOSTNAME Hostname –outdir OUTDIR Directory to store data –startdate STARTDATE

                    Startdate as YYYY-MM-DD

–enddate ENDDATE Enddate as YYYY-MM-DD`

Example:

./collector.py --startdate 2017-03-01 --enddate 2017-03-15

Train predictor

Example:

./predict.py --inputdir data/incoming --outdir data/output/ --featurename bow --featurefile data/models/feature_extractor_bow.pkl --predictor data/models/fasttext_svm.pkl

Predict hate speech

Example:

./predict.py --inputdir data/incoming --outdir data/output/ --featurename bow --featurefile data/models/feature_extractor_bow.pkl --predictor data/models/bow_svm.pkl

Sync data

Example:

./sync.py --inputdir data/output/

TODO

1. CNN on Embedding Matrix (c.f Willi)
2. Stemmings, stop words for BoW
3. Study SVM factors (with BoW)
4. Mezadona ? To Models
5. Plot TSNE manifolds for wikipedia model and twitter model
6. Highlight hatewords

DONE:

1. Try Naive Bayes-classifier with BoW
2. Naive Bayes (Gaussian) did perform comparable to RF, but worse than SVM
3. With FastText it performed poorly