By Jitendra Singh Malik, Guansong Pang, Anton van den Hengel
This research takes advantage of different embedding including Term Frequency - Inverse Document Frequency (TF-IDF), Glove (Global Vector) and transformers based embedding (eg. BERT, ELECTRA, AlBERT etc.) along with combination of CNN, Bi-LSTM and MLP to give a detailed comparison in order to find the best and efficient method to perform the hate speech detection. The research takes into account three data sets so any biasness in respect to any particular data set can be dealt with. Moreover, the study also compare the cross domain analysis by performing inter dataset experiments .In order to give out the best method we will look into the F1 score of different methods and the one with the highest score will be considered as outperforming its competition.
Folowing dataset were used by us to perform the experiments. These dataset were also modified by us to use the same models to perform the cross dimaib experiments.
Dataset 1 : @inproceedings{hateoffensive,
title = {Automated Hate Speech Detection and the Problem of Offensive Language},
author = {Davidson, Thomas and Warmsley, Dana and Macy, Michael and Weber, Ingmar},
booktitle = {Proceedings of the 11th International AAAI Conference on Web and Social Media},
series = {ICWSM '17},
year = {2017},
location = {Montreal, Canada},
pages = {512-515}
}Dataset 2 : @inproceedings{founta2018large,
title={Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior},
author={Founta, Antigoni-Maria and Djouvas, Constantinos and Chatzakou, Despoina and Leontiadis, Ilias and Blackburn, Jeremy and Stringhini, Gianluca and Vakali, Athena and Sirivianos, Michael and Kourtellis, Nicolas},
booktitle={11th International Conference on Web and Social Media, ICWSM 2018},
year={2018},
organization={AAAI Press}
}
Dataset 3: https://datahack.analyticsvidhya.com/contest/practice-problem-twitter-sentiment-analysis/
This code is written in Python and requires the packages listed in requirement.txt. Install with pip install in a virtualenv.
Each of the file :
Glove - BiLSTM+CNN+MLP- models for glove embeddingsTF-IDF + MLP + SVM + XGB- all models with TF-IDF embeddingsTransfomer- CNN & MLP- transfomer models
can be run directly by providing a dataset file path inside these files.
The given Transfomer- CNN & MLP file contains the implementation of the bert_en_uncased_L-2_H-128_A-2 model. In order to execute a different model the files needs to be prvided with the specific transfomer model name which are available upon the official transfomer website :
https://tfhub.dev/google/collections/transformer_encoders_text/1
Each of these models can be simply run by providing the code for the model available at above website. For more ease the following files give a dropdown option to select the models to run, when executes on google colaboratory without the need to past the code for each specific model.