covid_19_hate_speech

Data Gathering

The tweet dataset were acquired from a public dataset of tweet ids with various coronavirus related keywords from [COVID-19: The First Public Coronavirus Twitter Dataset](https://github.com/echen102/COVID-19-TweetIDs.

Hydrating Tweets

Install all requirements from requirements/hydration.txt

$ pip install requirements/hydration.txt

Apply for a Twitter developers account. Configure twarc with the developer keys from Twitter.

$ twarc configure

Hydration can be run using:

$ python3 hydrate_tweets.py

Default parameters for ZIP_TWEETS, raw tweet source dir, hydrated tweets target dir, month dir and language selections can be changed inside hydrate_tweets.py. By default they are:

ZIP_TWEETS             = False
raw_tweet_src          = 'raw_tweet_ids/'
hydrated_tweets_target = 'hydrated_tweets/'
month_dirs             = ['2020-01', '2020-02', '2020-03']
lang_set               = set(["en", "null", None])

Hate speech lexicon and labelled hate speech twitter data examples were acquired from Racial Bias in Hate Speech and Abusive Language Detection Datasets.

Data Preprocessing

To recreate results from Davidson et. al. and Zimmerman et. al., the preprocessing functions are detailed in util/preprocess_neural_network_ensemble.py and util/preprocess_log_regr_and_svm.py

The preprocessing steps in detail can be found at (davidson et al)[https://github.com/t-davidson/hate-speech-and-offensive-language] and (zimmerman et al)[https://github.com/stevenzim/lrec-2018]

The pre_trained_models folder must contain the Twitter Word2vec embeddings binary from (Federic Godin's GitHub repository)[https://github.com/FredericGodin/TwitterEmbeddings]

Hate Speech Classification Methodology

The hate speech Classification is achieved through an ensemble learning technique where the results from an SVM classifier, Logistic Regression classifier, and Neural Network ensemble are aggregated using a hard voting mechanism.

Performance on the labeled hate speech dataset from Davidson et. al.

Results

Total number of hate tweets, all tweets, and the ration of hateful tweets.

The normalized number of hate tweets, all tweets, and the ration of hateful tweets.

Discussion

The full report and discussion can be found in Research_COVID_19_and_Racism.pdf

References

• Chen E., Lerman K. and Ferrara E. (Mar 16 2020). COVID-19: The First Public Coronavirus Twitter Dataset. arXiv:2003.07372. https://arxiv.org/abs/2003.07372

• Zeerak Waseem and Dirk Hovy. 2016. Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter. In SRW@HLT-NAACL.

• Davidson Thomas , D. Warmsley , M. Macy and I. Weber. 2017. Auto- mated Hate Speech Detection and the Problem of Offensive Language.

• In AAAI Publications, Eleventh International AAAI Conference on Web and Social Media. https://arxiv.org/abs/1703.04009 [17] Steven Zimmerman, Udo Kruschwitz, and Chris Fox. 2018. Improving Hate Speech Detection with Deep Learning Ensembles. In LREC.