Hate speech is a crime that has been growing in recent years, especially in online communication. It can harm the individual or a group of people by targeting their conscious or unconscious intrinsic characteristics. Additionally, the psychological burden of manual moderation has necessitated the need for automated hate speech detection methods. In this notebook, we describe our profiling system to the PAN at CLEF 2021 lab “Profiling Hate Speech Spreaders on Twitter”. The aim of the task is to determine whether it is possible to identify hate speech spreaders on Twitter automatically. Our final submitted system uses character -grams as features in combination with an SVM and achieves an overall average accuracy of 69.5% for the English and Spanish datasets. Additionally, we experimented with a Bi-LSTM model and trained it with Sentence-BERT, achieving slightly worse performance results. The experiments show that it is dificult to detect solidly hate speech spreaders on Twitter as hate speech is not only the use of profanity.
The Cambridge Dictionary defines hate speech as abusive or threatening speech or writing
that expresses hate or prejudice towards a person or a particular group1, especially based on
ethnicity, religion, sex, or sexual orientation. Thus said, any characteristics of an individual can
become the target of hate be it gender, nationality, or even educational background. The Internet
and the possibility of communicating anonymously made it additionally an efective vehicle
for spreading hateful and ofensive content at an unprecedented rate [
Currently, social media companies such as Twitter and Facebook use human annotators to
manually detect hateful comments and posts2. Additionally, users are encouraged to report
ofensive and potentially harmful content. Given the high volume of messages posted on
social media websites, these methods are time-consuming, expensive, and depend on human
judgment. The evident harm and volume of the uncontrolled spread of hate speech [
This problem of detecting hate speech is addressed in this year’s author profiling shared
task of PAN at CLEF 2021 lab4 [
We propose two diferent learning experiments. Our final submitted system uses TF-IDF
weighted character -grams as features in combination with an SVM. As recurrent neural
networks (RNN) can preserve sequence information over time, and thereby integrate contextual
information better in classification tasks, we additionally experimented with a bidirectional
LSTM (Bi-LSTM) and trained it with Sentence-BERT (SBERT), a modification of the BERT
network. SBERT uses siamese and triplet network structures to derive semantically meaningful
sentence embeddings [
In the following sections, we describe our approach for the author profiling task at PAN 2021. After a brief review of related work in Section 2, Section 3 details the Twitter data provided by the PAN 2021 organizers. Additionally, we show some key statistics observed in the tweets. Section 4 details the preprocessing steps and features used to train our models. The methodology and classification results are discussed in Section 5. The last Section 6 concludes our work.
Mutanga et al. [
Kovács et al. [
A major challenge for the automatic detection of hate speech on social media is the separation
between hate speech and instances of ofensive language. Davidson et al. [
To train our system, we used the PAN 2021 author profiling corpus 5 proposed by Rangel et al. [
It is important to note that the classes are not predefined by the organizers. We assume that class 0 refers to hate speech spreaders. Nevertheless, since the organisers do not explicitly define classes 0 and 1, we have kept the class names as originally proposed. As can be seen in Table 1, the Twitter-specific tokens such as hashtags, URLs, and user mentions were replaced by the providers with the following placeholders: #HASHTAG#, #URL# and #USER#. The examples provided in Table 1 were chosen carefully to show that insults and profanities are used by hate speech spreaders as well as by other users. Additionally, Twitter-specific text significantly contributes to the dificulty of automatic hate speech detection, as the posts contain plenty of poorly written text and paralinguistic signals such as emoticons, @-mentions, and hashtags. Prior to feature engineering (described in Section 4), we analysed the distribution of diferent tokens. Table 2 shows some key insights for both languages.
We observed the distribution of specific tokens to see whether we could use these for the features engineering process. Unfortunately, we could not spot any significant diferences between the classes. Therefore, to train our model, we did not use features mentioned in Table 2.
6The selected tweets are used for demonstration and research purposes only and do not reflect the opinion of the authors. “#USER# #USER# Trump, that mother- “Kappa They gon be beating my fodder fucker is guilty of cowardice while being ninjas asses weak ass punks and i wont Commander-in-Chief #HASHTAG#.” even be laughing on the outside :-)” “RT #USER#: If a nigga taking care of me “Shut your fucking mouth i have no ill will i’m fasho taking care of him. it’s really that towards Kaep but he’s not even close lmao simple.” #URL#” “RT #USER#: Celebrities are so useless and “#USER# All the people shit talkin this are corny B*tch what the fuck does this even trippin, i’d pipe tf out if an old lady if she mean?” was payin for all my shit” “#USER# #USER# Mordes la mano de “#USER# Pos pa tu tierra sucnormal hiquien de da. De comer eres un cancer para jadeputa” nuestro pais #URL#” “Los varones opinando sobre el feminismo “#USER# Ostia tio que palo metió el jo#HASHTAG#. Nos sorprende? No nos sor- dido” prende” “Que pinches perras ganas de estar co- “RT #USER#: Qué horror. Condenado a 15 giendo con Ale” años de prisión por dejar embarazada a su hija tras un año de violaciones #URL#”
The preprocessing pipeline to clean and structure the data was performed for both languages (EN and ES) and models as follows: • The text from the original XML document was extracted and all 200 tweets per author were concatenated to one text. • The white-space between the tokens has been reduced to a single space. • The placeholders #USER#, #URL#, #HASHTAG#, and RT were removed. • HTML characters were converted to Unicode characters (e.g.: “>”, “<”, “&” to “&”, “>”, “<”). • Emojis were converted to text format by using Python’s emoji library. • The text was lowercased. • Irrelevant signs, e.g. “+,*,/” were deleted. • Alphanumeric tokens were separated (e.g. “Berlin2018” to “Berlin 2018”). • Sequences of repeated characters with a length greater than three were normalized to a maximum of two letters (e.g. “LOOOOOOOOL” to “LOOL”). • Words with less than three characters were ignored (except for the Bi-LSTM model for the English language). • Stopwords were deleted (except for the Bi-LSTM model for the English language). • As the last step, we lemmatized the English tweets for the TF-IDF character -gram SVM model using WordNetLemmatizer.
Besides the diferent preprocessing steps, we also experimented with diferent vectorization
techniques and hyperparameter tuning by employing scikit-learn’s grid search function. The
hyperparameters were tuned separately for English and Spanish. We experimented with
emotional signals and lists of hate words as handcrafted features as well as with automatically
learned features. The best results were achieved by using Scikit-learn’s term frequency-inverse
document frequency (TF-IDF) vectorizer and Sentence-BERT (SBERT), a BERT model
modification that uses siamese and triplet network structures to generate semantically meaningful
sentence embeddings [
For the SVM model, we employed TF-IDF weighted character -grams. In English, the best results were achieved using a maximum of 1,250 features (min_df=5) and character -grams with range [3;7]. For Spanish, we used top 2,350 features (min_df=5) and character -grams with range [2;7].
We defined this year’s PAN author profiling task “Hate Speech Spreaders on Twitter” as a binary classification problem. For each language (EN and ES) we trained two diferent models. We tested diferent features and vectorization techniques with a Support Vector Machine (SVM). Additionally, we experimented with bidirectional LSTM (Bi-LSTM) models as recurrent neural networks (RNN) have shown that they can preserve sequence information over time and thereby integrate contextual information in classification tasks.
For the final SVM model, we trained a linear kernel and set the penalty parameter C=10 for the English data. For the Spanish corpus, we trained the SVM with the radial basis function
kernel (RBF) and C=5. The performance was ranked by accuracy. Table 4 shows the scores
for our final system performed on the oficial PAN 2021 test set on the TIRA platform [
In this paper, we described our participation in the PAN 2021 author profiling task. The goal was to develop a system that can detect Twitter users who spread hate speech on a regular basis. First, we observed the distribution of specific tokens in the tweets like the usage of emojis or user mentions to see whether we could use these for the feature engineering process. Unfortunately, we could not spot any significant diferences between the two classes. Furthermore, we experimented with emotional signals and dictionaries listing hate words as handcrafted features in addition to automatically learned features. In relation to this, we could not detect any diference in emotions between the two classes and have shown that insults and profanities are not a discriminative features of hate speech spreaders and other users.
Our final submitted system uses an SVM with TF-IDF weighted character -grams. This model performed best for the English language. To detect hate speech spreaders in Spanish tweets, a bidirectional LSTM (Bi-LSTM) trained with Sentence-BERT achieved better classification results. The SVM model achieved an average accuracy of 69.5% for both languages which is slightly better than the Bi-LSTM model (69%).
The experiments show that it is challenging to detect hate speech spreaders on Twitter. It is challenging in diferent ways. First, we have shown that insults and profanities are not only used by hate speech spreaders, but also by users who do not ofend other individuals or groups. Additionally, Twitter posts contain plenty of poorly written text (spelling mistakes, abbreviations, etc.) and paralinguistic signals such as emoticons, @-mentions, and hashtags. In the future, we want to make the classification results interpretable to analyse how hate words and the context in which they are expressed contribute to the classification.
This work was supported by the German Federal Ministry of Education and Research and the Hessen State Ministry for Higher Education, Research and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE and under grant agreement "Lernlabor Cybersicherheit" (LLCS) for cyber security research and training.