A comparative study of different features for efficient automatic hate speech detection - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year :

A comparative study of different features for efficient automatic hate speech detection


Commonly, Hate Speech (HS) is defined as any communication that disparages a person or agroup on the basis of some characteristic (race, colour, ethnicity, gender, sexual orientation, na-tionality, etc. (Nockeby, 2000)). Due to the massive activities of user-generator on social networks(around 500 million tweets per day) Hate Speech is continuously increasing on the web.Recent initiatives, such as SemEval2019 shared task 5 Hateval2019 (Basile et al., 2019) contri-bute to the development of automatic hate speech detection systems (HSD) by making availableannotated hateful corpus. We focus our research on automatic classification of hateful tweets,which are the first sub-task of Hateval2019. The best Hateval2019 HSD system was FERMI (In-durthi et al., 2019) with 65.1 % macro-F1 score on the test corpus. This system used sentenceembeddings, Universal Sentence Encoder (USE) (Cer et al., 2018) as input of a Support VectorMachine classifier.In this article, we study the impact of different features on an HSD system. We use deep neu-ral network (DNN) based classifier with USE. We investigate the word level features, such aslexicon of hateful words (HFW), Part of Speech (POS), uppercase letters (UP), punctuationmarks (PUNCT), the ratio of the number of times a word appears in hateful tweets comparedto the total number of times that word appears (RatioHW) ; and the emojis (EMO). We think thatthese features are relevant because they carry feelings. For instance, cases (UP) and punctuations(PUNCT) can carry the intonation of the tweets and can be used to express a hateful content. ForHFW features, we tag each word of tweets as hateful or not using the Hatebase lexicon (Hate-base.org) and we associate a binary value to each word. For POS features, we use twpipe (Liu etal., 2018) for tagging the words and this information is coded as an one-hot vector. For emojis,we generate an embedding vector using emoji2vec tools (Eisner et al., 2016). The input of ourneural network consists of the USE vector and our additional features. We used convolutionalneural networks (CNN) as binary classifier. We performed the experiments on the HateEval2019corpus to study the influence of each proposed feature. Our baseline system without proposedfeatures achieves 65.7% of macro-F1 score on the test corpus. Surprisingly, HFW degrades thesystem performance and decreases the macro-F1 by 14 points compared to the baseline. Thiscan be due to the fact that some words are hateful only in a particular context. UP, RatioHWand PUNCT slightly degrade the baseline system. The POS features do not change the baselinesystem result and so are probably not correlated to the hate speech. The best result is obtainedusing EMO features with 66.0% of macro-F1. EMOs are largely used to transmit emotions. Inour system,they are modeled by a specific embedding vector. USE does not take into account theemojis. Therefore, EMOs give additional information to USE about the hateful content of tweets.
Fichier principal
Vignette du fichier
CFP___Offensive_language_on_social_media___International_Pragmatics_Conference_panel.pdf (60.53 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03115781 , version 1 (19-01-2021)


  • HAL Id : hal-03115781 , version 1


Nicolas Zampieri, Irina Illina, Dominique Fohr. A comparative study of different features for efficient automatic hate speech detection. IPrA 2021 - 17th International Pragmatics Conference, Jun 2021, Winterthur, Switzerland. ⟨hal-03115781⟩
225 View
146 Download


Gmail Facebook Twitter LinkedIn More