Current research on hate speech analysis is typically oriented towards monolingual and single classiifcation tasks. In this paper, we present a new multilingual hate speech analysis dataset for English, Hindi, Arabic, French, German and Spanish languages for multiple domains across hate speech - Abuse, Racism, Sexism, Religious Hate and Extremism. To the best of our knowledge, this paper is the first to address the problem of identifying various types of hate speech in these five wide domains in these six languages. In this work, we describe how we created the dataset, created annotations at high level and low level for diferent domains and how we use it to test the current state-of-the-art multilingual and multitask learning approaches. We evaluate our dataset in various monolingual, cross-lingual and machine translation classification settings and compare it against open source English datasets that we aggregated and merged for this task. Then we discuss how this approach can be used to create large scale hate-speech datasets and how to leverage our annotations in order to improve hate speech detection and classification in general.
Abusive language is an important and relevant issue in social media platforms such as Twitter. Social media is often exploited to propagate toxic content such as hate speech or other forms of abusive language. The amount of user-generated content produced every minute is very large, and manually monitoring abusive behavior in Twitter is not feasible and impractical. Twitter has made eforts to eliminate abusive content from their platform by providing clear policies on hateful conduct, user reporting and using moderators to filter content. Still, these manual eforts are not scalable enough and are not long term.
Several studies from the Natural Language Processing (NLP) field have been done to tackle the problem of hate speech detection in social media. Most studies proposed a supervised approach to detect abusive content automatically using various models ranging from traditional machine learning approaches to deep learning based approaches. However, the majority of work focused only on a single language, i.e., English, and a single abusive domain phenomenon, e.g., hate speech, sexism, racism, religious hate and so on, rather than multiple languages and multiple domains. Twitter supports content in 34 languages and user can use any one of them to express views. Thus the problem of tackling hateful content in real time in multiple languages becomes a challenge. We need robust models for hateful content detection across multiple languages and multiple domains.
In this paper we try to tackle two prominent challenges in hate speech detection1. Build a multilingual dataset for hate speech detection across 6 languages -: English,
Hindi, French, Arabic, German, Spanish. 2. Build a multi-domain dataset that covers these hate speech domains -: Abusive, Racism,
Sexism, Religious Hate and Extremism We define the diferent domains as follows-: 1. RACISM: Discrimination based on race, ethnicity, caste, nationality, culture, skin colour, hair texture, physical aspects. 2. SEXISM: Discrimination based on gender/sexual orientation. 3. RELIGIOUS HATE: Religious discrimination treating a person or group diferently because of the particular faith/belief which they hold about a religion. 4. ABUSE: Speech that causes or likely to cause distress, disrespect or mental pain, especially from vulgar and profane comments. 5. EXTREMISM: Speech that cause or is likely to cause, harm to individuals, communities or wider society, and where any political, civil issues can lead to extremist behaviour through violence.
There have been several studies on abusive language detection [
Along with that, there are several shared tasks that have focused on abusive language and
hate speech detection such as HASOC-2019 [
Waseem[
Wide variety of machine learning models have been used to deal with multi-domain hate
speech detection task. Some studies use traditional machine learning approaches such as, logistic
regression [
For cross-lingual abusive language detection, most studies utilized transformers based models.
Some traditional models used such as logistic regression [28] [
In this section, we describe the characteristics that we want to include in our dataset, our approach to collect diferent types of hate speech while covering all major hate speech domains and how to annotate data at large scale. We also give detailed statistics and analysis for the collected data.
Considering no such dataset is available at present that covers these five domains of hate speech in the six languages, our aim is to create a reliable multilingual and multifaceted hate speech dataset.
Multilingual data: Our dataset is created as multilingual resource to facilitate cross-lingual research. It contains hate speech in English, Hindi, French, Arabic, German and Spanish languages. Multi-Domain data: Our dataset consists of fine-grained labels for each hate sample per language. These aspects cover majority of hate domains such as racism, sexism, abuse, religious hate and extremism.
Considering the cultural diferences in the main regions where English, Hindi, French, Arabic, Spanish and German are spoken, we start by looking for the hateful keywords that are native to these languages. We use HateBase vocabulary dataset, which is a valuable lexicon for creating hateful dataset from public forums, as well as Hatebase’s sightings dataset, which is useful for trending analysis of keywords.
We targeted 6 languages. Additionally we tried some indic languages mainly Marathi and Bangla but hatebase has negligible coverage for these languages in terms of keywords. Additionally, we extracted the following: 1. Targeted groups for the keywords 2. Ofensive levels of keywords - Extremely, Highly and Mildly ofensive. 3. Recent sighting counts
Target groups help us categorise keywords in low level classes of ethnicity & nationality, religious hate, gender & sexual orientation and general abuse. Details of keywords per domain and language are provided in Table 1 in dataset statistics. Total keywords collected were 1022 . To extract data for extremism, we used a set of keywords related to extremism and terrorism (including terrorist organisation names) to retrieve news articles from BBC Monitoring1. 1. We further extracted comments from those articles related to extremism and terrorism. We also extracted tweets and comments which were extremist in nature from the counterextremism database2. 2. The comments extracted were said by members and suspects of terrorist and extremist organisations on Facebook, Twitter and YouTube.
We started our dataset collection by using the keywords built per language. We utilize twint3
API to collect 1000-2000 tweets per keyword as Twitter oficial API can only be utilized for
limited number of requests. We searched through the API for last one year of data. We also add
additional keywords from MLMA [
The collected raw data contains cross-lingual tweets and therefore language detection becomes a part of our process. For each keyword in each language we consider tweets only in that language and drop the rest. This helps us in training monolingual hate speech detection without any need to worry for code switching in languages.
We substituted all usernames with @USER and urls with @URL and cleaned any unnecessary symbols from the tweet. We also discarded short tweets with less than 4 words. In total we collected 497660 tweets. Details of tweets per language is given in section 3.3 Dataset Statistics.
For extremism class we carefully handcrafted a total of 88 keywords and collected the data from BBC monitoring. Additional data was collected from other websites for extremism.
Table 1 gives us the distribution of keywords across diferent languages and diferent domains. We have merged classes from hate base to arrive at those domains. Sexism contains keywords that belong to gender and sexual orientation. Racism includes classes of ethnicity and nationality. Religious Hate includes words belonging to Religion category on Hatebase 4. Extremism keywords were handcrafted as described in section 3.2.
English was the dominant language in terms of the total keywords (500) extracted. Besides Hindi, racism is the major domain for which all languages have highest number of hate keywords.
In order to build the dataset for diferent tasks with gold labels, we utilize 10 publicly available
datasets for diferent types (tasks) of hate speech detection. We sampled annotated examples
from diverse English datasets: GAO hatefulness [
We map labels from these datasets into diferent domains: abusive, non-abusive, sexism,
racism and religious hate. For OLID [37] OFFENSIVE maps to Abuse. GAO and WUL had binary
2https://www.counterextremism.com/daily-dose-archive
3https://github.com/twintproject/twint/
4https://hatebase.org/
labels, while the original TRAC uses three labels: non-aggressive, covertly-aggressive, and
openly-aggressive. We relabel the first as non-abusive, and the other two as abusive. For MLMA
we create labels on the basis of target groups: target group RELIGION becomes RELIGIOUS
HATE. For MLMA [
For some of these data we had to hydrate tweets from Twitter using oficial Twitter API as only the tweet ids had been provided. For Waseem and Founta [38], majority of the tweets were not available on Twitter which led to less number of tweets compared to the oficial dataset. We standardized these datasets into a single HSMerge dataset. Details of classes and total samples is available in Table 3.
In this section, we describe the various components of our pipeline as shown in figure 1 and the models used for semi-supervised annotation process in diferent settings. Most of the hate speech detection tasks depend on manual annotation that limits the number of samples that can be labeled. We used a hierarchical approach to validate and refine our initial keyword based dataset from Twitter.
1. Monolingual hate speech detection models from Hate-ALERT used as zero shot pipeline to get high level labels of Hate or NoHate in monolingual settings. 2. Multilingual mbert fine-tuned model to get high level labels. 3. Translate the raw data of five languages (other than English) to English. 4. Use multilingual binary classifier (developed in step 2) on translated data to classify tweets at high level. 5. Utilize Google Perspective API to predict toxicity of tweets for high level labels. 6. Use two way voting to predict final binary labels. 7. Fine-tuned "distilbert-base-uncased" model on only the English dataset, and used it to do annotations for the English translations of the data collected from Twitter in 5 languages. 8. To validate our domain specific labels we trained various models on HSMerge data shown in Table 3 and predicted on hateful samples obtained from step 6.
To evaluate the performance of the models, we used weighted average F1 for benchmarking on validation set. All the experiments were done on NVIDIA A100 GPU with up to 20 GiB RAM.
We utilised hate speech models to perform initial validation on the raw LAHM dataset; details can be found in Section 5.
We utilized BERT based language specific hate speech models from HuggingFace. The models used were trained, validated and tested on the same language.
We utilized the following language models for our experiments: 1. Hate-speech-CNERG/dehatebert-mono-english 5 2. l3cube-pune/hate-multi-roberta-hasoc-hindi6 3. Hate-speech-CNERG/dehatebert-mono-french7 4. Hate-speech-CNERG/dehatebert-mono-german8 5. Hate-speech-CNERG/dehatebert-mono-spanish9
Binary hate labels distribution results from monolingual experiments are presented in table 8
We adopted the methodology to utilize open source machine translation models for translation of English data to multilingual data and vice-versa. To select the best machine translation model for each of our languages, we evaluated a number of models on a small manually annotated dataset. The translations were carried out for each of the these languages: Hindi, Arabic, French, German and Spanish. Two types of translations were carried out using the translation models: 1. Translation of English dataset collected from various open source hate speech datasets.
All the English language samples in this dataset were translated to other 5 languages. 2. Translation of multilingual data collected from Twitter. This data was collected for the above 5 languages, and each of this language data was translated into corresponding English language data.
Consideration for choosing and evaluating the translation models was based on whether they were open source, free/easy to use, and the translation quality. Translation models used were: 1. Google sheets translation 2. m2m-100-1.2B10 3. IndicTrans11
The evaluation metrics used were bleu, rougeL and semantic similarity. For semantic similarity, all-mpnet-base-v2 model from sentence-transformers library was used to calculate the cosine score between input and translated sentence embeddings. Figure 4.3.2 shows comparisons of bleu, rougeL and semantic similarity scores for the 3 models on diferent languages. After comparing the performances and taking other considerations into account, indicTrans model was selected for all translations of Indic languages (Hindi), and m2m-100-1.2B was selected for the other languages.
5https://huggingface.co/Hate-speech-CNERG/dehatebert-mono-english 6l3cube-pune/hate-roberta-hasoc-hindi 7https://huggingface.co/Hate-speech-CNERG/dehatebert-mono-french 8https://huggingface.co/Hate-speech-CNERG/dehatebert-mono-german 9https://huggingface.co/Hate-speech-CNERG/dehatebert-mono-spanish 10https://github.com/UKPLab/EasyNMT 11https://github.com/AI4Bharat/indicTrans
We utilized mbert which is a pre-trained model on the top 104 languages with the largest Wikipedia corpus using a masked language modeling (MLM) objective, to fine-tune for binary classification task to perform few shot learning experiments. The aim was to make the classifier learn over samples from a few languages and test over other languages. The training dataset was predominant in English, curated from multiple open source resources. In addition to that we introduced language samples for a few other languages using a neural machine translation based open source model (m2m_100_1.2B)12. NMT is an approach to achieve machine translation using artificial neural nets to predict the best possible sequence of words.
The model was trained, tested and validated on the custom multilingual dataset curated with the class distribution shown in Table 4 per chosen language. We utilized this model to make predictions on the curated dataset.
Perspective API is a free to use API that uses machine learning to identify toxicity in the comments. It has API rate limitations of 1 request per second. Due to this limitation we utilized it on all languages except English. We utilized Google Cloud to generate the Perspective API key for sending requests, and specifically CommentAnalyzer API to get the toxicity score for each data point. The CommentAnalyzer API supports all the 5 languages we have considered. We used the Twitter cleaned data we had collected for the 5 languages to get the Toxicity score to classify each tweet as hate or not-hate. Numbers are reported in Table 8.
This section describes the multi-domain experiments to validate labels on the LAHM dataset.
HSMerge data present in Table 3 was used for the multi-domain classification of hate speech. This dataset was pre-processed by removing URLs, mentions/usernames, null values and duplicates.
The aim was to utilize transfer learning to fine-tune a model on only the English dataset, and use it to do annotations for the English translations of the data collected from Twitter in other 5 languages.
12https://github.com/pytorch/fairseq/tree/main/examples/m2m100 1. 20,174 training samples 2. 5,044 validation samples 3. 6,305 test samples
The model used for fine-tuning on this dataset was "distilbert-base-uncased"13. DistilBERT is a distilled version of the BERT base model. It has 40 percent less parameters than bert-baseuncased, is 60 percent faster and retains almost 97 percent of BERT’s performances14. A linear layer on top of the pooled output of the model was added for the multi-class classification. During training, the maximum sequence length was limited to 256, batch size was set to 32, learning rate 5e-5 and was trained for 4 epochs. The dataset was split into train, validation and test set using stratified random sampling. Dataset details :
The model evaluation metrics on the test set are given in Table 5.
To validate our fine-grained labels with keywords from hatebase, we experimented with crosslingual models trained on HSmerge dataset as shown in Table 3. In this experiment setting we fine tuned the mbert model for the multi-class classification task and trained the model on HSmerge data in Table 3 and do the prediction on the other languages. Empirical results can be found under Table 7. In this zero-shot setting, no other language samples were given to the model after fine-tuning on the English dataset and also machine translation was not involved at all in either the pre-training or fine-tuning.
We couldn’t manage to maintain the class distribution, and the extremism class was left with comparatively less data points. To counter the issue we introduced custom class weights to provide equal attention to the minority class which in our case was extremism. This was done using a weighted class random sampler WeightedSampler for class imbalance, so that all classes have equal probability. The empirical scores achieved by the cross-lingual model can be found in the Table 7.
13https://huggingface.co/distilbert-base-uncased 14https://arxiv.org/abs/1910.01108 1. 19298 training samples 2. 6432 validation samples 3. 6434 test samples
The open source mbert model we fine-tuned for our use case has previously been tested for zero-shot experimentation15 and did manage to achieve decent empirical scores shown in Table 6. Dataset details :
Domains
precision recall f1-score support racism 0.93 sexism 0.93 religious hate 0.93 abuse 0.93 extremism 0.85 accuracy macro avg weighted avg
We performed initial level of validation (Hatefulness) experiments on the raw LAHM dataset using open source monolingual models, Perspective API and the multilingual binary classifier we trained. We compared the results for monolingual and multilingual and noticed that our multilingual few-shot learning based classifier out-performed the open source monolingual BERT based models from HuggingFace except on Arabic and French languages where the latter did a better job; details can be found in Table 8. Based on the analysis we took at least 2 votes from the models which performed better, details shown in Table 9.
For Hindi, Perspective API performed better by predicting 31 percent as hate compared to 12 percent for binary classifier. For Arabic, the binary classifier performed significantly better.
For domain validation we utilised the zero-shot cross-lingual, multi-domain and machine translation based multi-domain models trained with HSmerge data in Table 3 and performed 15https://github.com/google-research/bert/blob/master/multilingual.md#results
Language Mono-Lingual Multi-lingual Perspective API English ✓ ✓ Hindi ✓ ✓ Arabic ✓ ✓ French ✓ ✓ German ✓ ✓
Spanish ✓ ✓ Language
Abuse
Sexism
Racism
Religious
Extremism English Arabic Hindi French German Spanish Total 4228 1895 1511 813 180 818 2934 994 440 779 123 483 1566 736 282 2393 30 139 predictions on the diferent languages. Label distribution from cross-lingual models are shown in Table 10. Label distribution of predictions for diferent languages on the LAHM dataset from multi-class translated model is in Table 11.
For Hindi, 31.6 percent samples were Abuse, while the lowest were 10.3 percent for Racism. For Arabic, Sexism and Extremism contributed to 73 percent of all hate samples. For French, 32 percent were Abuse and lowest 8.3 percent were Religious Hate. For German, 51 percent were Sexism while the lowest were Religious Hate and Extremism together amounting to 13 percent. For Spanish, Abuse and Sexism had 27 percent each, while Religious Hate had 5 percent of samples.
We have presented the LAHM dataset, a large scale semi-supervised training dataset for multilingual and multi-domain hate speech identification, we created by using 3 layer annotation pipeline and combination of monolingual, multilingual and cross-lingual models. To the best of our knowledge, LAHM is the largest of its kind, containing close to 300k tweets across 6 languages and 5 domains.
LAHM enables cross-lingual abusive language detection across five domains and in-depth interplay between language shift and domain shift. We have profiled LAHM as a comprehensive resource for evaluating hate speech detection through a series of cross-domain experiments in monolingual, multilingual and cross-lingual setups with state of the art transfer learning models.
We hope that LAHM will inspire more eforts in understanding and building semi-supervised large scale multilingual and multi-domain abusive language detection datasets.
For future work we explored leetspeak detection and identification on social networks. A lot of hate content on social media uses leetspeak to evade moderators and automated systems. We collected hate keywords belonging in this category for each of the 5 domains, and experimented with a set of leets for each keyword to extract the leetspeak hate content. We plan to use this for future work in multilingual and multi-domain settings. platform arabic news comment dataset for ofensive language detection, in: Proceedings of the 12th Language Resources and Evaluation Conference, 2020, pp. 6203–6212. [14] M. Wiegand, J. Ruppenhofer, A. Schmidt, C. Greenberg, Inducing a lexicon of abusive words–a feature-based approach (2018). [15] M. Karan, J. Šnajder, Cross-domain detection of abusive language online, in: Proceedings of the 2nd workshop on abusive language online (ALW2), 2018, pp. 132–137. [16] E. W. Pamungkas, V. Patti, Cross-domain and cross-lingual abusive language detection: A hybrid approach with deep learning and a multilingual lexicon, in: Proceedings of the 57th annual meeting of the association for computational linguistics: Student research workshop, 2019, pp. 363–370. [17] E. W. Pamungkas, V. Basile, V. Patti, Misogyny detection in twitter: a multilingual and cross-domain study, Information Processing & Management 57 (2020) 102360. [18] J. S. Meyer, B. Gambäck, A platform agnostic dual-strand hate speech detector, in: ACL 2019 The Third Workshop on Abusive Language Online Proceedings of the Workshop, Association for Computational Linguistics, 2019. [19] K. Wang, D. Lu, S. C. Han, S. Long, J. Poon, Detect all abuse! toward universal abusive language detection models, arXiv preprint arXiv:2010.03776 (2020). [20] A. Arango, J. Pérez, B. Poblete, Hate speech detection is not as easy as you may think: A closer look at model validation, in: Proceedings of the 42nd international acm sigir conference on research and development in information retrieval, 2019, pp. 45–54. [21] Z. Waseem, J. Thorne, J. Bingel, Bridging the gaps: Multi task learning for domain transfer of hate speech detection, in: Online harassment, Springer, 2018, pp. 29–55. [22] M.-A. Rizoiu, T. Wang, G. Ferraro, H. Suominen, Transfer learning for hate speech detection in social media, arXiv preprint arXiv:1906.03829 (2019). [23] T. Caselli, V. Basile, J. Mitrović, M. Granitzer, Hatebert: Retraining bert for abusive language detection in english, arXiv preprint arXiv:2010.12472 (2020). [24] A. Koufakou, E. W. Pamungkas, V. Basile, V. Patti, Hurtbert: incorporating lexical features with bert for the detection of abusive language, in: Proceedings of the fourth workshop on online abuse and harms, 2020, pp. 34–43. [25] G. Glavaš, M. Karan, I. Vulic, Xhate-999: Analyzing and detecting abusive language across domains and languages, Association for Computational Linguistics, 2020. [26] M. Mozafari, R. Farahbakhsh, N. Crespi, Hate speech detection and racial bias mitigation in social media based on bert model, PloS one 15 (2020) e0237861. [27] K. B. Ozler, K. Kenski, S. Rains, Y. Shmargad, K. Coe, S. Bethard, Fine-tuning bert for multi-domain and multi-label incivil language detection, in: Proceedings of the Fourth Workshop on Online Abuse and Harms, 2020, pp. 28–33. [28] S. S. Aluru, B. Mathew, P. Saha, A. Mukherjee, A deep dive into multilingual hate speech classification, in: Machine Learning and Knowledge Discovery in Databases. Applied Data Science and Demo Track: European Conference, ECML PKDD 2020, Ghent, Belgium, September 14–18, 2020, Proceedings, Part V, Springer International Publishing, 2021, pp. 423–439. [29] N. Vashistha, A. Zubiaga, Online multilingual hate speech detection: experimenting with hindi and english social media, Information 12 (2021) 5. [30] M. O. Ibrohim, I. Budi, Translated vs non-translated method for multilingual hate speech identification in twitter, Int. J. Adv. Sci. Eng. Inf. Technol 9 (2019) 1116–1123. [31] M. Corazza, S. Menini, E. Cabrio, S. Tonelli, S. Villata, Hybrid emoji-based masked language models for zero-shot abusive language detection, in: Findings of the Association for Computational Linguistics: EMNLP 2020, Association for Computational Linguistics, Online, 2020, pp. 943–949. URL: https://aclanthology.org/2020.findings-emnlp.84. doi: 10. 18653/v1/2020.findings-emnlp.84. [32] H. Ahn, J. Sun, C. Y. Park, J. Seo, NLPDove at SemEval-2020 task 12: Improving ofensive language detection with cross-lingual transfer, in: Proceedings of the Fourteenth Workshop on Semantic Evaluation, International Committee for Computational Linguistics, Barcelona (online), 2020, pp. 1576–1586. URL: https://aclanthology.org/2020.semeval-1.206. doi:10. 18653/v1/2020.semeval-1.206. [33] J. M. Pérez, A. Arango, F. Luque, ANDES at SemEval-2020 task 12: A jointly-trained BERT multilingual model for ofensive language detection, in: Proceedings of the Fourteenth Workshop on Semantic Evaluation, International Committee for Computational Linguistics, Barcelona (online), 2020, pp. 1524–1531. URL: https://aclanthology.org/2020.semeval-1.199. doi:10.18653/v1/2020.semeval-1.199. [34] L. Stappen, F. Brunn, B. Schuller, Cross-lingual zero-and few-shot hate speech detection utilising frozen transformer language models and axel, arXiv preprint arXiv:2004.13850 (2020). [35] T. Dadu, K. Pant, Team rouges at SemEval-2020 task 12: Cross-lingual inductive transfer to detect ofensive language, in: Proceedings of the Fourteenth Workshop on Semantic Evaluation, International Committee for Computational Linguistics, Barcelona (online), 2020, pp. 2183–2189. URL: https://aclanthology.org/2020.semeval-1.290. doi:10.18653/ v1/2020.semeval-1.290. [36] T. Ranasinghe, M. Zampieri, Multilingual ofensive language identification with crosslingual embeddings, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Association for Computational Linguistics, Online, 2020, pp. 5838–5844. URL: https://aclanthology.org/2020.emnlp-main.470. doi:10.18653/ v1/2020.emnlp-main.470. [37] M. Zampieri, P. Nakov, S. Rosenthal, P. Atanasova, G. Karadzhov, H. Mubarak, L. Derczynski, Z. Pitenis, Ç. Çöltekin, Semeval-2020 task 12: Multilingual ofensive language identification in social media (ofenseval 2020), arXiv preprint arXiv:2006.07235 (2020). [38] A. M. Founta, C. Djouvas, D. Chatzakou, I. Leontiadis, J. Blackburn, G. Stringhini, A. Vakali, M. Sirivianos, N. Kourtellis, Large scale crowdsourcing and characterization of twitter abusive behavior, in: Twelfth International AAAI Conference on Web and Social Media, 2018. [39] R. Gomez, J. Gibert, L. Gomez, D. Karatzas, Exploring hate speech detection in multimodal publications, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2020, pp. 1470–1478. [40] Fanton, Margherita and Bonaldi, Helena and Tekiroğlu, Serra Sinem and Guerini, Marco, Human-in-the-Loop for Data Collection: a Multi-Target Counter Narrative Dataset to Fight Online Hate Speech, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, 2021.