arXiv:2110.12687v1  [cs.CL]  25 Oct 2021Fine-tuning of Pre-trained Transformers for Hate, Offensi ve,
and Profane Content Detection in English and Marathi
Anna Glazkova1,*, Michael Kadantsev2, Maksim Glazkov3,
1 University of Tyumen, Tyumen, Russia
2 Thales Canada, Transportation Solutions, Toronto, Canad a
3 Neuro.net, Nizhny Novgorod, Russia
* a.v.glazkova@utmn.ru
Abstract
This paper describes neural models developed for the Hate Speech and Offensive Content
Identification in English and Indo-Aryan Languages Shared Task 20 21. Our team called
neuro-utmn-thales participated in two tasks on binary and fine-grained classification of
English tweets that contain hate, offensive, and profane conten t (English Subtasks A &
B) and one task on identification of problematic content in Marathi ( Marathi Subtask
A). For English subtasks, we investigate the impact of additional co rpora for hate speech
detection to fine-tune transformer models. We also apply a one-vs -rest approach based
on Twitter-RoBERTa to discrimination between hate, profane and o ffensive posts. Our
models ranked third in English Subtask A with the F1-score of 81.99% a nd ranked
second in English Subtask B with the F1-score of 65.77%. For the Mar athi tasks, we
propose a system based on the Language-Agnostic BERT Sentenc e Embedding (LaBSE).
This model achieved the second result in Marathi Subtask A obtainin g an F1 of 88.08%.
Introduction
Social media has a greater impact on our society. Social networks g ive us almost limitless
freedom of speech and contribute to the rapid dissemination of info rmation. However,
these positive properties often lead to unhealthy usage of social m edia. Thus, hate
speech spreading affects users’ psychological state, promote s violence, and reinforces
hateful sentiments [4, 5]. This problem attracts many scholars to a pply modern tech-
nologies in order to make social media safer. The Hate Speech and Of fensive Content
Identification in English and Indo-Aryan Languages Shared Task (H ASOC) 2021 [27]
aims to compare and analyze existing approaches to identifying hate speech not only
for English, but also for other languages. It focused on detecting hate, offensive, and
profane content in tweets, and offering six subtasks. We particip ated in three of them:
•English Subtask A: identifying hate, offensive, and profane content from the
post in English [24].
•English Subtask B: discrimination between hate, profane, and offensive posts
in English.
•Marathi Subtask A: identifying hate, offensive, and profane content from the
post in Marathi [14].
1/10The source code for our models is freely available1.
The paper is organized as follows. Section 2 contains a brief review of related works.
Next, we describe our experiments on the binary and fine-grained c lassification of En-
glish tweets in Section 3. In Section 4, we present our model for hat e, offensive, and
profane language identification in Marathi. We conclude this paper in S ection 5. Finally,
Section 6 contains acknowledgments.
1 Related Works
We briefly discuss works done related to harmful content detectio n in the past few
years. Shared tasks related to hate speech and offensive langua ge detection from tweets
was organized as a part of some workshops and conferences, suc h as FIRE [22, 23], Se-
mEval, [3, 10], GermEval [40, 43], IberLEF [42], and OSACT [29]. The pa rticipants
proposed a broad range of approaches from traditional machine le arning techniques
(for example, Support Vector Machines [15, 38], Random Forest [34 ]) to various neural
architectures (Convolutional Neural Networks, CNN [35]; Long Sh ort Term Memory,
LSTM [26, 28]; Embeddings from Language Models, ELMo [6]; and Bidirec tional En-
coder Representations from Transformers, BERT [19,36]). In mo st cases, BERT-based
systems outperformed other approaches.
Most research on hate speech detection continues to be based on English corpora.
Despite this, the harmful content is distributed in different langua ges. Therefore, there
have been previous attempts at creating corpora and developing m odels for hate speech
detection in common non-English languages, such as Arabic [1, 29], G erman [22, 23, 40,
43], Italian [7, 37], Spanish [3, 42], Hindi [22, 23], Tamil and Malayalam [22]. Several
studies have focused on collecting hate speech corpora for Chines e [41], Portuguese [11],
Polish [33], Turkish [8] and Russian [16] languages.
2 English Subtasks A & B: Identification and Fine-
grained Classification of Hate, Offensive, and Pro-
fane Tweets
The objective of English Subtasks A & B is to identify whether a tweet in English
contains harmful content (Subtask A) and perform a fine-graine d classification of posts
into three categories, including: hate, offensive, or profane (Su btask B).
2.1 Data
The dataset provided to the participants of the shared task cont ains 4355 manually
annotated social media posts divided into training (3074) and test ( 1281) sets. Table 1
presents the data description.
Further, we tested several data sampling techniques using differ ent hate speech cor-
pora as additional training data. Firstly, we evaluated the joint use of multilingual data
provided by the organizers of HASOC 2021, including the English, the Hindi, and the
Marathi training sets. Secondly, as the training sets were highly imb alanced, we applied
the positive class random oversampling technique so that each train ing batch contained
approximately the same number of samples. Besides, we experiment ed with the seq2seq-
based data augmentation technique [17]. For this purpose, we fine- tuned the BART-base
model for the denoising reconstruction task where 40% of tokens are masked and the
goal of the decoder is to reconstruct the original sequence. Sinc e the BART model [18]
1https://github.com/ixomaxip/hasoc
2/10Table 1. Data description.
Label DescriptionNumber of examples
(training set)
Subtask A
NOTNon Hate-Offensive: the post does not contain
hate speech, profane, offensive content1102
HOFHate and Offensive: the post contains hate,
offensive, or profane content.1972
Subtask B
NONEThe post does not contain hate speech, profane,
offensive content1102
HATE Hate speech: the post contains hate speech content. 542
OFFN Offensive: the post contains offensive content. 482
PRFN Profane: the post contains profane words. 948
Table 2. Hate-related dataset characteristics.
Dataset Size Labels
HASOC 2020 4522HOF - 50.4%
NOT - 49.6%
HatebaseTwitter 24783hate speech - 20.15%
offensive language - 85.98%
neither - 23.77%
HatEval 130001 (hate speech) - 42.08%
0 (not hate speech) - 57.92%
OLID 14100OFF - 32.91%
NOT - 67.09%
already contains the ¡mask¿ token, we use it to replace mask tokens . We generated
one synthetic example for every tweet in the training set. Thus, th e augmented data
size is the same size as the size of the original training set. Finally, we e valuated the
impact of additional training data, including: (a) the English dataset , used at HASOC
2020 [22]; (b) HatebaseTwitter, based on the hate speech lexicon f rom Hatebase2[10];
(c) HatEval, a dataset presented at Semeval-2019 Task 5 [3] ; (d) Offensive Language
Identification Dataset (OLID), used in the SemEval-2019 Task 6 (O ffensEval) [45]. All
corpora except the HatebaseTwitter dataset contain non-inter sective classes. Besides,
all listed datasets are collected from Twitter. A representative sa mpling of additional
data is shown in Table 2.
We preprocessed the datasets for Subtasks A & B in a similar manner . Inspired
by [2], we used the following text preprocessing technique3: (a) removed all URLs; (b)
replaced all user mentions with the $MENTION $placeholder.
2.2 Models
We conduct our experiments with neural models based on BERT [12] a s they have
achieved state-of-the-art results in harmful content detectio n. For example, BERT-
based models proved efficient at previous HASOC shared tasks [22 , 23] and SemEval
[32,45].
We used the following models:
2https://hatebase.org/
3https://pypi.org/project/tweet-preprocessor
3/10Table 3. Model validation results for English Subtask A, %.
Model F1 P R
BERT 79.24 79.74 78.82
BERTweet 78.65 79.36 78.08
Twitter-RoBERTa 81.1 80.01 82.65
LaBSE (English dataset) 78.83 79.5 78.29
LaBSE (English + Hindi) 79.32 79.95 78.8
LaBSE (English, Hindi, and Marathi) 79.27 81.74 77.79
Adding extra data to Twitter-RoBERTa
+ random oversampling 79.97 79.9 80.04
+ BART data augmentation 79.24 78.44 80.31
+ HASOC 2020 78.79 77.66 80.47
+ HatabaseTwitter 81.19 79.99 82.93
+ HatEval 74.31 75.53 73.64
+ OLID 79.29 78.17 80.93
•BERT base[12], a pre-trained model on BookCorpus [46] and English Wikipedia
using a masked language modeling objective.
•BERTweet base[30], a pre-trained language model for English tweets. The corpus
used to pre-train BERTweet consists of 850M English Tweets includin g 845M
Tweets streamed from 01/2012 to 08/2019 and 5M Tweets related to the COVID-
19 pandemic.
•Twitter-RoBERTa basefor Hate Speech Detection [2], a RoBERTa base[20] model
trained on 58M tweets and fine-tuned for hate speech detection w ith the Tweet-
Eval benchmark.
•LaBSE [13], a language-agnostic BERT sentence embedding model su pporting 109
languages.
2.3 Experiments
For both Subtask A and Subtask B, we adopted pre-trained models from HuggingFace
[44] and fine-tuned them using PyTorch [31]. We fine-tuned each pre -trained language
model for 3 epochs with the learning rate of 2e-5 using the AdamW op timizer [21]. We
set batch size to 32 and maximum sequence size to 64. To validate our models during
the development phase, we divided labelled data using the train and va lidation split in
the ratio 80:20.
Table 3 shows the performance of our models on the validation subse t for Subtask A
in terms of macro-averaging F1-score (F1), precision (P), and re call (R). As can be seen
from the table, BERT, BERTweet, and LaBSE show very close result s during validation.
Despite this, LaBSE jointly fine-tuned on three mixed multilingual dat asets shows the
highest precision score. The use of Twitter-RoBERTa increases th e F1-score by 1.5-2.5%
compared to other classification models. Based on this, we chose Tw itter-RoBERTa for
further experiments. We found out that neither the random over sampling technique
nor the use of the augmented and additional data shows a perform ance improvement,
except the joint use of the original dataset and the HatebaseTwit ter dataset that gives
an F1-score growth of 0.09% and a precision growth of 0.28% compar ed to basic Twitter-
RoBERTa.
For our official submission for Subtask A, we designed a soft-votin g ensemble of five
Twitter-RoBERTa jointly fine-tuned on the original training set and the HatebaseTwit-
4/10Table 4. Performance of our final models for English Subtasks A & B, official results,
%.
SubtaskF1 (our
model)F1 (winning
solution)P (our
model)P (winning
solution)Avg F1Number of
submitted
teamsRank
A 81.99 83.05 84.68 84.14 75.7 56 3
B 65.77 66.57 66.32 66.88 57.07 37 2
ter dataset (see Table 4). For Subtask B, we used the following one -vs-rest approach to
discrimination between hate, profane, and offensive posts.
•First, we applied our Subtask A binary models to identify non hate-of fensive
examples.
•Second, we fine-tuned three Twitter-RoBERTa binary models to de limit exam-
ples of hate-vs-profane, hate-vs-offensive, and offensive-v s-profane classes. The
training dataset was extended with the HatebaseTwitter dataset .
•Finally, we compared the results of binary models. If the result was d efined
uniquely, we used it as a predicted label. Otherwise, we chose the labe l in propor-
tion to the number of examples in the training set.
This can be illustrated briefly by the following examples.
–Let the models show the following results:
∗hate-vs-profane →hate;
∗hate-vs-offensive →hate;
∗offensive-vs-profane →offensive.
Thus, classes have the following votes: hate – 2, offensive - 1, pro fane – 0.
Then we predict the HATE label.
–If the results are:
∗hate-vs-profane →profane;
∗hate-vs-offensive →hate;
∗offensive-vs-profane →offensive,
we have the class votes, such as hate – 1, offensive - 1, profane – 1. Then we
choose the PRFN label as the most common label in the training set.
3 Marathi Subtask A: Identifying Hate, Offensive,
and Profane Content from the Post
3.1 Data
For the Marathi task, we used the original training and test sets p rovided by the orga-
nizers of the HASOC 2021. The whole dataset contains 2499 tweets , including: 1874
training and 625 test examples. The training set consists of 1205 te xts of the NOT
class and 669 texts of the HOF class. We used raw data as an input fo r our models.
Following [25,39], we experimented with the combination of the English, the Hindi, and
the Marathi training sets provided by the organizers.
5/10Table 5. Model validation results for Marathi Subtask A, %.
Model F1 P R
XLM-RoBERTa (Marathi dataset) 83.87 85.39 83.39
XLM-RoBERTa (Marathi + Hindi) 83.23 83.82 82.76
XLM-RoBERTa (Marathi + English) 84.83 85.03 84.64
XLM-RoBERTa (Marathi + Hindi + English) 84.35 84.82 83.95
LaBSE (Marathi) 87.76 87.82 87.68
LaBSE (Marathi + Hindi) 87.62 88.21 87.13
LaBSE (Marathi + English) 87.62 88.21 87.13
LaBSE (Marathi + Hindi + English) 86.34 86.63 86.08
Table 6. Performance of our final model for the Marathi Subtask A, offic ial results, %.
F1 (our
model)F1 (winning
solution)P (our
model)P (winning
solution)Avg F1Number of
submitted
teamsRank
88.08 91.44 87.58 91.82 82.55 25 2
3.2 Models
We evaluated the following models:
•XLM-RoBERTa base[9], a transformer-based multilingual masked language model
supporting 100 languages.
•LaBSE [13], a language-agnostic BERT sentence embedding model pr e-trained on
texts in 109 languages.
3.3 Experiments
We experimented with the above-mentioned language models fine-tu ned on monolingual
and multilingual data. For model evaluation during the development p hase, we used
the random train and validation split in the ratio 80:20 with a fixed seed. We set the
same model parameters as for English tasks.
Table 5 illustrates the results. It can be seen that LaBSE outperfo rms XLM-
RoBERTa in all cases. Moreover, the F1-score of LaBSE fine-tune d only on the Marathi
dataset are higher than the results of LaBSE fine-tuned on multiling ual data. XLM-
RoBERTa, on the other hand, mostly benefits from multilingual fine- tuning.
For our final submission, we used a soft-voting ensemble of five LaB SE fine-tuned on
the official Marathi dataset provided by the organizers of the co mpetition. The results
of this model on the test set are shown in Table 6.
Conclusion
In this paper, we have presented the details about our participatio n in the HASOC
Shared Task 2021. We have explored an application of domain-specif ic monolingual and
multilingual BERT-based models to the tasks of binary and fine-grain ed classification of
Twitter posts. We also proposed a one-vs-rest approach to discr imination between hate,
offensive, and profane tweets. Further research can focus on analyzing the effectiveness
of various text preprocessing techniques for harmful content d etection and exploring
how different transfer learning approaches can affect classifica tion performance.
6/10Acknowledgments
The work on multi-label text classification was carried out by Anna Gla zkova and sup-
ported by the grant of the President of the Russian Federation no . MK-637.2020.9.
References
1. N. Albadi, M. Kurdi, and S. Mishra. Are they our brothers? analys is and de-
tection of religious hate speech in the Arabic twittersphere. In 2018 IEEE/ACM
International Conference on Advances in Social Networks An alysis and Mining
(ASONAM) , pages 69–76. IEEE, 2018.
2. F. Barbieri, J. Camacho-Collados, L. E. Anke, and L. Neves. Twe etEval: Unified
benchmark and comparative evaluation for tweet classification. In Proceedings
of the 2020 Conference on Empirical Methods in Natural Langu age Processing:
Findings , pages 1644–1650, 2020.
3. V. Basile, C. Bosco, E. Fersini, N. Debora, V. Patti, F. M. R. Pard o, P. Rosso,
M. Sanguinetti, et al. Semeval-2019 task 5: Multilingual detection of hate speech
against immigrants and women in Twitter. In 13th International Workshop on
Semantic Evaluation , pages 54–63, 2019.
4. L. E. Beausoleil. Free, hateful, and posted: rethinking first ame ndment protection
of hate speech in a social media world. BCL Rev. , 60:2101, 2019.
5. M. Bilewicz and W. Soral. Hate speech epidemic. the dynamic effect s of deroga-
tory language on intergroup relations and political radicalization. Political Psy-
chology , 41:3–33, 2020.
6. M. Bojkovsky and M. Pikuliak. STUFIIT at SemEval-2019 task 5: M ultilingual
hate speech detection on twitter with MUSE and ELMo embeddings. I nProceed-
ings of the 13th International Workshop on Semantic Evaluat ion, pages 464–468,
2019.
7. C. Bosco, D. Felice, F. Poletto, M. Sanguinetti, and T. Maurizio. O verview of the
EVALITA 2018 hate speech detection task. In EVALITA 2018-Sixth Evaluation
Campaign of Natural Language Processing and Speech Tools fo r Italian , volume
2263, pages 1–9. CEUR, 2018.
8. C ¸ . C ¸ ¨ oltekin. A corpus of Turkish offensive language on social media. In Proceed-
ings of the 12th Language Resources and Evaluation Conferen ce, pages 6174–6184,
2020.
9. A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzm´ an,
´E. Grave, M. Ott, L. Zettlemoyer, and V. Stoyanov. Unsupervise d cross-lingual
representation learning at scale. In Proceedings of the 58th Annual Meeting of
the Association for Computational Linguistics , pages 8440–8451, 2020.
10. T. Davidson, D. Warmsley, M. Macy, and I. Weber. Automated h ate speech de-
tection and the problem of offensive language. In Proceedings of the International
AAAI Conference on Web and Social Media , volume 11, 2017.
11. R. P. de Pelle and V. P. Moreira. Offensive comments in the Brazilia n web:
a dataset and baseline results. In Anais do VI Brazilian Workshop on Social
Network Analysis and Mining . SBC, 2017.
7/1012. J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-t raining of
deep bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805 , 2018.
13. F. Feng, Y. Yang, D. Cer, N. Arivazhagan, and W. Wang. Langu age-agnostic
BERT sentence embedding. arXiv preprint arXiv:2007.01852 , 2020.
14. S. Gaikwad, T. Ranasinghe, M. Zampieri, and C. M. Homan. Cross -lingual offen-
sive language identification for low resource languages: The case of Marathi. In
Proceedings of RANLP , 2021.
15. S. Hassan, Y. Samih, H. Mubarak, A. Abdelali, A. Rashed, and S. A. Chowdhury.
Alt submission for osact shared task on offensive language detect ion. InProceed-
ings of the 4th Workshop on Open-Source Arabic Corpora and Pr ocessing Tools,
with a Shared Task on Offensive Language Detection , pages 61–65, 2020.
16. L. Komalova, A. Glazkova, D. Morozov, R. Epifanov, L. Motovs kikh, and E. May-
orova. Automated classification of potentially insulting speech acts on social net-
work sites. In International Conference on Digital Transformation and Gl obal
Society . Springer, 2021.
17. V. Kumar, A. Choudhary, and E. Cho. Data augmentation using pre-trained
transformer models. In Proceedings of the 2nd Workshop on Life-long Learning
for Spoken Language Systems , pages 18–26, 2020.
18. M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Lev y, V. Stoy-
anov, and L. Zettlemoyer. Bart: Denoising sequence-to-sequen ce pre-training for
natural language generation, translation, and comprehension. I nProceedings of
the 58th Annual Meeting of the Association for Computationa l Linguistics , pages
7871–7880, 2020.
19. P. Liu, W. Li, and L. Zou. NULI at SemEval-2019 task 6: Transfe r learning for
offensive language detection using bidirectional transformers. I nProceedings of
the 13th international workshop on semantic evaluation , pages 87–91, 2019.
20. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Le wis, L. Zettle-
moyer, and V. Stoyanov. Roberta: A robustly optimized bert pret raining ap-
proach.arXiv preprint arXiv:1907.11692 , 2019.
21. I. Loshchilov and F. Hutter. Decoupled weight decay regulariza tion. InInterna-
tional Conference on Learning Representations , 2018.
22. T. Mandl, S. Modha, A. Kumar M, and B. R. Chakravarthi. Overv iew of the
HASOC track at FIRE 2020: Hate speech and offensive language ide ntification
in Tamil, Malayalam, Hindi, English and German. In Forum for Information
Retrieval Evaluation , pages 29–32, 2020.
23. T. Mandl, S. Modha, P. Majumder, D. Patel, M. Dave, C. Mandlia, and A. Patel.
Overview of the HASOC track at FIRE 2019: Hate speech and offen sive content
identification in Indo-European languages. In Proceedings of the 11th forum for
information retrieval evaluation , pages 14–17, 2019.
24. T. Mandl, S. Modha, G. K. Shahi, H. Madhu, S. Satapara, P. Maj umder,
J. Sch¨ afer, T. Ranasinghe, M. Zampieri, D. Nandini, and A. K. Jaisw al. Overview
of the HASOC subtrack at FIRE 2021: Hate speech and offensive c ontent identi-
fication in English and Indo-Aryan languages. In Working Notes of FIRE 2021 -
Forum for Information Retrieval Evaluation . CEUR, December 2021.
8/1025. S. Mishra, S. Prasad, and S. Mishra. Multilingual joint fine-tunin g of transformer
models for identifying trolling, aggression and cyberbullying at TRAC 2 020. In
Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying ,
pages 120–125, 2020.
26. A. K. Mishraa, S. Saumyab, and A. Kumara. IIIT DWD@ HASOC 2020: Iden-
tifying offensive content in Indo-European languages. 2020.
27. S. Modha, T. Mandl, G. K. Shahi, H. Madhu, S. Satapara, T. Ran asinghe, and
M. Zampieri. Overview of the HASOC subtrack at FIRE 2021: Hate sp eech
and offensive content identification in English and Indo-Aryan langu ages and
conversational hate speech. In FIRE 2021: Forum for Information Retrieval
Evaluation, Virtual Event, 13th-17th December 2021 . ACM, December 2021.
28. A. Montejo-R´ aez, S. M. Jim´ enez-Zafra, M. A. Garc´ ıa-Cum breras, and M. C. D´ ıaz-
Galiano. SINAI-DL at SemEval-2019 task 5: Recurrent networks a nd data aug-
mentation by paraphrasing. In Proceedings of the 13th International Workshop
on Semantic Evaluation , pages 480–483, 2019.
29. H. Mubarak, K. Darwish, W. Magdy, T. Elsayed, and H. Al-Khalifa . Overview
of OSACT4 Arabic offensive language detection shared task. In Proceedings of
the 4th Workshop on Open-Source Arabic Corpora and Processi ng Tools, with a
Shared Task on Offensive Language Detection , pages 48–52, 2020.
30. D. Q. Nguyen, T. Vu, and A. T. Nguyen. BERTweet: A pre-train ed language
model for English tweets. In Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing: System Demonstrat ions, pages 9–14,
2020.
31. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Ch anan, T. Killeen,
Z. Lin, N. Gimelshein, L. Antiga, et al. Pytorch: An imperative style, h igh-
performance deep learning library. Advances in neural information processing
systems , 32:8026–8037, 2019.
32. J. Pavlopoulos, J. Sorensen, L. Laugier, and I. Androutsopo ulos. SemEval-2021
task 5: Toxic spans detection. In Proceedings of the 15th International Workshop
on Semantic Evaluation (SemEval-2021) , pages 59–69, 2021.
33. M. Ptaszynski, A. Pieciukiewicz, and P. Dyba/suppress la. Results of the P olEval 2019
shared task 6: First dataset and open shared task for automatic cyberbullying
detection in Polish Twitter. 2019.
34. B. Ray and A. Garain. JU at HASOC 2020: Deep learning with RoBER Ta
and random forest for hate speech and offensive content identif ication in Indo-
European languages. In FIRE (Working Notes) , pages 168–174, 2020.
35. A. Ribeiro and N. Silva. Inf-hateval at semeval-2019 task 5: Co nvolutional neural
networks for hate speech detection against women and immigrants on Twitter. In
Proceedings of the 13th International Workshop on Semantic Evaluation , pages
420–425, 2019.
36. J. Risch, A. Stoll, M. Ziegele, and R. Krestel. hpiDEDIS at GermEv al 2019:
Offensive language identification using a German BERT model. In KONVENS ,
2019.
9/1037. M. Sanguinetti, F. Poletto, C. Bosco, V. Patti, and M. Stranisc i. An Italian
Twitter corpus of hate speech against immigrants. In Proceedings of the Eleventh
International Conference on Language Resources and Evalua tion (LREC 2018) ,
2018.
38. F. Schmid, J. Thielemann, A. Mantwill, J. Xi, D. Labudde, and M. Sp ranger.
Fosil-offensive language classification of German tweets combining S VMs and
deep learning techniques. In KONVENS , 2019.
39. P. Singh and P. Bhattacharyya. CFILT IIT Bombay at HASOC 20 20: Joint
multitask learning of multilingual hate speech and offensive content detection
system. In FIRE (Working Notes) , pages 325–330, 2020.
40. J. M. Struß, M. Siegel, J. Ruppenhofer, M. Wiegand, M. Klenner , et al. Overview
of GermEval task 2, 2019 shared task on the identification of offe nsive language.
2019.
41. X. Tang, X. Shen, Y. Wang, and Y. Yang. Categorizing offensiv e language in so-
cial networks: A Chinese corpus, systems and an explanation tool. InChina Na-
tional Conference on Chinese Computational Linguistics , pages 300–315. Springer,
2020.
42. M. Taul´ e, A. Ariza, M. Nofre, E. Amig´ o, and P. Rosso. Overv iew of detoxis at
IberLEF 2021: Detection of toxicity in comments in Spanish. Procesamiento del
Lenguaje Natural , 67:209–221, 2021.
43. M. Wiegand, M. Siegel, and J. Ruppenhofer. Overview of the ger meval 2018
shared task on the identification of offensive language. In 14th Conference on
Natural Language Processing KONVENS 2018 , 2018.
44. T. Wolf, J. Chaumond, L. Debut, V. Sanh, C. Delangue, A. Moi, P . Cistac,
M. Funtowicz, J. Davison, S. Shleifer, et al. Transformers: State -of-the-art natural
language processing. In Proceedings of the 2020 Conference on Empirical Methods
in Natural Language Processing: System Demonstrations , pages 38–45, 2020.
45. M. Zampieri, S. Malmasi, P. Nakov, S. Rosenthal, N. Farra, and R . Kumar.
Semeval-2019 task 6: Identifying and categorizing offensive langu age in social
media (OffensEval). In Proceedings of the 13th International Workshop on Se-
mantic Evaluation , pages 75–86, 2019.
46. Y. Zhu, R. Kiros, R. Zemel, R. Salakhutdinov, R. Urtasun, A. To rralba, and
S. Fidler. Aligning books and movies: Towards story-like visual explan ations by
watching movies and reading books. In Proceedings of the IEEE international
conference on computer vision , pages 19–27, 2015.
10/10