he paper presents a method of hate speech spreaders recognition developed for the task of “Profiling Hate Speech Spreaders on Twitter” at the PAN@CLEF Conference 2021. Hate speech is an increasing problem nowadays. Due to the spread of the internet and the rise of social media, people with hostile views towards certain groups post hateful messages on social media resources. In this paper, we present a model to detect hate speech spreaders based on their Twitter posts. We aggregate contextualized embeddings of single tweets to form a vector representation for every user and employ classification methods to find users spreading hate speech. We analyze diferent embedding models based on BERT architecture for the problem. The submitted model achieves 67% in terms of accuracy for the English part of the dataset and 83% for the Spanish part of the dataset.
Hate speech detection in social networks is a significant social issue nowadays. A large number
of works is devoted to hate speech detection in the social media domain [
Hate Speech Spreaders profiling task considers the problem of hate speech detection in social
networks [
The main challenge of the current task is that the analysed object is not a single tweet, but a
collection of messages posted by a user. Even if the developed method detects a content without
hate speech in some of the user’s posts, the same user still can be classified as a potential hate
speech spreader. In this way, hate speech detection is similar to the previous fake news detection
task [
Recent studies show that approaches, based on embedding generation and their further
classification can be quite efective in hate speech detection. For example, in [
The following section describes the proposed method of hate speech spreader detection.
Similar to the previous year’s task [13], we consider the current task as a classification problem, where the classified object is a collection of tweets.
There is a given dataset D = (, ): = {1 , . . . , },
∈ W+, ∈ {1, . . . , }, ∈ {0, 1}, where W+ corresponds to all the possible strings written in the given language. The label = 1 corresponds to users that are likely to spread hate speech, = 0 corresponds to ordinary users.
Formally, the task is to find the binary classifier that minimizes an empirical risk on the dataset D: = arg min ∈F
∑︁ [ () ̸= ], ,∈D where F is a set of all considered classification models.
For the proposed problem solution we employ a deep learning-based approach: each tweet is vectorized using a deep learning model. After that, we aggregate the obtained vectors and use the resulting averaged vector as a feature set for the classifier.
The proposed method is illustrated in Figure 1. The method consists of 3 main components: 1. BERT model that extracts embedding vector e from every tweet . 2. Aggregation operation that averages the information across all the tweets of the author: e =
1 ∑︁ e.
=1 3. The resulting classifier that uses averaged tweet vector e as an input feature set.
To validate our models we conducted a computational experiment. Since we use a standard BERT tokenizer, we did not use any special preprocessing for our method. For all the analyzed models we used -fold cross-validation with = 10. The results are presented in Table 1. The test accuracy was evaluated using TIRA environment [14].
Inspired by the author profiling task from the previous year [
Since the dataset contains texts in two languages, we initially decided to use a pretrained multilingual BERT model [11] for the vectorization. We analyzed its performance with two classifiers: support vector machine [ 16] and logistic regression. For each classification algorithm, the optimal parameters for cross-validation were selected using GridSearch. Based on the results in Table 1, it is clear that this approach works well for Spanish, but shows poor performance on English data. We decided to use Multilingual BERT for the processing of Spanish and use a diferent model for English. In order to improve our quality in the English part of the dataset we analyzed the following models: • LaBSE [17], a multilingual model with BERT architecture. The main feature of this model is that all the languages supported by LaBSE share a common hidden space that can potentially improve our model generalization. • BERT-Base, the pretrained model that was trained only on the English texts; • BERT-base tuned on the external corpus of the tweets. The main idea of this approach is to fine-tune the lexical and semantic properties of the model on the tweet corpus in order to make the model more receptive to the specific Twitter lexicon.
Note that since the training dataset has rather small information for BERT model tuning, we did not consider the idea of fine-tuning straightforwardly on the current training dataset. Therefore we only tuned the model on the external tweet collection in an unsupervised manner it was originally trained. Some tokens in the sentences were masked, and the model tried to predict them. The predicted tokens were fed into a softmax layer to get the output words. Such procedure does not require additional labeling of external data and allows BERT to focus more on the Twitter posts than the English and Spanish texts in general.
LaBSE vectorization. Although the LaBSE model shares the common hidden space for all the languages including Spanish and English, we did not find any quality improvement in comparison to the basic architecture and other methods.
BERT-base. The pretrained BERT-base model was trained only on English texts, so we suggest that using it as a vectorizer for English tweets might improve accuracy. Our assumption that BERT-base single-language model shows better performance than the multilingual model on an English-only task is confirmed by cross-validation results. When using this model, we did not make any changes to the architecture.
BERT-based model fine-tuned on the external tweet corpus. Given that BERT model base was trained on a large collection of English texts from Wikipedia , we assume that it needs ifne-tuning on the data related to our problem, since tweets often use specific slang, syntax, emoji etc. In order to improve the quality of vector representation we took a large corpus of tweets [12], which contains 1,600,000 tweets in English, and fine-tuned the model on this dataset in masked language model task regime. For all 1,600,000 tweets from the additional dataset, we replaced links, hashtags and user mentions with tokens from the training dataset (#URL#, #HASHTAG#, #USER# respectively). We also added these tokens to the BERT tokenizer dictionary. The use of additional training improved the accuracy.
As we can see, the resulting model shows a competitive performance in comparison to the baseline model. The cross-validation accuracy has increased from 0.68 to 0.72 on English part of the dataset and from 0.79 to 0.8 on its Spanish part. We got a rather significant improvement on English data by tuning the model in an unsupervised manner on the external corpus of tweets. We also observe a slight increase in performance on Spanish tweets.
The submitted English model achieves 67% in terms of accuracy on the test set, while the Spanish model shows the accuracy 83%.
For further analysis, we made a t-SNE transformation [18] for each averaged user vector e in two dimensional space and visualized them according to their class. The result is shown in Figure 2,3 for English and Spanish, respectively.
As we can see, the obtained projection poorly separates averaged user vectors, which indicates that this method needs to be potentially improved.
We also built an LDA topic model [19] for the English part of the dataset in order to analyze the topic distributions across the tweets. For the topic model, we used 10 topics. The resulting t-SNE projection is shown in Figure 3. We analyzed the words mostly corresponding to each topic. Although we could not find a total interpretability of the obtained topics we noticed that some of them correspond to the US politics (“Biden”, “Trump”, “President”, “Wall”, “American” in the top words of the topic), US elections ( “Biden”, “Trump”, “Election”, “Voters”, “Democrats”, “Republicans”, “Court” in the top words of the topic), COVID-19 (“Covid”, “19”, “Virus” in the top words of the topic). This makes us believe that clustering tweets using some topic modeling approach and employing information about topic distribution across the user tweets for the decision about hate speech spreading.
The paper describes an approach to the problem of hate speech spreaders detection. We proposed a model to detect hate speech spreaders using contextualized embeddings of single tweets. We aggregate the obtained vectors and use their average vector as a feature set for the further classification. We also provide an analysis of diferent vectorization models based on the BERT architecture. The resulting model shows an accuracy of about 67% for the English test dataset and 83% for the Spanish test dataset. The future work includes a detailed analysis of topic distribution across the tweets and usage of this information in the final decision rule.
This research is funded by RFBR, grant 19-29-14100. tional Conference on Pattern Recognition Applications and Methods - ICPRAM„ INSTICC, SciTePress, 2020, pp. 453–460. doi:10.5220/0008954004530460. [11] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018). [12] A. Go, R. Bhayani, L. Huang, Twitter sentiment classification using distant supervision,
CS224N project report, Stanford 1 (2009) 2009. [13] O. Bakhteev, A. Ogaltsov, P. Ostroukhov, Fake news spreader detection using neural tweet aggregation, in: CLEF, 2020. [14] M. Potthast, T. Gollub, M. Wiegmann, B. Stein, TIRA Integrated Research Architecture, in: N. Ferro, C. Peters (Eds.), Information Retrieval Evaluation in a Changing World, The Information Retrieval Series, Springer, Berlin Heidelberg New York, 2019. doi:10.1007/ 978-3-030-22948-1\_5. [15] C. M. Bishop, Pattern recognition and machine learning, springer, 2006. [16] C. Cortes, V. Vapnik, Support vector machine, Machine learning 20 (1995) 273–297. [17] F. Feng, Y. Yang, D. Cer, N. Arivazhagan, W. Wang, Language-agnostic bert sentence embedding, arXiv preprint arXiv:2007.01852 (2020). [18] L. Van der Maaten, G. Hinton, Visualizing data using t-sne., Journal of machine learning research 9 (2008). [19] K. Vorontsov, O. Frei, M. Apishev, P. Romov, M. Dudarenko, Bigartm: Open source library for regularized multimodal topic modeling of large collections, in: International Conference on Analysis of Images, Social Networks and Texts, Springer, 2015, pp. 370–381.