1 Introduction

Nowadays social media has become an important part of people’s lives, where everyone can read news, communicate with friends and share their opinions. These advances also bring new challenges and risks associated with new technologies. With the increasing influence of social networks and online discussions, there is a problem of aggressive language in users’ messages increasing dramatically. It is no secret that on the Internet, in the territory where there is no real contact, people feel more free and allow themselves to use statements that can bring moral harm to other users.

Aggressive, offensive, and hate speech in social media can relate to various social aspects: the problems of immigration, race, gender, weight, and religion of other people. Hate speech messages often contain insults, but situations also rise when the message does not contain swear words and profanity, but the meaning of the message is offensive and humiliating to a group of people or a particular person.

However, as the intention is normally the same, we treat offensive and hate speech as synonyms.

Based on these facts, the problem of hate speech recognition in social networks in order to protect other users from such messages is very important one.

Currently, more and more attention is paid to solving this problem and experts from different fields (including computer scientists, psychologists and linguists) are making efforts to create approaches and technologies that are able to recognize offensive messages in social networks with a maximum accuracy and in the shortest period.

In this article, we present our approach to hate speech recognition based on the classical models of machine learning and the results we achieved by the experiments with two Twitter datasets. One of the experiment had the aim to identify the offensive language in messages, and another one was aimed to recognize the target of the offensive messages: an individual person or a group of people. This paper is organized as follows.

Some relevant related works in the area of hate speech recognition are described in Section 2. Section 3 describes the datasets we used for our experiments. Section 4 presents the preprocessing stage and the methodology for the models we used. In Section 5, the results are described and analyzed. In Section 6, we summarize our work and plan some steps for the future research.

2 Related Work

In order to classify a message as hate speech, it is necessary to understand clearly how to distinguish an offensive message from a non-offensive one. It seems most obvious to call the text offensive if it includes swearing, but this is not always correct, since the author may not have the aim to offend users, and he may have used profanity to express emotion or an another reason.

In [¹¹], the authors created a dataset to work with aggressive language identification in the context of two types of abuse: sexism and racism.

They collected and annotated 16,914 tweets over the course of two months: 3,383 of which (written by 613 unique users) were labeled as sexists, - 1,972 tweets by 9 users were racist and 11,559 tweets by 614 users were neither sexist nor racist.

The authors indicated some rules that help draw a clear line between offensive and no offensive tweets. The most important rules which help to indicate an offensive message are (i) a sexist or racial slur were used in a message, (ii) a minority were attacked or there were negatively stereotypes a minority, or there was the aim to seek to silence a minority in a message, and (iii) hate speech or violent crime were promoted but did not directly use or xenophobia or sexism were defended in a message.

Additionally, the authors extracted some features using the meta-data of messages. They highlighted a gender of users while looking at names in profiles and found that about half of all messages were written by men, 2.26 percent - by women and 47.64 percent - undefined users.

Using the time zones, which were marked in tweets authors, created geographic distribution feature and they calculated the length of tweets.

The authors created the model based on logistic regression, which showed best result in case when features of gender and geographic location were counted (F1-score 0.7362). It is interesting that in case when gender, geographic location and length of tweet were taking into account the result were not so good (F1-score 0.7347).

It was shown in [¹⁴] that detecting hateful content using linguistic characteristics is quite difficult because of the absence of unique discriminative features. Seven public Twitter datasets were analyzed, five of which included three different types of tweets labels: sexism, racism and neither, and another two datasets were contacted of tweets divided by hate and non-hate classes.

A uniqueness score measurement was created, which indicated the number of unique words corresponding to each class (i.e. not occurring in other classes) were included in the message. This measure is found as the intersection of words in a message with unique words from this class divided by the number of all words in that message.

The meaning of this measure takes a value from 0 to 1. The authors checked each dataset using this measure. They found that about half of all tweets did not contain the unique words of their classed or contained very few this words (the meaning of the measure less than 0.5) and it means that there are no discriminative features which could indicate hate speech because of the fact that people can write an offensive messages using different words.

In [²], the authors investigated the relationships between the user who posted an offensive message and the user who was the target of this message. The authors of the article analyzed two different groups of slurs: Sexist Derogatory Slurs (e.g., bitch) and Sexist Objectifying Slurs (e.g., hot chick) in case of different relationships (e.g., friends, partners, work-related context) and the gender of the user (man or woman) in the Italian language. Sexist Derogatory Slurs was the class with the aim of denigrating women in the context of stereotypes, sexual looseness and promiscuity, while Sexist Objectifying Slurs was the class of words, which reduced women to being objects of male sexual interests.

The authors showed that people tend to evaluate Sexist Derogatory Slurs as being more offensive when compared with Sexist Objectifying Slurs. Slurs directed at women were judged to be more offensive those directed at men. Correlation analyses was performed that the high level of frequency of use is connected with the high level of social acceptability. The results of evaluation in the specific social settings shows that the slurs have more social acceptability in a context of affected relationships than in an equal work-related context (a conversation between colleagues) and a higher work related context (a conversation between a superior and a subordinate). It is interesting that the Sexist Objectifying Slurs had the lowest social acceptability in the work-relation situation with unequal positions (supervisor- subordinate).

It should be noted that although it was not possible to indicate special linguistic characteristics for aggressive speech messages, the authors of [¹] conducted a study, which revealed some grammatical and lexical features, which are typical for aggressive posts. The article demonstrated the importance of functional linguistic variation in a corpus of racist and sexist Tweets. The authors analyzed 628 sexist tweets and 858 racist tweets and tried to establish the role of lexical and grammatical features using MDA (multidimensional analysis) in three different dimensions (interactive, antagonistic and attitudinal).

The first interactive dimension showed how interactive or informative the message was, the antagonistic dimension presented the attitude of the user to the reader: whether he agrees with them or not, and the attitudinal dimension represented the degree of attitudinal judgment exhibited by a tweet. Finally, the authors compared the difference for racist tweets and sexist ones along each of three dimensions.

The results showed that the sexist messages were more interactive and more attitudinal then racist ones, but had the same measure in the antagonistic dimension, and the most popular linguistic feature in offensive language were question marks (there were a lot of questions in this messages) and question DO (when a sentence begins with the word do).

When we speak about available approaches for models created for aggressive speech recognition it is important to note classical machine learning approaches. Models based on Support Vector Machine, Naive Bayes and Logistic Regression approaches and some ensembles of this classical models achieved quite good results [⁵, ⁹] in shared tasks AMI@IBEREVAL-2018 [³] and EVALITA-2018 [⁴]. The aim in these cases was to detect misogynistic behavior in English and Italian tweets, where the task was to multiclassify the messages according to the type of offensive language.

The classes for the classification included Stereotype and Objectification (a widely held but fixed and oversimplified image or idea of a woman), Dominance (an assertion the superiority of men over women), Derailing (a justify woman abuse), Sexual Harassment and Threats of Violence (a sexual advance, an intention to physically assert power over women through treats of violence) and Discredit (a slurring of women with no other larger intention) groups of misogynistic messages.

It should also be noted that the models, which are based on the approach of deep neural networks, have great prospects in the problems of identifying hate speech in messages. In [⁸], three different models were presented based on neural networks, which shows quite good results in hate speech identification.

These were a CNN-based model, which is the character-level convolutional network, a convolutional network where a sentence was segmented into words on input and finally a model, which combined the previous two with two inputs: characters and words. The idea behind creating this model was an observation that offensive tweets often contain either purposely or mistakenly misspelled words. All three models had 3 layers.

The classification was for three different classes: racism, sexism and none and in this case the best results was shown by the hybrid model of convolutional networks with 0.827 F1-score, while the best result by classical model (Logistic Regression) was 0.814.

In addition, authors created the combination of neural network based model and classic machine learning approach (Logistic Regression) and shown very good results for the multi-classification and this result looks encouraging.

3 Datasets

For our experiments, we used two datasets consisting of messages from Twitter. The first dataset was available in the framework of multilingual detection of hate speech against immigrants and women in Twitter (HatEval-2019 -one of the tasks in the frames of SemEval-2019 challenge^fn). The task consisted of two subtasks, one of which was the binary classification between offensive and non-offensive messages in case of hate speech detection against immigrants and women, and the second task proposed to make an aggressive/non-aggressive and individual target/a group target classification on the offensive messages.

Although were there two datasets (one in English, the other in Spanish) we used only the English one for our experiments. The training dataset included 10,000 tweets, of which 4,210 were labeled as hate speech tweets and 5790 not, and 1,560 tweets had individual target and 2,650 tweets had group target. The testing dataset contained of 3,000 tweets. Some examples of different types of messages are presented in Table 1.

The second dataset we chose was from the Identifying and Categorizing Offensive Language in Social Media shared task (OffensEval-2019 -one of the tasks in the frames of SemEval-2019 challenge).

The challenge had 3 different subtasks: the first two were binary classification for offensive language identification (is message offensive or not) and automatic categorization of offense types (is offensive message insult a person or a group of people or it is non-targeted profanity and swearing), and the last subtask had the aim to make the target classification for three groups: individual, group or other (in this case the target of the offensive post did not belong to any of the previous categories, e.g., a situation, an event, or an issue) target.

The training dataset we used consisted of 13,200 tweets, and 4,400 of them were offensive in the ratio: 2,407 tweets - individual target, 1,074 tweets - group target. Some examples from different types of tweets are presented in Table 2.

There were 860 tweets for testing in case of hate speech identification and 213 from them tweets for the target classification. It should also be noted that in the OffensEval dataset all references were anonymized and replaced with the string URL.

Although the datasets are not well balanced, this distribution can be perceived as the present state of affairs and the frequency of occurrence of such messages in reality.

For the HatEval dataset and OffensEval dataset we made experiments with the aim of hate speech identification and the target of hate speech recognition on the training data (10,000 tweets and 13,200 tweets accordingly) and after we tested our created model on the testing datasets (3,000 tweets and 860 tweets accordingly).

4 The Model

In this section, we explain two main steps of our experiments: preprocessing and modeling. Each stage of model creation was important for us because both preprocessing and modeling make a great contribution to the quality of the constructed classifiers and to the results of the research.

5 Conclusion

This article describes a possible approach to offensive language recognition. It involves a preprocessing step and a creation of models, which allows us to obtain quite good results in a solving the problem of a small number of messages classification with the aim to identify hate speech. The results were good for different types of classification: both in the case of hate speech identification, and in the target of hate speech identification. The results achieved by the proposed model are competitive in comparison with the best-published results, which were achieved in the processing of the same datasets.

The features which were used at the preprocessing stage indicate that the date preprocessing is very important and it is necessary to pay attention not only to the process of the model creating, but also to the analysis of the messages in the dataset and it is vital to try to identify certain patterns that occur in them.

Also, the use of messages that were referenced in the data improved the results of our model, so we can say that it makes sense to develop this area of research in the future and we should try to take into account not only the text of the message, which was referenced, but also other information, including:

In addition, in the future we plan to use additional sources of information, such as dictionaries of words of different tonality and lists of a swearing/harassment vocabulary for working with the identification of aggressive language.

Despite the fact that the message may contain, for example, swearing words, but be not offensive, it is intuitively clear that the combination of insult in the message - the message is offensive in nature is more common, so this direction in the continuation of the research seems promising to us.

In addition, as described above, the use of models based on deep machine learning, in particular CNN, and the combination of such models with models based on the classical machine learning approach, increases the accuracy of the research. In the future, we plan a developing of our ensemble model in this direction in order to improve the accuracy of the classification.