In this paper we present HATECHECKER, a tool for the automatic detection of hater users in online social networks which has been developed within the activities of ”Contro L'Odio” research project. In a nutshell, our tool implements a methodology based on three steps: (i) all the Tweets posted by a target user are gathered and processed. (ii) sentiment analysis techniques are exploited to automatically label intolerant Tweets as hate speeches. (iii) a lexicon is used to classify hate speeches against a set of specific categories that can describe the target user (e.g., racist, homophobic, antisemitic, etc.). Finally, the output of the tool, that is to say, a set of labels describing (if any) the intolerant traits of the target user, are shown through an interactive user interface and exposed through a REST web service for the integration in third-party applications.
In the experimental evaluation we crawled and annotated a set of 200 Twitter profiles and we investigated to what extent our tool is able to correctly identify hater users. The results confirmed the validity of our methodology and paved the way for several future research directions.
Copyright 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). According to a recent study1, 58% of the Italian population regularly uses online social networks as Twitter, Facebook, Instagram and LinkedIn.
Such a huge diffusion of these platforms is providing the users with many new opportunities and services, just think that almost everyone now uses social media to get information, discuss, express opinions and stay in touch with friends. Unfortunately, due to the lack of control and the absence of a clear management of the concept of identity of the users, social networks have become the perfect place to spread hate against minorities and people having different cultures, values and opinions.
As pointed out by several works
Similar research attempts were also proposed for the Italian language. One of the most popular initiative is the Italian HateMap project (Musto 1https://wearesocial.com/it/blog/2018/01/global-digitalreport-2018
2https://www.cnbc.com/2019/02/04/facebook-googleand-twitter-are-getting-faster-at-removing-hate-speechonline-eu-finds–.html
et al., 2016), a research project that exploits
semantic analysis and opinion mining to identify
the most-at-risk areas of the Italian country, that
is to say, the areas where the users more
frequently publish hate speeches. The interest of the
research community for the topic was confirmed
by the recent work by Bosco et al.
In order to continue the investigation in this research line ACMOS3, a no-profit association based in Torino, recently launched ”Contro l’Odio4”, a joint research project with the University of Bari, University of Torino and several local associations. The project aims to develop tools and methodologies to monitor (and hopefully tackle) online hate speeches and intolerant behaviors.
One of the outcomes of the research is HATECHECKER, a tool that aims to automatically identify hater users on Twitter by exploiting sentiment analysis and natural language processing techniques. The distinguishing aspect of the tool with respect to the work we have previously introduced is the focus of the tool itself. Indeed, differently from most of the literature, that focused on the analysis of single Tweets, HATECHECKER aims to analyze the users as a whole, and to identify hater users rather than hate speeches. Clearly, both the tasks are in close correlation, since techniques to detect hate speeches can be used to detect hater users as well.
However, through this work we want to move the focus on the latter since, up to our knowledge, this a poorly investigated research direction. Just think that no datasets of hater users is currently publicly available.
To sum up, the contributions of the work can be summarized as follows:
We present a workflow that allows to detect hater users in online social networks; 3http://www.acmos.net 4http://www.controlodio.it We evaluate several configurations (on varying of lexicons and sentiment analysis algorithms) of the pipeline and we identified the most effective one to tackle our specific task; We share the first publicly available dataset for automatic detection of hater users on Twitter.
In the following, we will first describe the methodology we designed to implement our system, then we will discuss the effectiveness of the approach by analyzing the results we obtained on a (publicly available) dataset of 200 Twitter users. 2
The workflow carried out by the HATECHECKER tool is reported in Figure 1.
Generally speaking, the pipeline consists of four different modules, that is to say, a SOCIAL DATA EXTRACTOR, a SENTIMENT ANALYZER, a PROFILE CLASSIFIER anda SOCIAL NETWORK PROCESSOR. All these components use a NoSQL database to store the information they hold and expose the output returned by the tool through a REST interface as well as through a Web Application. In the following, a description of the single modules that compose the workflow is provided.
The whole pipeline implemented in the HATECHECKER tool needs some textual content posted by the target user to label the user as a hater or not. In absence of textual content, it is not possible provide such a classification. To this end, the first and mandatory step carried out by the tool is the extraction of the Tweets posted by the user we want to analyze. In this case, we used the official Twitter APIs to gather the available Tweets and to forward it to the next modules of the workflow.
Given that the real-time execution of the workflow is one of the constraints of the project, we limited the extraction to the 200 most recent Tweets posted by the user. This is a reasonable choice, since we aim to detect users who recently showed an intolerant behavior, rather than users who posted hate speeches one or two years ago.
Once the Tweets have been collected, it is necessary to provide the tool with the ability to go through the content posted by the target and to automatically identify the hate speeches.
To this end, the SENTIMENT ANALYZER
modules exploits Sentiment Analysis techniques
SentiPolC:
A complete overview of the algorithms is out of the scope of this paper and we suggest to go through the references for a thorough discussion. For the sake of simplicity, we can state that the output of both the algorithms is a binary classification of each Tweet posted by the target user as negative (that is to say, conveying hate speeches) or positive. Such an output is then passed to the PROFILE CLASSIFIER module whose goal is to assign a more precise label to the user, on the ground of the nature of the hate speeches she posted (if any). 2.3
In such a specific setting, the simple exploitation of sentiment analysis techniques that provide a rough binary classification of the single Tweets (conveying/not conveying hate) is not enough. Indeed, the answers to two fundamental questions are still lacking:
How can we label the user as hater or nonhater on the ground of the Tweets she posted? How can we return a more fine-grained classification of the user (e.g., racist, homofobe, etc.) on the ground of the Tweets she posted? Both these issues are tackled by the PROFILE CLASSIFIER module. As for the first question, a very simple strategy based on thresholding is implemented. In particular, we defined a parameter , and whether the user posted a number of Tweets labeled as hate speeches higher than , the user herself is labeled as an hater. Of course, several values for the parameter can be taken into account to run the tool.
As for the second question, we used a lexiconbased approach to provide a fine-grained classification of users’ profiles. The intuition behind our methodology is that for each category a specific lexicon can be defined, and whether a Tweet posted by the user contains one of the terms in the lexicon, the user is labeled with the name of the category.
Formally, let C = fc1; c2 : : : cng be the set of the categories (e.g., racism, homophobia, sexism, etc.) and let VCi = ft1; t2 : : : tmg be the vocabulary of the category Ci. Given a Tweet T written by a user u, if one of the terms in VCi is contained in T , the user u is labeled with the category Ci.
To define the lexicon for each category, we
relied on the research results of the Italian Hate Map
In order to (hopefully) enrich and improve the
lexicon used in the Italian Hate Map project, we
exploited Hurtlex, a multilingual lexicon of hate
words
Obviously, in the experimental session the effectiveness of the tool on varying of different lexicons and on different configuration of the workflow will be evaluated. 2.4
At the end of the previous step, the target user is labeled with a set of categories describing the facets of her intolerant behavior.
However, one of the goals of the project was also to investigate the role and the impact of the social network of the users in the dynamics of online haters. Accordingly, the SOCIAL NETWORK PROCESSOR gathers the entire social network of the target user and runs again (in background, of course) the whole pipeline on all the following and followers of the target user, in order to detect whether other people in the social network of the target user can be labeled as haters as well. The goal of this step is to further enhance the comprehension of network dynamics and to understand whether online haters tend to follow and be followed by other haters.
Unfortunately, due to space reasons, the discussion of this part of the workflow is out of the scope of this paper and is left for future discussions. 2.5
Finally, the output of the platform is made available to third-party services and to the user itself. In the first case, a REST web service makes available the output of the tool (that it so say, the hate categories and the number of haters in her own social network), while in the latter the same data are shown through an interactive user interface.
A screenshot of the working prototype of the platform is reported in Figure 2. As shown in the Figure, a user interacting with the platform can query the system by interactively providing a Twitter user name. In a few seconds, the interface shows a report of the target user containing a set of emojis reporting the behavior of the user for each of the categories we analyzed, a snapshot of her own Tweets labeled as hate speeches and some information about the percentage of hater profiles that are in the social network of the target user.
It is worth to note that such a web application is very useful for both monitoring tasks (e.g., to verify whether a third-party account is an online hater) as well as for Quantified Self scenarios (Swan, 2013), that is to say, to improve the selfawareness and the self-consciousness of the user towards the dynamics of her social network. Our intuition is that a user who is aware of not being an hater, can use the system to identify (if any) the haters that are still in her own social network, and maybe decide to unfollow them. 3
The goal of the experimental session was to evaluate the effectiveness of the tool on varying of different configurations of the pipeline.
To this end, due to the lack of a dataset of hater profiles, we manually crawled and annotated a set of 200 Twitter users, which we made available5 for the sake of reproducibility and to foster the research in the area.
In particular, we compared four different strategies to run our tool, on varying on two different parameters, such as the lexicon and the sentiment analysis algorithm. In particular, we exploited the following combinations of parameters:
HanSEL, as previously explained and Lexicons: HateMap lexicon and complete lexicon (HateMap+Hurtlex)
As for the parameters, the threshold was set equal to 3 and both the sentiment analysis algorithms were run with the standard parameters introduced in the original papers. To evaluate the effectiveness of the approaches, we calculate the number of correctly classified user profiles over the total of hater users in the dataset.
5https://tinyurl.com/uniba-haters-dataset
The results of the experiments are reported in Table 1. In general, we can state that our approach to automatically detect hater users in online social network provided us with encouraging results, since more a percentage between 78% and 97% of the online haters were correctly detected by the algorithm, regardless of the specific category.
It is worth to note that the worse results (both of them are beyond 70%, through) were obtained for racism and xenophobia, that is to say, two facets characterized by a lexicon that quickly evolves and often adopts terms that are not conventional and not necessarily conveying hate (e.g.,, expressions as ’Aiutiamoli a casa loro’ or terms as ’clandestini’). However, even for these categories the results we obtained were encouraging.
Conversely, results were particularly outstanding for facets such as anti-semitism and homophobia, that have a quite fixed lexicon of terms that can be used to hurt or offend such minorities.
As for the different configurations, we noted that HANSEL tended to obtain better results than SENTIPOLC. This is a quite expected outcome, since it exploits more novel and effective methods as those based on word embeddings and deep learning techniques. Moreover, we can state that the results can be further improved since no particular tuning of the parameters was carried out in this work.
As for the lexicons, the extension of the original Italian Hate Map lexicons with new terms led to an improvement of the results for all the facets (except for homophobia) for at least one of the comparisons. Such improvement are often tiny, but this is an expected outcome since just a few terms coming from Hurtlex were added. However, even these preliminary results provided us with encouraging findings, since they showed that the integration and the extension of sensible terms with the information coming from recently developed lexical resources can lead to a further improvement of the accuracy of the system. 4
In this work we have presented HATECHECKER, a tool that exploits sentiment analysis and natural language processing techniques to automatically detect hater users in online social networks.
Given a target user, the workflow we implemented in our system uses sentiment analysis techniques to identify hate speeches posted by the user and exploits a lexicon that extends that of the Italian Hate Map project to assign to the person one or more labels that describe the nature of the hate speeches she posted.
As future work, we plan to arrange a user study,
specifically designed for young people, to evaluate
the effectiveness of the system as a Quantified Self
tool
Marco Polignano and Pierpaolo Basile. 2018. Hansel: Italian hate speech detection through ensemble learning and deep neural networks. In EVALITA@ CLiC-it.