Towards a Corpus of Violence Acts in Arabic Social Media

Towards a corpus of violence acts in Arabic social media Ayman Alhelbawy,♣,♠,♦ Udo Kruschwitz,♠ Massimo Poesio ♠ ♣ Minority Rights Group, London, UK ♠ University of Essex, Colchester, UK ♦ Fayoum University, Fayoum, Egypt aalhel, udo, poesio@essex.ac.uk Abstract In this paper we present a new corpus of Arabic tweets that mention some form of violent event, developed to support the automatic identification of human rights abuses and different violent acts. The dataset was manually labelled for seven classes of violence using crowdsourcing. Only tweets classified with a high degree of agreement were included in the final dataset. Keywords: Arabic NLP, Human Rights Abuse, Violence, Arabic Twitter Corpus, Classification 1. Introduction Unfortunately, a great number of Human Rights Abuse (HRA) incidents take place every day in Arab countries in a conflict or quasi-conflict situation like Syria, Iraq, and Libya. Human Right Organizations do their best to find out about such incidents, but they have very few represen- tatives and these cannot be everywhere. Social media, to- gether with text messages and emails, provide an alternative channel to report such abuse. And Human Language Tech- nology could provide a means to retrieve large quantities of such reports from a variety of media and assess their plau- sibility. In order to support such activity, we created a cor- pus of tweets reporting violent activities, including HRAs as well as other types of crime and conflict. To our knowl- edge, there is no resource to support such research either in English or Arabic. A further issue is that people posting on Twitter normally do not use Modern Standard Arabic (MSA)–only news organizations do so–but the ‘dialectal’ forms of Arabic used in everyday life. This means, for in- stance, that different words are used to express the same sense depending on where the posters are located. In this paper, we will describe the corpus and the method- ology we adopted to create it. In Section 2. we discuss how we collected an initial 557,576 tweets using a semi- automatically prepared list of violence words and a twitter word filter. Additional filters were deployed to filter posts that contained violence words but do not really report vi- olent accidents. In Section 3. we discuss the classification scheme we developed to classify the tweets reporting vio- lence. This scheme was used to manually annotate a dataset of 20,151 tweets using crowdsourcing. Finally, we investi- gate the degree of agreement achieved using the classifica- tion scheme, and report that we observe a substantial level of inter-rater agreement. 2. Data Collection As we are only interested in Arabic tweets, we use the Twit- ter language filter to receive only Arabic tweets. But there are still millions of tweets in Arabic, a small fraction of which may mention some kind of violence. So, further fil- ters are required to find an initial set of tweets that may mention some kind of violence to use for manual classifi- cation. We developed two filters for this purpose that we discuss in this …

Content truncated to 500 words