Browsing by Issue Date, starting with "2020-02-21"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Unsupervised and Language Independent Approach to Extremism and Collective Radicalization UnderstandingPublication . Albardeiro, Miguel Ângelo Serra; Pais, Sebastião Augusto Rodrigues Figueiredo; Cordeiro, João Paulo da CostaIncreasingly in social media, we find cases where groups are organized to protest against something, often in those groups, members with extremist ideologies are inserted. These cases are happing more often, groups are created for the organization of peaceful protests and someone starts a topic with an extremist language leading, sometimes, to a radicalisation of the group. This research aims to create an approach that allows the detection of cases of extremism and collective radicalisation within social networks, this should be done in an unsupervised and independent of language way. The methods used to achieve the intended objectives are the creation of a lexicon of extreme sentiment terms named ExtremeSentiLex and a classifier of extreme sentiment in which the input is the extreme sentiment terms and the social network post. For the development of these tools were used purely statistical natural language processing methods. To validate the ExtremeSentiLex it was applied using the extreme sentiment classifier, the input posts that are analysed are posts from a dataset already validated by the scientific community. For a comparative study, word embeddings are used to expand the first ExtremeSentiLex obtained and a test is also performed in which the ExtremeSentiLex is balanced and applied to a balanced polarity dataset. The results obtained in this content level research that will be available to the scientific community are the ExtremeSentiLex and several datasets that were evaluated by us regarding the presence of extreme sentiment. At the level of tests performed when the ExtremeSentiLex was validated, the level of precision in finding extreme sentiment at the correct polarity was very high. When applying word embeddings the results dropped. Regarding the ExtremeSentiLex and balanced dataset, the results were very positive. It has been concluded that our dataset is suitable for the application in detecting extreme sentiments in text. Furthermore, it was found that with the help of linguistic and psychological experts the ExtremeSentiLex could be improved. However, this investigation aimed to do so using purely statistical methods. This goal has been successfully achieved.