par Leroux, Alexandre ;Gagliolo, Matteo
Référence Sunbelt 2020 (13-17 june 2020: Virtual Conference)
Publication Non publié, 2020-06-14
Communication à un colloque
Résumé : Belgium, and Europe in general, experienced an increased attention of publicopinion to immigration in the last few years, accompanied by outbursts ofactivity on online social networks, ranging from solidarity with migrants andrefugees, to xeno- and Islamophobia. These digital traces provide ground tostudy citizens' opinions and attitudes towards refugees and migrants. Combiningtextual information with Facebook network features allows us to piece together atopography of the disparate citizen narratives about migration. Our ongoingresearch aims to investigate those opinions and their evolution, as discussed onFacebook.The analysed corpus consists of 24.8 million comments written on a set of 15 000Facebook pages related to migration, between 2014 and 2018, collected via theplatform's API. Out of the 6000 pages reporting geographic data, 83% are locatedin Belgium. For this work we will focus on the comments to posts published onFrench-speaking pages.As we are interested only in migration related discourse, after lemmatizing andfiltering out stop words, we further filter the dataset through string matching,using /word2vec/, keeping only comments containing terms with the highestlexical and semantic similarity to ``refugees'' and ``migrants''.In order to observe the evolution of discursive patterns over time, we follow anapproach similar to Rule, Cointet, and Bearman (2015): we split the collectionof comments in four-month intervals and define the corresponding co-occurrencenetworks, where the nodes represent lemmas, and the weighted edges represent thesimilarity between each pair of lemmas, evaluated as the cosine similarity amongcolumns of the lemma-comment matrix. We then identify themes by clustering eachnetwork, and labeling the obtained communities of words .In the last step of our analysis, we will look upon the relationships amongthese communities, and their evolution over time. Using Bhattacharyya distance,we intend to measure the temporal evolution of the distribution of lemmas withineach cluster, as well as define a macro temporal network connecting the themes,and varying across the 4-months periods. This should allow us to observe theimpact of punctual events, and identify trends in the public discourse over this5-year period.