Discourse on Middle Eastern Refugee Waves and War, Mainly in 2015, through "Greek" Tweets. The Semi-Fuzzy and Semi-Automated temporaLDA and TransGIS-LDA-SVM Approaches

Abstract

Stathis G. Arapostathis

Current article is an extended version of a paper presented at the ITDRR-2023 conference on the spatiotemporal archiving of Greek tweets regarding the 2015 Middle Eastern refugee waves and related war conflicts. The paper further explores the spatial factor. Specifically, a semi-fuzzy approach based on Machine Learning and GIS is introduced. Location Entity Recognition (LER) using a transformer, geocoding, extensive GIS processing for geolocation extraction, Latent Dirichlet Allocation (LDA) models, and Support Vector Machine (SVM) classification were combined, creating an innovative methodology. Both the temporal aspect (Phase 1) and the spatial aspect (Phase 2) were examined through subsets generated from locations refered to specific countries. In total, 1,780 topics (Phase 1, temporal factor) and 1,131 topics (Phase 2, spatial factor) were generated from temporal and spatial subsets of a corpus comprising approximately 1.4 million Greek tweets.

Through GIS processing, 99.7% of approximately one million geolocations were validated or revised, resulting in a unique geodatabase containing the locations referred to the majority of Greek tweets discussing the Middle Eastern refugee waves and related war conflicts, primarily those of 2015.

GIS processing also provided location frequencies and subsets of tweets referring to each location. The importance of the spatial factor emerged from the unequal distribution of location references across the discussions in terms of both topic diversity and frequency. These distributions are presented in thematic maps, while the proportions of the two main classes, refugees and war, are presented for 15 countries, including Greece, Syria, Turkey, Iraq, North Macedonia, Jordan, Israel (including Palestine), Russia, and the USA.

The current research can be regarded as a highly innovative interdisciplinary approach, also considering the significance of the geographical area, as Greece was the main entry point for Middle Eastern refugees during that period. The proposed methodology successfully combines machine learning, other state-of-the-art methods, GIS techniques, and human expert assessment, resulting to the effective semi-fuzzy processing of more than one million social media posts concerning a topic of considerable social and geopolitical significance.

PDF