The February 2023 Turkey-Syria earthquake sequence, impacting 350,000 square kilometers and resulting in over 59,000 fatalities, generated an unprecedented volume of multilingually rich digital discourse. This study examines a corpus of 478,052 tweets from Turkish, Arabic, and English streams during the acute crisis phase using a computational social science methodology. This study uses the mDeBERTa-v3 architecture to overcome the limitations of conventional lexiconbased models in managing such language volatility, utilizing its disentangled attention mechanism to extract subtle emotional insights. Our results show a clear division in public opinion: a highly connected Distress Cluster (Fear and Anger, r = 0.72) and a dominant Pro-social Cluster (35.04% Solidarity). The transition from Honeymoon to Disillusionment phases, as well as a notable Second Shock on February 20 caused by a huge 6.4 magnitude aftershock, are empirically identified by mapping these temporal fluctuations against the Zunin and Myers (2000) Model of Disaster Phases. The findings show that pro-social and distress sentiments had a near-zero correlation (r = -0.01), indicating a parallel, dual-track psychological reaction. This work provides a consistent benchmark for future multilingual disaster-response analytics by validating the effectiveness of transformer-based Knowledge Discovery in producing highresolution blueprints of community trauma.

mDeBERTa-v3,entiment analysis,disaster analytic,multilingual social media,Turkey-Syria earthquake