Tsvi Kuflik

In this study, we investigated the potential of text mining techniques to identify the emotive components of political articles written in Arabic. We employed pretrained word embedding models alongside a precompiled emotive lexicon to represent political terms and documents, constructing contextual, domain-specific emotive profiles based on the classical spectrum of emotions (anger, disgust, fear, joy, surprise, and sadness). This approach introduces an additional emotive layer that researchers can use to enhance text and intent interpretation. We hypothesized that this methodology would enable the identification of significant emotive differences in political lexicons across political groups and time periods, thereby contributing to a deeper understanding of historical and political phenomena. To demonstrate the methodology, we conducted a case study examining articles related to the Muslim Brotherhood during the Arab Spring (2011–2012) and the subsequent period (2013–2020). Specifically, we analyzed emotive elements in Egyptian political discourse as reflected in both pro– and anti–Muslim Brotherhood newspapers. The findings indicate that both overt and covert emotions in pro–Muslim Brotherhood newspapers shifted from predominantly positive to predominantly negative following the political transition from the period of Muslim Brotherhood governance under Mohamed Morsi to the presidency of Abdel Fattah El-Sisi. Conversely, opposition newspapers exhibited a shift from predominantly negative to predominantly positive emotions across the same periods. The contribution of this study extends beyond the specific case examined. The proposed methodology offers a framework for identifying changes in overt and covert emotions over time in media discourse, not only in Arabic but potentially in other languages as well. The ability to classify media outlets by emotion profile suggests potential for computational detection of ideological bias in Arabic texts. Nevertheless, the current results remain somewhat exploratory and warrant further investigation in future research.

Proactive health interfaces require accurate forecasting of physiological states to deliver timely interventions. In this work-in-progress, we investigate the feasibility of predicting short-term glucose dynamics in older adults with Type 2 Diabetes using "activity fingerprints" - frequency-domain features derived from wearable posture signals. Analyzing data from 103 participants, we utilize Fast Fourier Transform to capture individualized movement rhythms. Granger causality analysis identifies these rhythms as leading indicators of glucose changes with a predictive horizon of approximately 40 minutes. We demonstrate that an XGBoost model utilizing these fingerprints outperforms standard baselines (Mean Prediction and Last-Value-Carried-Forward). These preliminary findings validate the potential of posture dynamics as a non-invasive predictor, laying the technical groundwork for future intelligent interfaces capable of prompting timely glucose measurements.