Sentiment Shift Pattern Recognition for Anomaly Detection in Temporally Ordered News Streams

This study proposes a rule-based pattern recognition framework for identifying sentiment shift anomalies in time-ordered textual data. The method combines lexicon-based sentiment scoring with a temporal fluctuation threshold to detect significant tonal transitions between adjacent news entries. A 60 -month dataset comprising 155 curated English-language news articles was compiled using a two-stage relevance filtering process. This filtering removed ambiguous and off-topic entries based on lexical and semantic criteria. Each article was analyzed using the VADER sentiment analyzer to compute both absolute tone scores and relative tone shifts. Alerts were flagged if sentiment polarity fell below a predefined threshold or if the tone shift between consecutive articles exceeded a volatility criterion. Six distinct alert events were identified and confirmed through manual validation. The framework requires no supervised learning, offers full interpretability, and supports real-time anomaly recognition in structured media streams. This work contributes to pattern recognition research by formalizing sentiment shift as a detectable pattern for affective anomaly identification in chronological textual corpora.