Social media is full of misinformation that tugs at the heartstrings but doesn’t say anything definite. This study introduces the Polarity-Weighted Semantic Expansion coupled with a Multinomial Naive Bayes Ensemble (PWSE-MNB-E) as a resilient approach for fake-content identification. Core components sentiment-sensitive lexical extraction, semantic condensation, polar-transfer, and frequency-weighting are harmonised within a Naive Bayes framework, supplemented by redundancy curtailment and ensemble moderation to enhance predictive accuracy. The Kaggle Fake-Content repository has been subjected to multiple empirical tests and it is evaluated that PWSE-MNB-E attains an accuracy of 63.36% and an F-Measure of 64.35%, advancing over the MMFND and SARD-FN benchmark models. The F Measure Index and Matthews Correlation Coefficient Indexes also register improvement indexes at 64.48 and 27.21 respectively indicating improvement in reliability in the case of highly unbalanced class distributions. The experiment strongly supports the architecture's ability to capture subtle variation in affect-phraseing and thus offers an effective, low-cost means of counteracting deliberate disinformation.