International Journal of Science, Engineering and TechnologyAuthors: Dr. Raj Kumar, Vishal Kumar, Arjun Kumar Shah, Pradeep Kr. Yadav, Priya Saini
Abstract: Modern sentiment classification systems face severe performance degradation when deployed over private peer-to-peer (P2P) communication streams. This drop is primarily driven by the stark divergence between casual conversational language and the structured datasets historically used to train core language models. This study explores the structural and algorithmic requirements needed to sustain high-accuracy affective computing within contemporary messaging networks. We identify three distinct operational failure modes: tokenization failures caused by widespread leetspeak and evolving youth vernacular; polarity flips triggered by contextual emoji usage that traditional text-cleaning steps omit; and the inherent latency-privacy bottleneck brought on by demanding a sub-50-millisecond execution window under end-to-end encryption. To mitigate these vulnerabilities, we introduce the Hybrid Affective Reasoning System (HARS), an architecture that pairs a Parameter-Efficient Fine-Tuned DistilRoBERTa-v2 core with a deterministic Semantic Compositional Lexicon (SCL) and a dependency parser via spaCy for localized Aspect-Based Sentiment Analysis (ABSA). Additionally, we establish Sentiment Velocity (SV) as a dynamic temporal derivative designed for proactive escalation tracking, demonstrating that early-fusion multi-modal attention preserves accuracy against emoji distortion, containing drop rates to under 3%. Our accompanying three-tiered, privacy-first data ingestion framework utilizes HMAC-based tokenization, structural generalization, and differential privacy validated against EDPB Guidelines 1/2026. Testing against a dataset of 100,000 anonymized P2P interactions demonstrates a 91.4% macro-F1 score, yielding an 8.2% improvement over isolated, single-model configurations. Zero-shot evaluations leveraging Pragmatic Metacognitive Prompting (PMP) on large scale models further prove its viability for identifying nuanced linguistic irony.
