Hybrid Lexicon-LLM Arbitration for Nuanced Sentiment and Dual-Affect Parsing in Customer Experience Analytics

The majority of sentiment analysis architectures in enterprise Customer Experience (CX) pipelines compress natural language data into a single scalar point: positive, negative, or neutral. Although computationally efficient, this reductionist paradigm fails to decode modern consumer feedback involving sarcasm, irony, frequent emoji sequences, and simultaneous conflicting affects (e.g., product excitement mixed with pricing anxiety). In this paper, we introduce SentimentDetector, an affective natural language processing pipeline designed for mid-market B2B SaaS deployments. SentimentDetector pairs rule-based lexical affective engines—specifically an augmented Valence Aware Dictionary and sEntiment Reasoner (VADER) and the National Research Council Canada (NRC) Emotion Lexicon—with a multimodal, high-reasoning large language model gateway powered by Claude Sonnet 4.5. The architecture features an operational mode selector that dynamically routes text, adjudicates lexicon-model discrepancies, and outputs multi-label primary and secondary affects alongside qualitative rationale. Evaluated via an empirical, single-blind human validation benchmark (N = 30, 240 evaluation pairs) across four linguistic challenge strata, the system demonstrated substantial inter-annotator agreement (κ= 0.74 for polarity, κ= 0.69 for dual affect) and an overall alignment score of µ = 4.41 ± 0.68 on a 5-point Likert scale. Our findings illustrate that hybrid arbitration outperforms rigid lexical heuristics and unguided prompt architectures, offering enterprise CX teams an interpretable framework for post-launch sentiment discovery.