An Uncertainty-Aware Hybrid Symbolic-Lexicon Framework for Multi-Domain Sentiment Classification
Year: 2026
Type: article
Abstract: Multi-domain sentiment classifiers may return a polarity label even when the evidence behind that decision becomes fragile under domain shift. This study evaluates an uncertainty-aware hybrid symbolic-lexicon framework designed to pair each polarity prediction with a traceable accept-or-defer status. The framework combines a lexical branch adapted only on training data with training-only fitted symbolic feature states, variable-precision rough evidence, and nested score fusion. Evaluation uses the UCI Sentiment Labelled Sentences dataset, comprising 3,000 balanced English sentences from Amazon Cells, IMDb, and Yelp, under repeated within-source, pooled multi-domain, and leave-one-source-out protocols. In pooled evaluation, the framework achieved 83.73% accuracy and 83.76% F1-score, improving F1 by 0.50 points over the adaptive lexical branch. The boundary mechanism deferred 53.55% of cases and retained 92.54% accuracy among accepted predictions; however, matched-coverage logistic regression and SVM confidence rankings produced lower selective risk. The contribution is therefore a reproducible, algorithmically traceable polarity-and-deferment mechanism rather than a claim of universal predictive or uncertainty-ranking superiority.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $32
OpenAlex ID: W7223269718