From cognitive asymmetries to NLP optimization: Quantifying type-frequency interactions in emotional text processing through eye-tracking experiments

Current sentiment analysis models in Natural Language Processing (NLP) inadequately distinguish emotion-label words which directly represents emotional meaning (e.g., “sad”) from emotion-laden words that indirectly represents emotional meaning (e.g., “tomb”), despite their fundamentally different semantic architectures. Through a 2 × 2 factorial design simultaneously manipulating emotional word type (label vs. laden) and word frequency (high vs. low), this eye-tracking study dissects how these variables interactively shape emotional text processing across valences: Experiment 1 (N = 40, positive valence) and Experiment 2 (N = 40, negative valence). Results revealed three key findings: First, emotion-label words elicited stronger emotional activation and sustained attention than emotion-laden words across processing stages. Second, frequency effects exhibited valence-dependent patterns—high-frequency enhanced semantic extraction for positive label words but reduced perceptual attention to negative label words. Third, low-frequency amplified emotional salience in negative label words while diminishing activation in positive counterparts. These findings establish that emotional word type and word frequency interactively shape affective text comprehension through distinct cognitive mechanisms. Our empirical evidence advances NLP sentiment analysis by 1) demonstrating the necessity of semantic differentiation between emotional word types for accurate sentiment modeling, and 2) providing cognitive processing benchmarks to optimize emotion lexicon construction in machine learning applications. The methodological integration of experimental psycholinguistics with computational linguistics offers new pathways for developing neurologically grounded NLP systems.