From reviews to empathy: Natural language processing-driven automated empathy mapping and its methodological implications

Traditional empathy mapping (EM) in human–computer interaction suffers from subjectivity, limited scalability, and poor reproducibility. This study introduces a multi-layered natural language processing framework that automatically generates EMs from large-scale product reviews by integrating Think–Feel–Say–Do with customer journey mapping (CJM). A dataset of 4,845 Amazon robot vacuum reviews (30,642 sentences) was analyzed using zero-shot classification (e.g., BART-Large-MNLI and RoBERTa-Large-MNLI), interpretability (e.g., local interpretable model-agnostic explanations), aspect-based sentiment analysis (e.g., ABSAbank-RoBERTa, local context-focused-Bidirectional Encoder Representations from Transformers [BERT]), and CJM alignment (e.g., sentence-BERT). The findings highlight a predominance of Think (46.6%) and Do (42.6%), while Feel (9.4%) and Say (1.5%)—though less frequent—convey strong emotional polarity, with 56% of content at extremes. “Device experience” dominates as the key touchpoint (66%) and the CJM experience stage (80.7%). Aspect analysis emphasizes technical (73.2%) and commercial (24.7%) drivers, particularly cleaning performance, battery life, price, and warranty. Cluster analysis identifies three profiles: action-intensive, rational-evaluation, and narrative-emotion. The framework advances EM as scalable, reproducible, and evidence-based, supporting user experience optimization, persona design, and real-time monitoring.