Answers are grounded in this paper and cite their sources.
Uncovers factors driving opposite user evaluations in e-commerce reviews, suggesting improved sentiment analysis techniques.
This study proposes a systematic technique for identifying why users assign opposite evaluations to the same function of the same product. Focusing on 369 Japanese e-commerce reviews of household rice cookers, we first removed unsuitable texts, segmented each review at line breaks to obtain 1 005 topic-level sentences, and fine-tuned a pretrained Japanese BERT model to classify every sentence as positive or negative. Each sentence was then tokenized with MeCab; evaluation-related nouns were embedded by BERT, reduced in dimensionality with t-SNE, and grouped by hierarchical clustering. Representative terms from positive and negative clusters were contrasted to isolate topics that produced divergent sentiments. To capture richer functional context, we further extracted verb–object pairs via dependency parsing with spaCy + GiNZA and limited analysis to pairs appearing at least six times. Qualitative comparison of the five most frequent pairs—“put water”, “unplug outlet”, “press button”, “cook rice”, and “open lid”—revealed three common drivers of polarity shifts: (1) the expectation benchmark set by price or prior models, (2) users’ willingness to tolerate or self-mitigate inconveniences, and (3) perceived long-term risks such as safety and durability. These findings demonstrate that coupling sentiment classification with context-aware clustering of functional expressions effectively uncovers the latent explanatory factors behind divergent user evaluations.
No takes yet. Share an insight, caveat, or question.
HIRAI et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: