Conventionally, impression quantification through review analysis relies on frequently appearing impression words, making it difficult to directly estimate impression scales that appear infrequently in review texts. This study proposes a method to predict impression evaluation scores for low-frequency scales in apparel products by leveraging existing scores for four frequent scales (“Dark- Bright,” “Tight-Loose,” “Formal-Casual,” and “Simple-Fancy”). Specifically, four high-frequency scales were vectorized using BERT with input sentence templates to construct semantic axes reflecting the fashion context. The position of the target scale on each semantic axis was calculated using cosine similarity, and the final score was estimated by applying weights based on the resulting direction and scores of the high-frequency scales. Using “Muted-Vivid” as an example, the estimated scores were shown to reflect semantically related scales, such as “Dark-Bright.” This method enables consistent score estimation across diverse impression scales, regardless of their frequency of occurrence in reviews.
ZHU et al. (Thu,) studied this question.