Generating Experiential Descriptions and Estimating Evidence Using Generative Language Model and User Products Reviews
Shen Chenfu, Yoshiyuki Shoji, Takehiro Yamamoto, Katsumi Tanaka, Martin J. Dürst · 2024
This paper introduces a method to transform technical product descriptions into user-friendly experiential descriptions, while also highlighting relevant parts of the original description. Product descriptions often are hard to understand without prior knowledge. For example, a beginner with a camera cannot understand technical descriptions like “ISO sensitivity 51,200”. Our method translated this description to more relatable phrases such as “captures clear faces even at night.” Our method adopts a generative language model to enable such experiential description generation and evidence estimation. Our method first trains a model with pairs of product descriptions and reviews. The trained model generates many candidate experiential descriptions when given product descriptions. After training, our method uses an ablation-based approach to estimate the evident description of the generated candidates. It checks for the frequency of words in the generated narrative when a portion of the description is removed. For example, terms like “night” or “clear” became less prevalent in reviews when “ISO sensitivity” was removed from the input description. Subject experiments with the actual review dataset verified our method's effectiveness in generating accurate narratives highlighting product features.