This study aims to introduce a methodology of building evaluation-annotated datasets based on evaluation triples that we extracted from reviews in fashion e-commerce applications. The evaluation triples defined in this study consist of a target, an aspect, and a value. We classified the aspect/value pairs into 35 categories, within three types: related to information, judgment or suggestion. The triples were represented under a set of Local Grammar Graphs based on the domain-specific dictionary DECO-DOM. By applying these linguistic resources to the target review texts, we generated a large-scale annotated dataset. In this study, the Semi-automatic Symbolic Propagation (SSP) methodology proposed by Nam (2021) was adopted to generate the EVAD dataset of about 200,000 review texts. The evaluation of the SSP approach shows an F1-score of 0.91, which confirms the reliability of the process.