<![CDATA[Online review platforms such as Yelp play an important role in consumer decision-making, but the growing prevalence of fake reviews undermines their reliability. This study proposes a hybrid approach for fake review detection by integrating stylometric features, language model signals, and semantic embeddings within a unified classification framework. The proposed method combines linguistic indicators, including GPT-2 perplexity, lexical diversity, sentence burstiness, punctuation ratio, and sentiment intensity, with TF-IDF representations and Sentence-BERT embeddings. A composite feature, namely the Semantic Reliability Index (SRI), is introduced to capture interactions between semantic similarity and linguistic characteristics, serving as an auxiliary feature within the hybrid model rather than a standalone classifier. Experiments on a Yelp hotel review dataset demonstrate that the hybrid model outperforms baseline methods in terms of F1-score and AUC, indicating improved discriminative capability. It should be noted that the classification setting is based on a binary transformation of ordinal labels, which may simplify the underlying label structure and influence performance interpretation. Overall, this work's contribution lies in a systematic feature-integration strategy that enhances fake review detection in the evaluated dataset.]]>
Copyrights © 2026
