| dc.description.abstract | With the rise of Google restaurant reviews, consumers can avoid the disappointment of dining experiences without the need for first-hand trials, simultaneously driving industry competition and facilitating the natural elimination of unfit businesses from the market. However, as time goes by, this intended purpose of the review system seems to have been distorted, with many operators inducing consumers to provide non-objective reviews through marketing activities, leading to a gradual decline in people′s trust in Google reviews. Therefore, this study aims to explore how to use machine learning techniques to construct a mechanism that can filter out inducive reviews, thereby enhancing the effectiveness of community management strategies.
This research takes Google restaurant reviews in the Hsinchu area as the subject of study, using four feature engineering methods, namely Word2Vec model, FastText model, TF-IDF, and three composite methods, to screen out a collection of inducive feature words from a large number of reviews. Subsequently, we use three different text vectorization methods, including TF-IDF, BoW, and BERT models, and finally, we construct an inducive review detection model with the SVM model as the classifier.
The research results reveal that the best predictive effect is achieved by using TF-IDF to identify similar feature words, using TF-IDF for text vectorization, and then using the SVM model for prediction. The Precision, Recall, and F1 Score of this model all exceed 80%, and the AUC-ROC reaches as high as 93.31%. According to our experimental results, if the review content of a restaurant in the Hsinchu area contains these feature words in Google reviews, then the credibility of the review deserves to be questioned, and the star rating of the restaurant on Google reviews may also lose its objectivity. | en_US |