A Multi-Model Collaborative Sentiment Analysis Framework for Tourism Reviews Enhanced by Adversarial Learning

Abstract

Sentiment analysis holds significant value in processing tourism reviews, aiming to automatically identify emotional tendencies from user-generated texts to support service quality evaluation and product optimization. Existing approaches predominantly rely on single-model architectures, which often exhibit limited generalization capabilities when confronted with complex and implicit emotional expressions. Moreover, they generally lack mechanisms for credibility assessment and dynamic optimization of analysis results, making it difficult to simultaneously improve the accuracy, stability, and interpretability of sentiment judgment. This paper proposes a sentiment analysis method based on multi-level model collaboration and adversarial learning. The method begins by preprocessing user tourism review texts to construct a feature matrix. This matrix is then fed into two trained models: a random forest sub-model and an attention-based Long Short-Term Memory (LSTM) model, which output the first and second sentiment probabilities, respectively. These are weighted and fused to generate the third sentiment probability. Furthermore, a generator-discriminator adversarial architecture is designed: the feature matrix along with the first two sentiment probabilities are input into the generator to produce a sentiment analysis report. The discriminator then evaluates the authenticity of the report using a confidence threshold and outputs a dynamically optimized fourth sentiment probability. Finally, the third and fourth sentiment probabilities are fused to obtain the final sentiment probability. Experimental results demonstrate that compared to traditional single-model or simple model fusion methods, the proposed approach achieves higher accuracy and F1 scores across multiple sentiment classification tasks. It exhibits particularly strong robustness when handling implicit emotions, contradictory expressions, and cross-domain tourism texts. The introduction of the adversarial learning mechanism significantly enhances the model's adaptability to noisy and sparse data, effectively enabling dynamic calibration of sentiment analysis outcomes. By integrating a hybrid architecture of statistical learning and deep learning, along with multi-level probability fusion and adversarial optimization, this method provides a solution for tourism review sentiment analysis that offers higher precision and stronger interpretability, thereby contributing to the evolution of related intelligent systems toward collaborative and adaptive judgment paradigms.

Keywords: Sentiment Analysis, Multi-level Model Collaboration, Adversarial Learning

DOI: 10.54941/ahfe1007530