Scholarlink

<p>User reviews on gaming platforms contain valuable indicators of satisfaction, but sentiment classification remains challenging due to slangs and domain-specific terminology. In this paper a lightweight hybrid ensemble is proposed for Steam review sentiment analysis that integrates three complementary components: (i) a Naïve Bayes classifier trained on NLTK token features, (ii) a logistic regression model using TF–IDF n-grams, and (iii) a fine-tuned Sentence-BERT (SBERT) transformer classifier for contextual semantic sentiment prediction. By combining symbolic, statistical, and semantic representations through an adaptive performance-weighted probability fusion strategy, the ensemble improves robustness to linguistic variability while remaining highly efficient compared to transformer-based end-to-end models. Experiments on a dataset of 147,586 Steam reviews across 160 games show that the proposed ensemble achieves an F1-score of 94.0% and ROC-AUC of 95.0%, outperforming classical baselines (Naïve Bayes, Random Forest, XGBoost) and achieving competitive results relative to the OPT-175B large language model, despite being orders of magnitude smaller. The proposed weighted fusion mechanism assigns dynamic importance to individual classifiers based on validation performance, enabling the ensemble to better exploit complementary strengths among lexical, statistical, and semantic representations. Ablation studies confirm that each component contributes distinct strengths: SBERT provides semantic depth, TF–IDF captures discriminative n-grams, and Naïve Bayes reinforces frequent lexical cues. The findings demonstrate that hybridizing classical machine learning with transformer embeddings offers an effective and computationally efficient strategy for domain-specific sentiment analysis of game reviews.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight hybrid adaptive weighted ensemble integrating fine-tuned SBERT with lexical and statistical models for robust steam review sentiment classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Nima Etemad Golestani,
  • Mohammad Hossein Moattar

摘要

User reviews on gaming platforms contain valuable indicators of satisfaction, but sentiment classification remains challenging due to slangs and domain-specific terminology. In this paper a lightweight hybrid ensemble is proposed for Steam review sentiment analysis that integrates three complementary components: (i) a Naïve Bayes classifier trained on NLTK token features, (ii) a logistic regression model using TF–IDF n-grams, and (iii) a fine-tuned Sentence-BERT (SBERT) transformer classifier for contextual semantic sentiment prediction. By combining symbolic, statistical, and semantic representations through an adaptive performance-weighted probability fusion strategy, the ensemble improves robustness to linguistic variability while remaining highly efficient compared to transformer-based end-to-end models. Experiments on a dataset of 147,586 Steam reviews across 160 games show that the proposed ensemble achieves an F1-score of 94.0% and ROC-AUC of 95.0%, outperforming classical baselines (Naïve Bayes, Random Forest, XGBoost) and achieving competitive results relative to the OPT-175B large language model, despite being orders of magnitude smaller. The proposed weighted fusion mechanism assigns dynamic importance to individual classifiers based on validation performance, enabling the ensemble to better exploit complementary strengths among lexical, statistical, and semantic representations. Ablation studies confirm that each component contributes distinct strengths: SBERT provides semantic depth, TF–IDF captures discriminative n-grams, and Naïve Bayes reinforces frequent lexical cues. The findings demonstrate that hybridizing classical machine learning with transformer embeddings offers an effective and computationally efficient strategy for domain-specific sentiment analysis of game reviews.