Elamur Research⌕
☰
home / papers / 2026 / Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods

Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods

Paper · 2026 · Krishna Thakar, Mohamed Abu Sheha and Emmanuel Thompson

Summary

Which model family works best for review-level sentiment classification is still disputed. Earlier Yelp studies disagree on whether classical learners, neural networks or transformers come out ahead, and few compare all three under identical conditions. This student workshop paper runs that controlled comparison on a balanced sample of 21,000 recent reviews from the Yelp Open Dataset. Each review gets a three-way label (negative, neutral, positive) derived from its star rating. All texts pass through one shared preprocessing pipeline, including spaCy lemmatization, stop-word removal and a rule that marks words following a negator with a not_ prefix. Ten models are then trained: four classical classifiers on TF-IDF features, five neural networks with embeddings learned from scratch, and a fine-tuned RoBERTa. RoBERTa reaches the highest accuracy. SVM and Logistic Regression beat CNN and the recurrent networks, while vanilla LSTM, GRU and RNN drop to chance level, which the authors attribute to untuned training settings. The resulting guidance for practitioners: use transformers when GPUs are available and fall back on linear TF-IDF models as a cheap, reliable option.

Key points

  • Data and splits:
    • Labels: 4–5 stars are positive, 3 stars neutral, 1–2 stars negative.
    • Classes are balanced by random undersampling to about 6,667 reviews each.
    • The data are split 70/30 (about 14,700 training and 6,300 test reviews), with 5-fold stratified cross-validation used for model selection.
  • Model tiers:
    • Classical: Logistic Regression, SVM, Naïve Bayes, Random Forest.
    • Deep learning: CNN, LSTM, BiLSTM, GRU, RNN.
    • Transformer: RoBERTa.
  • Input representations:
    • TF-IDF with 5,000 features for the classical models.
    • Keras tokenizer (10,000-word vocabulary, 100-dimensional embeddings, sequences of 200 tokens) for the neural models.
    • RobertaTokenizerFast with a 128-token limit for RoBERTa.
  • Results:
    • SVM reaches 76.10% accuracy and CNN 74.75%.
    • Vanilla LSTM, RNN and GRU collapse to about 33%. Each was trained with only one configuration (lr 0.001, 5 epochs, batch size 128).
    • The neutral class is the hardest for every model, partly because star-based labels are noisy near class boundaries.
  • Limitations and future work:
    • Limitations: no significance testing, only recent reviews sampled (temporal bias), and no pre-trained word vectors for the neural models.
    • Planned follow-ups: GloVe/FastText embeddings, aspect-level analysis, and larger Yelp subsets.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Southeast Missouri State University.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Sentiment & emotion · 972 works

Cite

Cite as Krishna Thakar, Mohamed Abu Sheha and Emmanuel Thompson. 2026. Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.38

@inproceedings{thakar2026sentiment,
  title = {{Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods}},
  author = {Krishna Thakar, Mohamed Abu Sheha and Emmanuel Thompson},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.38},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.38}
}

References

13 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ali Areshey and Hassan Mathkour. 2023. Transfer learning for sentiment classification using bidirectional encoder representations from transformers (BERT) model. Sensors, 23(11):5232. find
  2. Viacheslav Bochkarev. 2024. Yelp reviews sentiment analysis. GitHub repository. Last Publish January 4, 2024. find
  3. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  4. Md. Shofiqul Islam, Muhammad Nomani Kabir, Ngahzaifa Ab Ghani, Kamal Zuhairi Zamli, Nor Sara-datul Akmar Zulkifli, Md Mustafizur Rahman, and Mohammad Ali Moni. 2024. Challenges and future in deep learning for sentiment analysis: a comprehensive review and a proposed novel hybrid approach. Artificial… find
  5. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751, Doha, Qatar. Association for Computational Linguistics.
  6. Siqi Liu. 2020. Sentiment analysis of yelp reviews: A comparison of techniques and models. Computing Research Repository, arXiv:2004.13851. — [arXiv:2004.13851](https://arxiv.org/abs/2004.13851) find
  7. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. Computing Research Repository, arXiv:1907.11692.
  8. Sepehr Mostafavi, Yeganeh Yahyavi, and Reza Ravanmehr. 2026. Systematic literature review on sentiment analysis using transformers. International Journal of Data Science and Analytics, 22(1):49. find
  9. Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP 2002), pages 79–86. Association for Computational Linguistics.
  10. Ralph Shad, Kaledio Potter, and Abram Gracias. 2024. Natural language processing (NLP) for sentiment analysis: A comparative study of machine learning algorithms. Preprints.org. find
  11. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 16…
  12. Peter D. Turney. 2002. Thumbs up or thumbs down? semantic orientation applied to unsupervised classification of reviews. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 417–424.
  13. Sida Wang and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 90–94, Jeju Island, Korea. Association for Computational Linguis…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/thakar-2026-sentiment-analysis-yelp-review-dataset/