A Rewrite That Keeps the Label Has Already Read It: Why an LLM Paraphrase Can Raise a Sentiment Classifier's Accuracy Without the Classifier Getting Better, Four Accounts of the Gain, and the Re-Annotation and Neutral-Rate Controls That Would Separate Them

APA

Mahendrakar, P. (2026). A Rewrite That Keeps the Label Has Already Read It: Why an LLM Paraphrase Can Raise a Sentiment Classifier's Accuracy Without the Classifier Getting Better, Four Accounts of the Gain, and the Re-Annotation and Neutral-Rate Controls That Would Separate Them [Journal article]. Life of Research. https://doi.org/10.5281/zenodo.23406462

Copy APA

BibTeX

@misc{lor1102026,
  title        = {A Rewrite That Keeps the Label Has Already Read It: Why an LLM Paraphrase Can Raise a Sentiment Classifier's Accuracy Without the Classifier Getting Better, Four Accounts of the Gain, and the Re-Annotation and Neutral-Rate Controls That Would Separate Them},
  author       = {Pranay Mahendrakar},
  year         = {2026},
  publisher    = {Life of Research},
  doi          = {10.5281/zenodo.23406462},
  url          = {https://lifeofresearch.com/articles/110},
  keywords     = {sentiment analysis, AI-generated text, paraphrase robustness, label preservation, distribution shift, evaluation methodology, selective classification, sarcasm},
  note         = {LOR-00110}
}

Copy BibTeX More on the record page