Benchmark for Neural Paraphrase Detection

这是神经释义检测的基准，用于区分原始内容和机器生成的内容。训练：从 4,012 篇（英文）维基百科文章中提取 1,474,230 个对齐的段落（98,282 个原始段落，1,375,948 个用 3 个模型和 5 个超参数配置进行释义的段落，每个 98,282 个）。测试：BERT-large（加壳）：arXiv - Original - 20,966；释义 - 20,966；论文 - 原创 - 5,226；释义 - 5,226；维基百科 - 原始 - 39,241；释义 - 39,241； RoBERTa-large (case)：arXiv - Original - 20,966；释义 - 20,966；论文 - 原创 - 5,226；释义 - 5,226；维基百科 - 原始 - 39,241；释义 - 39,241； Longformer-large（未加壳）：arXiv - Original - 20,966；释义 - 20,966；论文 - 原创 - 5,226；释义 - 5,226；维基百科 - 原始 - 39,241；释义 - 39,241；

This is a benchmark for neural paraphrase detection, designed to differentiate between original content and machine-generated content.
Training set: 1,474,230 aligned paragraphs were extracted from 4,012 English Wikipedia articles, including 98,282 original paragraphs and 1,375,948 paraphrased paragraphs. The paraphrased paragraphs were generated using 3 models with 5 hyperparameter configurations, with each configuration producing 98,282 paraphrased paragraphs.
Test sets:
- BERT-large (cased):
* arXiv dataset: 20,966 original paragraphs, 20,966 paraphrased paragraphs
* Paper dataset: 5,226 original paragraphs, 5,226 paraphrased paragraphs
* Wikipedia dataset: 39,241 original paragraphs, 39,241 paraphrased paragraphs
- RoBERTa-large (cased):
* arXiv dataset: 20,966 original paragraphs, 20,966 paraphrased paragraphs
* Paper dataset: 5,226 original paragraphs, 5,226 paraphrased paragraphs
* Wikipedia dataset: 39,241 original paragraphs, 39,241 paraphrased paragraphs
- Longformer-large (uncased):
* arXiv dataset: 20,966 original paragraphs, 20,966 paraphrased paragraphs
* Paper dataset: 5,226 original paragraphs, 5,226 paraphrased paragraphs
* Wikipedia dataset: 39,241 original paragraphs, 39,241 paraphrased paragraphs

搜集汇总

背景与挑战

背景概述

该数据集是用于神经释义检测的基准，旨在区分原始内容和机器生成的内容。其训练数据包含从英文维基百科文章中提取的约147万个对齐段落，测试数据则涉及BERT-large、RoBERTa-large和Longformer-large等模型在arXiv、论文和维基百科数据集上的评估。