d1n910 / argugpt

ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models

This project is still a work in progress. We will keep updating our corpus, AIGC detectors, and research paper. So, you are welcomed to star and watch this repo for future update. 【本项目还在进行中。后续我们持续更新语料、探测器以及论文。请持续关注我们！】

If you are confused by any of the following contents or you have any suggestion, please contact us at [email protected]. 【如果有任何疑问以及建议，欢迎通过邮箱联系我们。】

In this repo, you will see 【内容如下】:

• A corpus of 4k argumentative essays written by 7 GPT models. 【我们用7个GPT模型生成了4k议论文。】
• Several classifiers for detecting GPT-generated argumentative essays (99% accuracy at doc-level; 93% at sentence-level). 【训练了相关模型检测议论文是否由GPT模型生成（文章检测达到99%的准确率，句子检测达到93%的准确率）。】

And we make following resources public 【我们公布了以下资源】:

• Models: We fine-tuned RoBERTa to detect machine-generated argumentative essays. We trained both doc-level and sentence-level classifiers. 【我们微调了RoBERTa以检测机器生成的议论文，我们在文档级别和句子级别上都进行了训练。】
• App-demo: We built an app in huggingface space which bases on the sentence-level detector. 【基于句子级别的检测模型，我们搭建了一个App，可以在网页中直接使用。】
• Paper: We released our research paper in arxiv. 【我们的论文预印在了arxiv中。】

Introduction to ArguGPT corpus 语料清单

We compiled a 8k human-machine (4k human v.s. 4k machine) comparison argumentative essays. Sub-corpora are listed in the following table. 【ArguGPT包含了8k议论文文章，包括4k人类议论文和4k机器生成议论文，语料清单如下。】

sub-corpus	# essays	# tokens	mean len	source	access
WECCL-human	1,845	450,657	244	SWECCL 2.0	SWECCL (Wen & Wang, 2008)
WECCL-machine	1,813	442,531	244	GPT models	Released. See in data/argugpt folder
TOEFL-human	1,680	503,504	299	TOEFL11	Purchased in LDC
TOEFL-machine	1,635	442,963	270	GPT models	Released. See in data/argugpt folder
GRE-human	590	341,495	578	GRE-prep materials	No copyright to release
GRE-machine	590	268,640	455	GPT models	Released. See in data/argugpt folder

More detailed information about the ArguGPT corpus can be seen in our paper. 【有关ArguGPT的更多信息，请参考我们的论文。】

Data split and baseline 数据划分及基准模型

We first split the data into train/dev/test sets. The test split of TOEFL essays are as well evaluated by human participants in the Turing test. Then we established baselines by training detectors based on SVMs and RoBERTa. Moreover, we conducted ablation study to see the effect of reducing training data points. Finally, we evaluated two detectors on our own test set, namely GPTZero and RoBERTa trained by Guo et al. (2023). 【我们将数据集进行了划分。其中，图灵测试（人类测评）中使用的数据为测试集中的托福作文。在此数据集上，我们训练了SVM和RoBERTa模型，作为该数据集的基准模型。此外，我们在不同大小的训练集上进行了训练。】

split	TOEFL	WECCL	GRE	total
train	3,058	2,715	980	6,753
dev	300	300	100	700
test	300	300	100	700

Accuracy of human evaluators on the TOEFL split of test set is only 64.65%, far lagging behind ML detectors/classifiers. 【对测试集中的托福文章进行图灵测试，人类参与者的准确率仅有64.65%，远低于基于机器学习的检测器。】

train data	test data	maj. bsln	RoBERTa	Best SVM	GPTZero	Guo et al. (2023)
doc-all	doc test	50	99.38	95.14	98.86	89.86
doc-50%	doc test	50	99.76	94.14	-	-
doc-25%	doc test	50	99.14	93.86	-	-
doc-10%	doc test	50	97.67	92.29	-	-
doc-all	para test	52.62	74.58	83.61	-	-
para-all	para test	52.62	97.88	90.55	92.11	79.95
doc-all	sent test	54.18	49.73	72.15	-
sent-all	sent test	54.18	93.84	81	90.10	71.44

Notes: We broke down the dataset of essays into sentences and paragraphs, trained the models from document-, paragraph-, and sentence-level, and evaluated these models. More experiments will be carried out in this regard. 【我们把文章拆分为段落和句子，并且分别从文章、段落和句子的层次训练了模型，并进行了一些测试。后续会开展更多相关的实验。】

Team members 团队介绍

We are a group of students who are interested in language, linguistics, and NLP as well. Led by Hai Hu, an assistant professor from English Department of Shanghai Jiao Tong University (SJTU), whose research interest is computational linguistics, we hope to contribute something interesting to CL and NLP commuities from the perspective of language leaners. 【大家好，我们是一群热爱语言、同时也对NLP技术感兴趣的本科/研究生，团队由交大外院的胡海老师指导。我们会从语言学习者的角度，进行一些有趣的实验与研究。如果有幸的话，希望我们的研究能够对计算语言学和NLP社群做出一些小小的贡献！】

Name	姓名	Affiliation	所属机构	Status
Hai Hu	胡海	SFL, SJTU	上海交通大学外院	Assistant Professor
Yiwen Zhang	张伊文	Amazon	亚马逊	Language Engineer
Shisen Yue	岳士森	SFL, SJTU	上海交通大学外院	Undergraduate
Wanyang Zhang	章万扬	SFL, HUST	华中科技大学外院	Undergraduate
Xiaojing Zhao	赵晓婧	SFL, SJTU	上海交通大学外院	Graduate
Xinyuan Cheng	程心远	SFL, SJTU	上海交通大学外院	Undergraduate
Yikang Liu	刘逸康	SFL, HUST	华中科技大学外院	Undergraduate
Ziyin Zhang	张子殷	SEIEE, SJTU	上海交通大学电院	Undergraduate

Citation

Please cite our work as

@misc{liu2023argugpt,
      title={ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models}, 
      author={Yikang Liu and Ziyin Zhang and Wanyang Zhang and Shisen Yue and Xiaojing Zhao and Xinyuan Cheng and Yiwen Zhang and Hai Hu},
      year={2023},
      eprint={2304.07666},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}