2023年数维杯C题思路——AI生成文本的智能识别与检测

近年来，随着信息技术的快速发展，人工智能的各种应用相继涌现。典型应用包括机器人导航、语音识别、图像识别、自然语言处理和智能推荐等。在这些应用中，由ChatGPT领衔的大型语言模型（Large Language Models，LLMs）在全球范围内受到欢迎，并被广泛推广和使用。我们充分认识到这些模型为人们带来的丰富、智能和便利的体验。同时，也需要意识到AI文本生成等工具可能带来的许多风险。

首先，这些大型语言模型是基于文本训练的。不同类型的语言和不同领域的文化背景可能对生成结果产生显著影响。其次，基于数据生成的AI结果可能存在语义偏见，缺乏逻辑连贯性，缺乏创造力。最后，由于学生使用AI生成论文而产生的隐私保护、版权保护和相关学术不端行为的定义等问题，给本科和研究生的教学和培训过程带来了重大困难和挑战。为了防止AI生成文本的滥用，确保生成内容的质量，并讨论如何解决AI生成论文带来的问题，有必要根据主题要求识别和检测AI生成文本的模式，包括领域、模型、图像和公式。

为了确定文本是否由AI生成，除了考虑符合字数要求、生成次数和是否为中英文翻译等因素外，还应注意AI目前缺乏人类情感和判断。这可能导致文本生成中出现“更多短语缺乏实例、缺乏情感、结构单一”等现象或风格。

请使用数学建模解决以下四个问题：

注：赛题发布1小时后，本题的第二问和第三问被组委会更正，下文已更新。

Problem C：Intelligent Recognition and Detection of AI-Generated Text问题更正说明

问题一：请使用AI根据附录I中提供的二十个科学网博客链接的一些文章部分进行改写。并寻找AI文本生成的基本规则，这些规则可以从AI生成所需字数（例如200字、500字等）、生成次数（首次生成后点击“重新生成”按钮）、是否为中英翻译以及生成文本的风格等要求中统计推断出来。

问题二（更正）：根据附录II中提供的十段AI生成的文章，基于第一个问题中获得的模式，请对这些段落的生成次数（不超过五次）、从中文翻译成英文的次数（不超过一次）、从英文翻译成中文的次数（不超过一次）以及输出段落是否有字数要求进行判断。

问题三（更正）：对于AI生成的理论和方法，请仔细考虑每个文章段落是否基于不同的生成语言、是否为翻译、生成次数以及是否存在输出字数限制等因素由AI生成。然后，标记附录III中提供的十篇文章中的每个段落是否由AI生成。

问题四：请建立相关理论和方法，进一步确定文章中的数学模型、图像和公式是否为剽窃内容。并使用附录IV中的示例来演示这一点，并评估所建立的模型。

问题二（原版）：请根据不同生成语言、是否为翻译、生成次数、是否存在输出字数限制等因素，仔细考虑每个文章段落是否由AI生成。然后标记附录III中提供的十篇文章中的每个段落是否由AI生成。

问题三（原版）：对于AI生成的理论和方法，请根据不同生成语言、是否为翻译、生成次数、是否存在输出字数限制等因素，仔细考虑每个文章段落是否由AI生成。然后，标记附录III中提供的十篇文章中的每个段落是否由AI生成。

更新了什么：

问题二
更专注于评估生成和翻译的具体次数，而不仅仅是判断是否由AI生成。（难度增加）
在两个版本中，问题三的描述基本相同。

原题：

Problem C：Intelligent Recognition and Detection of AI-Generated Text

In recent years, with the rapid development of information technology, various applications of artificial intelligence have emerged. Typical applications include robot navigation, speech recognition, image recognition, natural language processing, and intelligent recommendation, among others. Among these applications, large language models (Large Language Models, LLMs) led by ChatGPT have gained popularity worldwide and have been widely promoted and used. While we fully recognize the rich, intelligent, and convenient experiences that these models bring to people. It is also important to be aware of the many risks associated with tools such as AI text generation.

First, these large language models are trained based on text. And different types of languages and cultural backgrounds of different domains can have a significant impact on the generated results. Second, AI-generated results based on data may have semantic biases, lack logical coherence, and lack creativity. Finally, issues such as privacy protection, copyright protection, and the definition of related academic misconduct resulting from students using AI to generate papers pose significant difficulties and challenges to the teaching and training process for undergraduate and graduate students. In order to prevent the misuse of

AI-generated text, ensure the quality of generated content, and discuss how to address the problems caused by AI-generated papers, it is necessary to identify and detect the patterns of AI-generated text according to the topic requirements, including fields, models, images, and formulas.

To determine whether the text is AI-generated, in addition to considering factors such as meeting the word count requirement, the number of times generated, and whether it is Chinese-English translation. It is also important to note that AI currently lacks human emotion and judgment. This can lead to phenomena or styles in text generation such as "more phrases lack of examples, lack of emotion, structure, such as a single".

Please use mathematical modeling to solve the following four problems:

**Problem one: **Please use AI to rewrite some parts of the articles according to the links of twenty blogs on the Web of Science provided in Appendix I. And look for the basic rules of AI text generation, which can be inferred statistically inferred from the requirements of the number of words to be generated by AI (e.g. 200 words, 500 words, etc.), the number of times to be generated (the first time to be generated and then click on the button of "Regenerate"), whether it is a translation of the Chinese and English, and the style of the generated text.

**Problem two: **Please carefully consider whether each paragraph in the articles is generated by AI based on factors. This includes different generated languages, whether it is a translation, number of times generated, and whether there are restrictions on the number of output words, etc. for the theory and method of AI generation. Then mark the results of whether each paragraph in the ten articles provided in Appendix III is generated by AI.

**Problem three: **For the theory and method of AI generation, please carefully consider whether each paragraph in the articles is generated by AI based on factors. It includes different generated languages, whether it is a translation, number of times generated, and whether there are constraints on the output word count, etc. for the theory and method of AI generation. Then, mark the results of whether each paragraph in the ten articles provided in Appendix III is generated by AI.

**Problem four: **Please establish relevant theories and methods for further determining whether mathematical models, images, and formulas in the articles are plagiarized content. And use the examples in Appendix IV to demonstrate this and evaluate the established model.

更新公告转载：

Problem C：Intelligent Recognition and Detection of

AI-Generated Text问题更正说明

** 问题更正说明如下：**

Problem two:** **Based on the ten AI-generated passages provided in Appendix II, please make a judgment on the number of times these passages have been generated (no more than five), the number of times they have been translated from Chinese to English (no more than one), the number of times they have been translated from English to Chinese (no more than one), and whether or not there has been a word requirement for the output passages in light of the pattern obtained in the first question.

Problem three: For the theory and method of AI generation, please carefully consider whether each paragraph in the articles is generated by AI based on factors. It includes different generated languages, whether it is a translation, number of times generated, and whether there are constraints on the output word count, etc., for the theory and method of AI generation. Then, mark the results of whether each paragraph in the ten articles provided in Appendix III is generated by AI.