大型语言模型可以检测社交媒体上的谣言吗?

Qiang Liu1,2   Xiang Tao1,2   Junfei Wu1,2   Shu Wu1,2   Liang Wang1,2
1Center for Research on Intelligent Perception and Computing,
State Key Laboratory of Multimodal Artificial Intelligence Systems,
Institute of Automation, Chinese Academy of Sciences
2School of Artificial Intelligence, University of Chinese Academy of Sciences
qiang.liu@nlpr.ia.ac.cn, {xiang.tao,junfei.wu}@cripac.ia.ac.cn
{shu.wu,wangliang}@nlpr.ia.ac.cn
摘要

在这项工作中,我们研究使用大型语言模型(大语言模型)进行社交媒体上的谣言检测。 然而,大语言模型对社交媒体上包含新闻内容和大量评论的整个传播信息进行推理具有挑战性,因为大语言模型可能无法集中在复杂的传播信息中的关键线索,并且在推理时遇到困难当面对海量、冗余的信息时。 因此,我们提出了一种LLM授权的谣言检测(LeRuD)方法,其中我们设计提示来教大语言模型推理新闻和评论中的重要线索,并将整个传播信息划分为传播链以减少大语言模型的负担。 我们在 Twitter 和微博数据集上进行了广泛的实验,LeRuD 的性能优于几种最先进的谣言检测模型 3.2%7.7% 同时,通过应用大语言模型,LeRuD不需要数据进行训练,因此在少样本或零样本场景中显示出更有前景的谣言检测能力。

大型语言模型可以检测社交媒体上的谣言吗?


Qiang Liu1,2   Xiang Tao1,2   Junfei Wu1,2   Shu Wu1,2   Liang Wang1,2 1Center for Research on Intelligent Perception and Computing, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences 2School of Artificial Intelligence, University of Chinese Academy of Sciences qiang.liu@nlpr.ia.ac.cn, {xiang.tao,junfei.wu}@cripac.ia.ac.cn {shu.wu,wangliang}@nlpr.ia.ac.cn


1简介

随着社交媒体的发展,用户可以更方便地获取信息,信息可以更迅速地传播。 但与此同时,谣言也可以更容易、更广泛地向公众传播。 有必要开展社交媒体谣言自动检测的研究Castillo 等人(2011); Islam等人(2020),还是很有挑战性的。 最近,大语言模型Zhao等人(2023)取得了成功并得到广泛应用。 这给人们带来了极大的便利。 然而,人们也可以利用大语言模型故意生成大量不明信息Vykopal 等人 (2023); Chen and Shu (2024),这加剧了社交媒体上谣言的传播情况。 因此,大语言模型的成功使得自动谣言检测变得更加具有挑战性和紧迫性。

Refer to caption
图1: 大语言模型能检测社交媒体上的谣言吗?

谣言检测的关键在于对社交媒体传播模式的建模,即聚合原始新闻和用户评论的特征。 早期的方法将新闻和评论视为文本序列,并应用循环网络 Ma 等人 (2016)、卷积网络 Yu 等人 (2017) 或注意力网络 Liu 等人 (2018) 用于特征提取。 然后,基于图神经网络的谣言检测模型Lu and Li (2020); Bian 等人 (2020) 已被广泛提出并取得了最先进的性能。

最近,一些研究工作尝试应用大语言模型来识别信息的真实性Chen and Shu (2023) 他们大多使用大语言模型从文本Yang等人(2023)中提取知识或单独分析新闻内容Hu等人(2024) 大语言模型分析社交媒体上包含新闻和评论的传播信息的能力尚未被探索。 同时,正如胡等人(2024)介绍的; Chen and Shu (2024)认为,仅仅依靠新闻的文本特征,大语言模型很难获得有希望的检测结果。 这就需要我们考察大语言模型对社交媒体上整个传播信息进行建模的能力。

然而,利用大语言模型对传播信息进行建模并不容易。 首先,我们发现大语言模型无法集中于传播信息中的关键线索,因此可能会产生错误的预测(第3.23.3)。 其次,注释量通常很大,这给大语言模型的推理带来很大的负担(第3.4),因为大语言模型通常有输入长度的限制,在面对长上下文时会遇到困难或冗余信息 Huang 等人 (2023);谢(2023)

为了教会大语言模型对传播信息进行推理并克服上述缺点,我们提出了一种LLM-empowered Rumor D检测方法,称为LeRuD 首先,我们借鉴了以往的谣言检测模型Ma等人(2016)的经验;刘等人 (2018);马等人 (2018);杨等人 (2022);金等人 (2022);胡等人 (2024); Przybyla (2020),其中突出显示了区分谣言的有用信息。 我们设计适当的提示,教导大语言模型集中注意新闻的写作风格和常识性错误,以及评论中的反驳或冲突。 其次,我们将传播信息划分为传播链,使得大语言模型能够按照时间顺序以合理数量的注释进行逐步推理。

然后,为了评估 LeRuD 方法在检测谣言方面的有效性,我们在 Twitter 和微博数据集上进行了实验(第 4 节)。 我们使用广泛应用的 GPT 3.5 111https://chat.openai.com/ 在我们的实现中。 为了顺利使用 GPT 3.5,我们过滤了存在道德问题的样本。 同时,为了避免数据泄露,我们还过滤了GPT 3.5已经了解新闻事实的样本。 在过滤后的数据集上,LeRuD 的性能优于多种最先进的谣言检测模型 3.2%7.7% 需要注意的是,LeRuD 在零样本设置中实现了这些结果,而其他模型则需要训练数据。 总之,通过我们的努力,大语言模型能够很好地检测社交媒体上的谣言

我们的主要贡献如下:

  • 我们建议应用大语言模型来检测社交媒体上的谣言,其中包括新闻和评论的完整传播信息。

  • 我们设计适当的提示,教会大语言模型集中注意力于新闻和评论中的关键线索,使大语言模型能够从多个角度进行推理。

  • 我们将一条新闻的传播信息划分为一个传播链(Chain-of-Propagation),使得大语言模型每次都能利用合理的信息量进行逐步推理。

  • 我们进行了广泛的实验,比较了几种最先进的谣言检测模型。 结果证明了我们的 LeRuD 方法的优越性。

Vanilla Prompt There is a piece of news: “Samsung paid Apple one-billion fine, with 20 billion coins in 30 trucks!” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
Rational Prompt There is a piece of news: “Samsung paid Apple one-billion fine, with 20 billion coins in 30 trucks!Based on the writing style and the commonsense knowledge, estimate the credibility of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output The news about Samsung paying Apple a one-billion fine with 20 billion coins in 30 trucks seems highly implausible and unrealistic. It lacks specific details, and the use of such an unconventional method for payment raises doubts about the credibility of the information. Therefore, the most appropriate answer would be: A. Fake.
Vanilla Prompt There is a piece of news: “Around 7 o’clock this morning, a major traffic accident occurred on the Yuncheng to Houma section of the Shanxi Dayun Expressway, 9.3 kilometers away from Houma. A total of 47 vehicles collided in a series, resulting in 3 deaths and 7 injuries.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
Rational Prompt There is a piece of news: “Around 7 o’clock this morning, a major traffic accident occurred on the Yuncheng to Houma section of the Shanxi Dayun Expressway, 9.3 kilometers away from Houma. A total of 47 vehicles collided in a series, resulting in 3 deaths and 7 injuries.Based on the writing style and the commonsense knowledge, estimate the credibility of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output The provided news about a major traffic accident on the Yuncheng to Houma section of the Shanxi Dayun Expressway appears to be presented in a straightforward and factual manner. It includes specific details such as the time of the incident, the location, and the number of vehicles involved, as well as the reported casualties (3 deaths and 7 injuries). While I cannot independently verify the information, the news seems plausible and is presented in a manner consistent with typical reporting of traffic accidents. Therefore, based on the writing style and common sense, it appears to be: B. Real.
表格1: 使用 GPT 3.5 在两个样本上分析新闻内容。 如果没有适当的提示,大语言模型很难判断新闻的可信度。 因此,我们设计了提示(红色斜体字)来教导大语言模型专注于写作风格和常识性错误,使大语言模型能够在写作细节中找到重要线索并进行常识性错误推理。 推理过程中的关键线索用红色下划线的单词突出显示。

2相关工作

社交媒体上的谣言检测是自动区分不可信信息、维护互联网和社会健康发展的一项重大文本挖掘任务。 人们已经开展了大量工作来从不同角度检测谣言 Castillo 等人 (2011);伊斯兰教等人 (2020)

谣言检测的早期尝试主要集中在提取传播过程的统计特征Kwon 等人 (2013);马等人(2015). 然后,用于建模序列数据的深度神经网络已广泛应用于聚合传播信息和检测社交媒体上的谣言。 其中,基于循环神经网络Ma等人(2016,2018)、卷积神经网络Yu等人(2017)、注意力网络Liu等的检测模型人 (2018) 和生成对抗网络 Ma 等人 (2019) 已被广泛研究。 同时,他们的组合也被用于谣言检测Liu and Wu (2018);于等人(2019) 最近,基于图神经网络的检测模型Bian等人(2020);卢和李(2020); Nguyen 等人 (2020); Xu 等人 (2022) 取得了最先进的表现。 此外,一些工作研究了结合图对比学习以获得更好的谣言检测性能Sun等人(2022b);吴等人(2023) 为了更好的可解释性,提出了子图推理 Jin 等人 (2022);杨等人(2022) 同时,为了捕捉时间特征,构建了动态图并将其应用于谣言检测Sun等人(2022a) 如今,考虑到实际应用,去偏、少样本或零样本场景中的谣言检测备受关注朱等人(2022);徐等人 (2023);岳等人 (2023);林等人 (2023);刘等人(2023)

大语言模型取得巨大成功的同时,也给错误信息识别带来了机遇和挑战Chen and Shu (2023) 大语言模型能够轻松生成错误信息Vykopal 等人 (2023); Chen and Shu (2024),仅根据文本特征很难区分人类或大语言模型Chen and Shu (2024) CICAN Yang 等人 (2023) 使用 GPT 3.5 作为辅助工具,用于从新闻和评论中提取实体,并构建关系图。 ARG Hu 等人 (2024) 使用大语言模型来分析新闻内容,并得出结论,大语言模型在检测假新闻方面很难击败微调的小模型。 因此,ARG 将大语言模型生成的见解融入到用于假新闻检测的小模型中。

Vanilla Prompt There is a piece of news: “Around 7 o’clock this morning, a major traffic accident occurred on the Yuncheng to Houma section of the Shanxi Dayun Expressway, 9.3 kilometers away from Houma. A total of 47 vehicles collided in a series, resulting in 3 deaths and 7 injuries.” There are comments for the news: “My god!” “Blessing…” “Why isn’t the highway closed on such a foggy day?” “Repost” “Cherishing life!” “How scary!” “Blessing.” “Drive carefully and pay attention to safety.” “Shocking…” Based on the comments, verify the authenticity of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output Based on the provided comments, it’s challenging to determine the authenticity of the news directly. The comments express shock, concern, and questioning regarding the incident, but they don’t provide concrete evidence or information about the news source. Therefore, the answer is: A. Fake.
Conflicting Prompt There is a piece of news: “Around 7 o’clock this morning, a major traffic accident occurred on the Yuncheng to Houma section of the Shanxi Dayun Expressway, 9.3 kilometers away from Houma. A total of 47 vehicles collided in a series, resulting in 3 deaths and 7 injuries.” There are comments for the news: “My god!” “Blessing…” “Why isn’t the highway closed on such a foggy day?” “Repost” “Cherishing life!” “How scary!” “Blessing.” “Drive carefully and pay attention to safety.” “Shocking…Based on the comments, analyze whether there are any rebuttals or conflicts, and accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output The comments mainly express shock, concern, and questions about the responsibility for the accident. Some comments criticize the decision not to close the expressway in thick fog, suggesting potential negligence. There is also a call for attention to safety in such weather conditions. There doesn’t seem to be any direct rebuttal or conflict among the comments. Instead, the comments generally focus on expressing emotions, seeking information about the incident, and questioning the decision-making around highway closure during foggy conditions. Based on the analysis of comments, there is no clear evidence of conflicting viewpoints that would cast doubt on the authenticity of the news. The concerns raised by commenters align with the details provided in the news report. Therefore, the answer is: B. Real.
表2: 使用 GPT 3.5 分析示例的评论。 如果没有适当的提示,在缺乏明确支持材料的情况下,大语言模型往往会说该新闻是假的。 然而,在大多数情况下,尤其是在传播的早期阶段,很难有明确的证据。 因此,我们设计了提示(红色斜体字)来教导大语言模型专注于评论中的反驳或冲突,从而引导大语言模型利用群体智慧来判断信息的真实性。 推理过程中的关键线索用红色下划线的单词突出显示。

3方法论

在本节中,我们将详细介绍我们提出的 LeRuD 方法,该方法包括对新闻和评论进行推理的提示,以及传播链推理过程。 LeRuD的整个流程如图3所示。

3.1问题表述

对于社交媒体上传播的一条新闻,我们将新闻内容表示为n,即一段文本。 新闻在传播过程中收到评论,记为C={(c1,t1),(c2,t2),},其中ci表示第i条评论的文本内容,ti表示与原始新闻相比的相对发布时间。 我们的目标是利用大语言模型对新闻n和评论C进行建模,并检测谣言,即识别新闻是假还是真的。

3.2 RP:新闻的理性提示

我们需要设计适当的提示来教授大语言模型分析新闻内容以进行谣言检测。 在这里,我们应用广泛使用的 GPT 3.5 作为我们提出的方法中使用的大语言模型。 如Tab中的两种情况所示。 1,我们尝试了普通的提示,直接要求大语言模型推断新闻内容的可信度。 然而,如果没有适当的提示,大语言模型很难对真实性做出猜测。 大语言模型不知道如何推理内容,因为它不了解所描述的事实。 这主要是因为大语言模型无法集中关注新闻内容中的一些关键线索。

根据之前的一些研究,突出新闻的写作风格对于错误信息识别Przybyla (2020),而常识推理对于区分假新闻胡等人(2024)至关重要。 因此,我们需要从这些方面来教授大语言模型推理。 据此,我们设计了理性提示,要求大语言模型回答“根据写作风格和常识,估计新闻的可信度。”如表​​所示。 1,借助理性提示,大语言模型能够发现写作细节中的重要线索并进行常识推理,并对新闻内容的真实性做出判断。 在第一种情况下,大语言模型通过注意到该事件从常识的角度来看是不合理的,正确地预测了该新闻是假的。 在第二种情况下,大语言模型通过注意到一些事故细节,正确预测了新闻的真实性,并得出写作风格与典型报道一致的结论。

3.3 CP:意见冲突提示

此外,我们还需要设计提示来教大语言模型对传播中的评论进行推理。 如表所示。 2,我们尝试了普通的提示,直接要求大语言模型根据给定的评论来验证新闻的真实性。 然而,在这样的提示下,在缺乏明确支撑材料的情况下,大语言模型往往会预测该新闻是假新闻。 大多数情况下,很难有足够明确的证据,尤其是在传播初期。 这使得大语言模型很容易将真实新闻误认为是假新闻。

评论中的冲突、反驳和怀疑已被证明是区分错误信息的关键特征 Ma 等人 (2016);刘等人 (2018);马等人 (2018);卢和李(2020);于等人 (2017);杨等人 (2022);金等人 (2022). 考虑到缺乏足够支撑材料的困难,我们设计了冲突提示,要求大语言模型回答“根据评论分析是否存在反驳或冲突的情况,据此验证消息的真实性。”如表​​所示。 2,通过冲突提示,大语言模型通过注意到人们都在讨论事件本身,并且没有发布冲突信息,正确预测了新闻是真实的。

Refer to caption
(a) Results on Twitter.
Refer to caption
(b) Results on Weibo.
图2: 大语言模型有太多评论的麻烦。
Refer to caption
图3: LeRuD 方法的过程。 LeRuD将整个传播信息划分为一个传播链,并在每一步中添加简单的注释信息以进行推理。 在每个推理步骤中,都会执行用红色斜体字书写的理性和冲突提示。 LeRuD 推理过程中的主要发现用红色下划线字突出显示。

3.4 CoP:传播链

虽然理性和冲突提示可以教会大语言模型推理检测谣言,但面对太多评论时,大语言模型仍然遇到麻烦。 在真实的社交媒体场景中,评论数量通常非常大。 然而,大语言模型通常对输入长度有限制。 更严重的是,大语言模型不能很好地推理长上下文或冗余信息Huang 等人 (2023);谢(2023) 我们在 Twitter 和微博数据集上使用 GPT 3.5 测试上述理性和冲突提示(第 4.1 节),结果如图 2 所示。 当样本的输入长度超过限制时,我们随机选择属于该样本的部分评论作为输入。 很明显,当样本中的评论过多时,检测精度急剧下降,这意味着大语言模型无法很好地推理过多的信息。

在最近的大语言模型研究中,将一个对大语言模型来说很难的复杂任务分解为一系列简单的任务,让大语言模型来执行Wei等人(2022);王等人 (2023);姚等人 (2023); Besta 等人 (2024) 这增强了大语言模型处理复杂任务的能力。 因此,我们建议将命题信息划分为传播链(CoP),这使得大语言模型更容易进行推理。 考虑到一个时间段内的评论通常具有相似的态度,我们按照评论发布时间的顺序构建CoP,并在每个推理步骤中选择k条评论。 如图2所示,大语言模型最多可以轻松处理100条评论,因此我们设置k=100 CoP中的推理步骤在大语言模型的同一个会话中运行,以便后面的推理步骤可以参考前面的推理步骤的结果。 此外,考虑到最后一步聚合了所有信息,我们使用最后一步的结果作为最终估计。

表3: 性能比较和消融研究的结果。
Approach Twitter Weibo
Accuracy Precision Recall F1-score Accuracy Precision Recall F1-score
BIGCN 0.8882 0.8739 0.8940 0.8838 0.9258 0.9133 0.9335 0.9233
DDGCN 0.8982 0.8834 0.9078 0.8954 0.9414 0.9390 0.9483 0.9436
GACL 0.9042 0.8884 0.9171 0.9025 0.9354 0.9294 0.9409 0.9351
CICAN 0.9082 0.8855 0.9263 0.9054 0.9390 0.9298 0.9458 0.9377
ARG 0.8902 0.8628 0.8986 0.8803 0.9043 0.9022 0.9089 0.9055
LeRuD 0.9401 0.9156 0.9493 0.9321 0.9809 0.9732 0.9852 0.9792
w/o RP 0.9261 0.9107 0.9401 0.9252 0.9689 0.9635 0.9754 0.9694
w/o CP 0.9142 0.8525 0.9585 0.9024 0.9342 0.8777 0.9901 0.9306
w/o CoP 0.9281 0.9148 0.9401 0.9273 0.9330 0.9249 0.9409 0.9328

4实验

在本节中,我们进行了大量的实验,以验证大语言模型是否能够通过我们的努力来检测社交媒体上的谣言。

4.1数据准备

首先,我们介绍数据准备过程,这使我们能够进行充分的评估。

数据集:为了评估,我们将 Twitter15 和 Twitter16 合并为 Twitter 数据集 Ma 等人 (2017),同时还合并了 Weibo 数据集Ma等人(2018) 它们是从两个流行的社交媒体平台收集的,即推特和微博。 Twitter和微博中每条新闻的平均评论数分别为232816 我们仅在 Twitter 中使用经过验证的虚假谣言(假新闻)和非谣言(真实新闻)。

数据过滤:为了顺利且低成本地使用GPT 3.5,我们过滤了存在道德问题(例如暴力、色情和歧视)和重复新闻内容的样本 我们还过滤主要包含主观表达的样本,因为它们实际上并不真实或虚假。 同时,为了公平比较,我们需要过滤掉因使用大语言模型而产生数据泄露风险的样本,其中GPT 3.5已经准确了解了新闻中的事实。 因此,我们迫切要求 GPT 3.5 使用选项卡中的香草提示来验证每条新闻的准确性。 1,并且只保留GPT 3.5不了解事实且无法给出答案的样本。 相比之下,GPT 3.5在了解一条新闻中的事实时,会详细说明事实并明确判断其真实性。

4.2设置

然后,我们详细说明了实验的设置。

基线:首先,我们与一些基于图神经网络的最先进的谣言检测模型进行比较:BIGCN Bian 等人 (2020)DDGCN Sun 等人 (2022a)GACL Sun 等人 (2022b) 其次,我们还与大语言模型的两个谣言检测模型进行了比较:CICAN Yang 等人 (2023)ARG Hu 等人(2024)

指标:我们使用准确度精确度召回率F1-score来衡量性能评估。 根据谣言检测结果计算 Precision、Recall 和 F1-score。

零样本设置:LeRuD在零样本设置中进行,不需要任何数据,对数据集中的所有样本进行评估。 我们使用GPT 3.5作为LeRuD中的大语言模型。

5倍交叉验证设置:对于其他比较基线,我们执行5倍交叉验证。 整个过程运行5次,每次使用3倍样本进行训练,使用1倍样本进行验证, 1 倍样本用于测试。 运行5次后,所有样本均已用于测试,与零样本设置相同。

Prompt There is a piece of news: “Samsung paid Apple one-billion fine, with 20 billion coins in 30 trucks!” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
表 4: 当被要求估计微博数据集中样本“3484403570896812”的准确性时,GPT 3.5 对此一无所知。
Prompt There is a piece of news: “Ex-New York Giants safety Tyler Sash, who died at 27, found to have degenerative brain disease.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output As of my last update in January 2022, the information you provided was accurate. Tyler Sash, a former New York Giants safety, passed away at the age of 27 and was found to have chronic traumatic encephalopathy (CTE), a degenerative brain disease often associated with head trauma in contact sports. Therefore, the answer is: B. Real.
表 5: 当被要求估计 Twitter 数据集中样本“692142338890661888”的准确性时,GPT 3.5 提供了有关事实的详细信息,并确认了该新闻的真实性。

4.3性能比较

在选项卡中。 3,我们将 LeRuD 与几种最先进的谣言检测模型进行比较。 在三种基于图的检测模型中,DDGCN和GACL明显优于BIGCN。 借助 GPT 3.5 提取实体,CICAN 与基于图的模型相比实现了相似或更好的性能。 ARG 仅考虑新闻内容,因此它的性能不如其他同时捕获新闻和评论的检测模型。 然而,根据大语言模型生成的见解,ARG 仍然达到约 90% 的高精度。 同时,我们提出的 LeRuD 方法在所有指标评估的两个数据集上都实现了最佳性能。 具体来说,LeRuD 的性能分别优于 Twitter 和微博上的最佳基线 3.19%3.95% 这些结果有力地证明了LeRuD的优越性。 此外,LeRuD不需要训练数据,因此它可以在少样本或零样本场景中始终获得有希望的谣言检测性能,因此更适合现实世界的谣言检测。

4.4消融研究

在选项卡中。 3,我们还展示了LeRuD的消融研究,包括三种变体:LeRuD w/o RP、LeRuD w/o CP和LeRuD w/o CoP。 LeRuD w/o RP 表示没有 Rational Prompts 的 LeRuD。 它使 LeRuD 的性能降低了大约 1%,但仍然优于其他基线。 LeRuD w/o CP 表示没有冲突提示的 LeRuD。 准确率和 f1-score 的下降幅度为 2.5%4.8% LeRuD w/o CP 实现了最高的召回值,但也实现了最低的精度值。 这是因为,如果没有冲突提示,大语言模型可能会错误地将真实新闻归类为假新闻。 LeRuD w/o CoP 表示没有传播链的 LeRuD。 Twitter 和微博上的准确率和 f1-score 下降幅度分别约为 1%4.5% 根据第二节的统计,微博上的性能下降要大得多,因为微博数据集中的平均评论数要大得多。 4.1 这有力地证明了,大语言模型在过多信息的情况下无法进行推理。 上述观察证明了 LeRuD 中提出的三个模块的必要性。

Refer to caption
(a) Results on Twitter.
Refer to caption
(b) Results on Weibo.
图4: 通过不同检测期限的准确性来评估早期谣言检测结果。

4.5早期检测

如图4所示,我们使用不同的检测期限进行早期谣言检测。 不包括 ARG,因为它不对传播信息建模。 LeRuD 在所有检测期限内始终显着优于所有基线。 这说明,LeRuD在传播初期能够表现良好,并且不依赖于评论信息的充分性。

Refer to caption
(a) Results on Twitter.
Refer to caption
(b) Results on Weibo.
图5: 不同评论数量下 LeRuD 的准确性。 更多评论,更好的表演。

4.6 评论数量的影响

在图5中,我们说明了LeRuD在不同评论数量下的准确性。 在CoP的帮助下,LeRuD在面对大量评论时推理不再困难。 同时,评论越多,LeRuD 的谣言检测准确率就越高,可供分析的信息也就越多。 上述观察结果与图2完全不同。 总之,将传播信息划分为Chain-of-Propagation有助于大语言模型更容易地进行推理。

Prompt of Step 1 There is a piece of news: “Samsung paid Apple one-billion fine, with 20 billion coins in 30 trucks!” There are comments for the news: “Naughty!” “It’s funny!” “Weigh it ” “Samsung has to count first!” “How creativity!” “Verification needed.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the writing style and commonsense knowledge, the news seems highly unlikely and lacks credibility. Paying a fine with 20 billion coins in 30 trucks sounds more like a creative or satirical piece rather than a factual event. Additionally, such a method of payment raises practical concerns and seems impractical. Analyzing the comments, we can observe a mix of skepticism, humor, and requests for verification. Comments like “Naughty!,” “Weigh it,” and “Samsung has to count first!” suggest doubts and a need for verification. The comment “How creativity!” acknowledges the imaginative aspect, and “Verification needed” explicitly calls for confirmation. Therefore, the answer is: A. Fake.
Prompt of Step 2 There is a piece of news: “Samsung paid Apple one-billion fine, with 20 billion coins in 30 trucks!” There are comments for the news: “It’s hard to say whether it is true or not.” “Using $1 billion in cash is illegal.” “Repost” “A truck with a load capacity of 3000 tons is impossible.” “Well done!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 Based on the writing style and commonsense knowledge, the news still appears unlikely and lacks credibility. Paying a one-billion dollar fine with 20 billion coins in 30 trucks seems impractical and raises legal and logistical issues. Analyzing the comments, there are several expressions of doubt and skepticism. Comments like “It’s hard to say whether it is true or not,” “Using $1 billion in cash is illegal,” and “A truck with a load capacity of 3000 tons is impossible” highlight concerns about the feasibility and legality of the described scenario. The comment “Well done!” seems to be sarcastic or ironic, suggesting a lack of belief in the news. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 6: 对微博数据集中的样本“3484403570896812”执行LeRuD方法的提示和输出。 LeRuD 推理过程中的关键线索用红色下划线的单词突出显示。 LeRuD 正确地将样本识别为假样本,从常识角度来看,解释也相当合理。

4.7数据泄露风险

正如在第二节中一样。 4.1,我们过滤了有数据泄露风险的样本,其中GPT 3.5已经准确地了解了新闻中的事实。 有和没有数据泄露风险的示例如表 1 所示。 分别为 45 更多示例可以在 App 中找到。 A 根据这些例子,我们可以清楚地观察到有数据泄露风险和没有数据泄露风险的样本之间大语言模型的输出完全不同 这使得我们很容易区分有和没有数据泄露风险的样本。 也就是说,对这些样本进行过滤后,我们可以避免数据泄露的风险,并且LeRuD和基线之间可以公平地进行实验比较。

4.8说明

我们可以进一步参考大语言模型生成的解释,以确保LeRuD不会利用确切的事实知识来检测谣言。 Tab 中显示了一个示例。 6,更多示例可以在App中找到。 B 根据这些解释,我们可以将LeRuD推理过程中的关键线索总结如下:

书写的规律性:当书写不拘一格,或者语气显得可疑时,LeRuD倾向于判断样本是假的(表1)。 2124)。 当写作是正式的,或者与典型的报告风格一致时,LeRuD 倾向于判断样本是真实的(表 1)。 2527-30)。

细节的充分性:当新闻中缺乏足够的事件细节时,LeRuD倾向于判断样本是假的(表1)。 17212224)。 当细节丰富时,LeRuD 倾向于判断样本是真实的(表 1)。 272831)。

内容的合理性:当内容看起来不切实际,或者内容内部存在冲突,或者事件不符合常识逻辑时,LeRuD倾向于判断样本是假的(表1)。 624)。 当内容从常识的角度来看似乎合理且合理时,或者事件在某些情况下很有可能发生时,LeRuD倾向于判断样本是真实的(表1)。 27 - 29)。

公众的态度:当评论中的态度包含大量的怀疑、质疑和反驳,甚至有的评论直接声称该新闻是假的时,LeRuD倾向于判断样本是假的(表1)。 617 - 24)。 当评论很少表达怀疑或反驳,并且人们正在讨论和评论新闻事件本身时,LeRuD 倾向于判断样本是真实的(表 1)。 25 - 32)。

评论的一致性:当评论中的细节与新闻事件相冲突,或者表明新闻的不合理性时,LeRuD倾向于判断样本是假的(表1)。 6171920)。 当评论中的细节与新闻相符时,LeRuD 倾向于判断样本是真实的(表 1)。 2526)。

综上所述,LeRuD 通过从常识角度进行推理来检测社交媒体上的谣言,没有利用大语言模型对新闻事实的精确先验知识

5结论

在本文中,我们研究使用大语言模型进行社交媒体上的谣言检测,并提出了一种称为 LeRuD 的新方法。 LeRuD设计了适当的提示来教大语言模型推理新闻和评论中的重要线索。 同时,我们将一条新闻的整个传播信息划分为一个传播链(Chain-of-Propagation),引导大语言模型每次以合理的信息量进行逐步推理。 然后,我们使用广泛使用的 GPT 3.5 来实现 LeRud,并在 Twitter 和微博数据集上进行了广泛的实验。 LeRud 可以大幅超越几种最先进的谣言检测模型。 综上所述,通过我们的努力,大语言模型能够很好地检测社交媒体上的谣言。

局限性

在本文中,尽管我们教会了大语言模型来检测社交媒体上的谣言,但我们的工作在以下两个方面仍然存在局限性。 首先,运行大语言模型进行谣言检测需要较高的成本,例如高性能GPU的使用,或者API的调用,或者Web界面的输入。 其次,图神经网络对结构数据建模具有很强的能力,而结构数据很难在文本中得到很好的描述。 因此,使用大语言模型来分析复杂的传播结构,或者将大语言模型和图神经网络的能力结合起来,还没有被探索过。 未来,我们将更加努力克服上述局限性。

参考

  • Besta et al. (2024) Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Michal Podstawski, Hubert Niewiadomski, Piotr Nyczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. In AAAI.
  • Bian et al. (2020) Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, and Junzhou Huang. 2020. Rumor detection on social media with bi-directional graph convolutional networks. In AAAI.
  • Castillo et al. (2011) Carlos Castillo, Marcelo Mendoza, and Barbara Poblete. 2011. Information credibility on twitter. In WWW.
  • Chen and Shu (2023) Canyu Chen and Kai Shu. 2023. Combating misinformation in the age of llms: Opportunities and challenges. arXiv preprint arXiv:2311.05656.
  • Chen and Shu (2024) Canyu Chen and Kai Shu. 2024. Can llm-generated misinformation be detected? In ICLR.
  • Hu et al. (2024) Beizhe Hu, Qiang Sheng, Juan Cao, Yuhui Shi, Yang Li, Danding Wang, and Peng Qi. 2024. Bad actor, good advisor: Exploring the role of large language models in fake news detection. In AAAI.
  • Huang et al. (2023) Yunpeng Huang, Jingwei Xu, Zixu Jiang, Junyu Lai, Zenan Li, Yuan Yao, Taolue Chen, Lijuan Yang, Zhou Xin, and Xiaoxing Ma. 2023. Advancing transformer architecture in long-context large language models: A comprehensive survey. arXiv preprint arXiv:2311.12351.
  • Islam et al. (2020) Md Rafiqul Islam, Shaowu Liu, Xianzhi Wang, and Guandong Xu. 2020. Deep learning for misinformation detection on online social networks: a survey and new perspectives. Social Network Analysis and Mining, 10:1–20.
  • Jin et al. (2022) Yiqiao Jin, Xiting Wang, Ruichao Yang, Yizhou Sun, Wei Wang, Hao Liao, and Xing Xie. 2022. Towards fine-grained reasoning for fake news detection. In AAAI.
  • Kwon et al. (2013) Sejeong Kwon, Meeyoung Cha, Kyomin Jung, Wei Chen, and Yajun Wang. 2013. Prominent features of rumor propagation in online social media. In ICDM.
  • Lin et al. (2023) Hongzhan Lin, Pengyao Yi, Jing Ma, Haiyun Jiang, Ziyang Luo, Shuming Shi, and Ruifang Liu. 2023. Zero-shot rumor detection with propagation structure via prompt learning. In AAAI.
  • Liu et al. (2023) Qiang Liu, Junfei Wu, Shu Wu, and Liang Wang. 2023. Out-of-distribution evidence-aware fake news detection via dual adversarial debiasing. arXiv preprint arXiv:2304.12888.
  • Liu et al. (2018) Qiang Liu, Feng Yu, Shu Wu, and Liang Wang. 2018. Mining significant microblogs for misinformation identification: an attention-based approach. ACM Transactions on Intelligent Systems and Technology (TIST), 9(5):1–20.
  • Liu and Wu (2018) Yang Liu and Yi-Fang Wu. 2018. Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks. In AAAI.
  • Lu and Li (2020) Yi-Ju Lu and Cheng-Te Li. 2020. Gcan: Graph-aware co-attention networks for explainable fake news detection on social media. In ACL.
  • Ma et al. (2016) Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon, Bernard J Jansen, Kam-Fai Wong, and Meeyoung Cha. 2016. Detecting rumors from microblogs with recurrent neural networks. In IJCAI.
  • Ma et al. (2015) Jing Ma, Wei Gao, Zhongyu Wei, Yueming Lu, and Kam-Fai Wong. 2015. Detect rumors using time series of social context information on microblogging websites. In CIKM.
  • Ma et al. (2017) Jing Ma, Wei Gao, and Kam-Fai Wong. 2017. Detect rumors in microblog posts using propagation structure via kernel learning. In ACL.
  • Ma et al. (2018) Jing Ma, Wei Gao, and Kam-Fai Wong. 2018. Rumor detection on twitter with tree-structured recursive neural networks. In ACL.
  • Ma et al. (2019) Jing Ma, Wei Gao, and Kam-Fai Wong. 2019. Detect rumors on twitter by promoting information campaigns with generative adversarial learning. In WWW.
  • Nguyen et al. (2020) Van-Hoang Nguyen, Kazunari Sugiyama, Preslav Nakov, and Min-Yen Kan. 2020. Fang: Leveraging social context for fake news detection using graph representation. In CIKM.
  • Przybyla (2020) Piotr Przybyla. 2020. Capturing the style of fake news. In AAAI.
  • Sun et al. (2022a) Mengzhu Sun, Xi Zhang, Jiaqi Zheng, and Guixiang Ma. 2022a. Ddgcn: Dual dynamic graph convolutional networks for rumor detection on social media. In AAAI.
  • Sun et al. (2022b) Tiening Sun, Zhong Qian, Sujun Dong, Peifeng Li, and Qiaoming Zhu. 2022b. Rumor detection on social media with graph adversarial contrastive learning. In WWW.
  • Vykopal et al. (2023) Ivan Vykopal, Matúš Pikuliak, Ivan Srba, Robert Moro, Dominik Macko, and Maria Bielikova. 2023. Disinformation capabilities of large language models. arXiv preprint arXiv:2311.08838.
  • Wang et al. (2023) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In ICLR.
  • Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS.
  • Wu et al. (2023) Junfei Wu, Weizhi Xu, Qiang Liu, Shu Wu, and Liang Wang. 2023. Adversarial contrastive learning for evidence-aware fake news detection with graph neural networks. IEEE Transactions on Knowledge and Data Engineering (TKDE).
  • Xie (2023) Wenbei Xie. 2023. Analysis of the reasoning with redundant information provided ability of large language models. arXiv preprint arXiv:2310.04039.
  • Xu et al. (2023) Weizhi Xu, Qiang Liu, Shu Wu, and Liang Wang. 2023. Counterfactual debiasing for fact verification. In ACL.
  • Xu et al. (2022) Weizhi Xu, Junfei Wu, Qiang Liu, Shu Wu, and Liang Wang. 2022. Evidence-aware fake news detection with graph neural networks. In WWW.
  • Yang et al. (2023) Chang Yang, Peng Zhang, Wenbo Qiao, Hui Gao, and Jiaming Zhao. 2023. Rumor detection on social media with crowd intelligence and chatgpt-assisted networks. In EMNLP.
  • Yang et al. (2022) Ruichao Yang, Xiting Wang, Yiqiao Jin, Chaozhuo Li, Jianxun Lian, and Xing Xie. 2022. Reinforcement subgraph reasoning for fake news detection. In KDD.
  • Yao et al. (2023) Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. In NeurIPS.
  • Yu et al. (2019) Feng Yu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2019. Attention-based convolutional approach for misinformation identification from massive and noisy microblog posts. Computers and Security, 83:106–121.
  • Yu et al. (2017) Feng Yu, Qiang Liu, Shu Wu, Liang Wang, Tieniu Tan, et al. 2017. A convolutional approach for misinformation identification. In IJCAI.
  • Yue et al. (2023) Zhenrui Yue, Huimin Zeng, Yang Zhang, Lanyu Shang, and Dong Wang. 2023. Metaadapt: Domain adaptive few-shot misinformation detection via meta learning. In ACL.
  • Zhao et al. (2023) Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.
  • Zhu et al. (2022) Yongchun Zhu, Qiang Sheng, Juan Cao, Shuokai Li, Danding Wang, and Fuzhen Zhuang. 2022. Generalizing to the future: Mitigating entity bias in fake news detection. In SIGIR.
表 7: Twitter和微博数据集上GPT 3.5已经了解新闻事实的样本中存在数据泄露风险的样本比例。
Twitter Weibo
Ratio of Data Leakage 0.1880 0.0434

附录A数据泄露风险示例

在选项卡中。 7,我们计算有数据泄露风险的样本,即需要过滤的样本的比例。 Twitter数据集上的比例远大于微博数据集上的比例。 这可能是由训练 GPT 3.5 使用的数据源引起的。 英文新闻已被广泛收集,而中文新闻很少用于 GPT 3.5 的训练过程。

此外,我们进一步说明了几个样本,其中一些是 GPT 3.5 确切知道的事实,而另一些则是 GPT 3.5 不知道的。 在选项卡中。 8-10,存在GPT 3.5不了解事实的样本,大语言模型无法对新闻的真实性给出答案。 在选项卡中。 11-13,大语言模型对事实进行了验证描述,确认其为假新闻。 在选项卡中。 14-16,大语言模型提供了更多有关事实的细节,并确认它们是真实的新闻。

根据这些例子,存在数据泄露风险的样本主要有两类:(1)实时新闻,已记录在训练GPT 3.5的数据中;(2)实时新闻。 (2)非实时常识性知识,主要集中在健康、历史等领域。 在实际的在线谣言检测系统中,非实时知识(类型2)实际上对检测过程有帮助,而实时知识(类型1)很难被大语言模型及时记住。

附录 B LeRuD 示例

在本节中,我们在 Tab 中提供了 LeRuD 运行结果的更多示例。 17 - 32

Prompt There is a piece of news: “Famous actor Morgan Freeman passed away in Los Angeles due to a heart attack.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
表8: 微博数据集中的样本“3547977337564354”。 GPT 3.5 不知道这个事实。
Prompt There is a piece of news: “Top North Korean diplomat who disappeared from London embassy has defected to South Korea, Seoul.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
表 9: Twitter 数据集中的示例“765859710503378944”。 GPT 3.5 不知道这个事实。
Prompt There is a piece of news: “ALERT: Tsunami advisory issued for Southern California from 3-7 a.m. local time. Stay away from ocean and shoreline.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output I cannot verify the credibility of real-time or specific news events as my training only includes information up until January 2022, and I do not have the capability to access the internet for real-time updates.
表 10: Twitter 数据集中的示例“765859710503378944”。 GPT 3.5 不知道这个事实。
Prompt There is a piece of news: “GMO foods weaken reproductive ability. Eating them mean chronic suicide.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output As of my last knowledge update in January 2022, there is no widespread scientific consensus that GMO (Genetically Modified Organism) foods weaken reproductive ability or lead to chronic suicide. Numerous studies have been conducted on the safety of GMOs, and the majority of scientific organizations, including the World Health Organization (WHO) and the National Academy of Sciences, have stated that GMOs currently approved for sale are safe to eat. Therefore, the answer is: A. Fake.
表 11: 微博数据集中的样本“8535326047”。 GPT 3.5确认该消息是假的。
Prompt There is a piece of news: “The screen car god Paul Walker was not dead and has been rescued.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output As of my last knowledge update in January 2022, Paul Walker, the actor known for his role in the "Fast and Furious" film series, passed away in a car accident on November 30, 2013. Any news claiming that he has been rescued and is not dead would be false. Therefore, the answer is: A. Fake.
表 12: 微博数据集中的样本“3651171602494589”。 GPT 3.5确认该消息是假的。
Prompt There is a piece of news: “NASA Confirms Earth Will Experience 15 Days Of Complete Darkness.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output The news stating "NASA Confirms Earth Will Experience 15 Days Of Complete Darkness" is fake. There is no credible information or scientific basis to support such a claim. If this news were true, it would have had widespread and legitimate coverage from reputable sources. Therefore, the answer is: A. Fake.
表 13: Twitter 数据集中的示例“620916279608651776”。 GPT 3.5确认该消息是假的。
Prompt There is a piece of news: “The Lushun Massacre, an unforgettable Japanese crime! After the Japanese army captured the fortress of Lushun, they carried out inhumane robbery, massacre, and rape inside the city. Over four days and three nights, approximately 20000 innocent civilians were brutally killed by the Japanese army’s knives.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output The Lushun Massacre did occur, and the information in the news is generally accurate. Lushun, also known as Port Arthur, was a Chinese fortress that was captured by the Japanese army during the First Sino-Japanese War in 1894. There were reports of inhumane acts, including robbery, massacre, and rape, during the Japanese occupation of the city. The details may vary, and historical accounts might differ, but the basic premise of the massacre is real. Therefore, the answer is: B. Real.
表 14: 微博数据集中的样本“3911526290906659”。 GPT 3.5证实了该消息的真实性。
Prompt There is a piece of news: “Colombian government and left-wing Farc rebel movement agree ceasefire to end decades-long conflict.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output As of my last knowledge update in January 2022, a historic peace agreement between the Colombian government and the Revolutionary Armed Forces of Colombia (FARC) was indeed reached in 2016. The peace deal aimed to end the decades-long armed conflict in Colombia. The agreement included a ceasefire and paved the way for negotiations on various issues, such as disarmament, political participation, and rural development. Therefore, the answer is: B. Real.
表 15: Twitter 数据集中的示例“689529955294777344”。 GPT 3.5证实了该消息的真实性。
Prompt There is a piece of news: “RIP Kenny Baker, the actor behind legendary Star Wars droid R2-D2.” Verify the credibility of the news. Please choose the answer from the following options: A. Fake, B. Real.
Output As of my last knowledge update in January 2022, Kenny Baker, the actor who portrayed R2-D2 in the Star Wars films, passed away in 2016. Therefore, the answer is: B. Real.
表 16: Twitter 数据集中的示例“764505291853627392”。 GPT 3.5证实了该消息的真实性。
Prompt of Step 1 There is a piece of news: “The State Council will officially approve Shanghai’s request to abolish the establishment of Jing’an District and divide the original administrative regions of Jing’an District into Changning District and Huangpu District in the near future.” There are comments for the news: “Changning District has expanded, and we hope for the great development of Changning culture!!!” “But I really like Jing’an District, the most beautiful district name, don’t merge it!” “Really?” “I don’t understand, but it seems reasonable.” “No way, right?!” “Is Jing’an District going to change its past tense???” “Didn’t we already refute the rumors a few months ago” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news seems to be presented as a statement of fact, mentioning that the State Council will officially approve Shanghai’s request to abolish the establishment of Jing’an District. However, it lacks specific details or sources. The use of phrases like “in the near future” might suggest a lack of concrete information or immediacy. The comments reflect a variety of opinions, ranging from support for the development of Changning District to concerns about losing the name “Jing’an District.” Some comments express confusion or skepticism about the news, indicating a lack of clarity or consensus among the readers. The comment mentioning the refutation of rumors a few months ago raises questions about the consistency of the information. Therefore, the answer is: A. Fake.
Prompt of Step 2 There is a piece of news: “The State Council will officially approve Shanghai’s request to abolish the establishment of Jing’an District and divide the original administrative regions of Jing’an District into Changning District and Huangpu District in the near future.” There are comments for the news: “No I object” “Why merge?” “At that time, the government even stepped forward to refute rumors” “I saw this post three months ago.” “I have become a resident of Changning District” “What is the significance of it” “Such fake news would make so many people believe it!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 The news is presented as a statement, but it lacks specific details or sources, making it challenging to verify its accuracy. Phrases like “in the near futur” suggest a lack of immediacy and concrete information. The comments include objections, questions about the merger, references to previous government refutations, and skepticism about the news, indicating potential credibility issues. Several comments express objections, questioning the reasons behind the merger and citing past government refutations. A comment mentions having seen a similar post three months ago, suggesting a lack of novelty or consistency in the information. One commenter claims to have become a resident of Changning District, potentially providing a firsthand perspective. Therefore, the answer is: A. Fake.
Prompt of Step 3 There is a piece of news: “The State Council will officially approve Shanghai’s request to abolish the establishment of Jing’an District and divide the original administrative regions of Jing’an District into Changning District and Huangpu District in the near future.” There are comments for the news: “Jing’an District is going to become history” “The trend is inevitable, we’ll just wait for time.” “Well, I’ve heard of this specific plan, but the problem is it’s already a month now, how can we make it.” “Really?” “Fake, at least currently fake.” “This is not true, is it…” “Nonsense.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 The news is presented as a statement, asserting that the State Council will officially approve Shanghai’s request to abolish Jing’an District and divide it into Changning District and Huangpu District. Phrases like “Jing’an District is going to become history” and “The trend is inevitable” suggest a certain level of acceptance or resignation. Comments expressing doubts like “Fake, at least currently fake” and “This is not true, is it…” indicate skepticism. The comment labeling it as “Fake, at least currently fake” and others expressing disbelief or calling it “Nonsense” add weight to doubts about the authenticity. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 17: 微博数据集中的样本“3498854609883720”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “Famous actor Morgan Freeman passed away in Los Angeles at 8am local time on Tuesday, February 19th due to heart disease.” There are comments for the news: “A great actor!” “Blessing.” “So suddenly.” “One of my favorite actors…” “Oh my goodness, I really like him!” “A great actor” “I hope it’s fake news!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news about Morgan Freeman’s death seems credible based on the provided information. The details include the specific time, cause of death, and location, which adds a level of specificity often found in legitimate news reports. Additionally, the news is written in a straightforward and factual manner, contributing to its credibility. The comments express sentiments of admiration and surprise, with no apparent rebuttals or conflicts. However, the comment "I hope it’s fake news!" suggests some skepticism. It’s important to note that public figures’ deaths can sometimes be subject to rumors or hoaxes, but without additional information, it’s challenging to determine the veracity solely based on the comments. Therefore, the answer is: B. Real.
Prompt of Step 2 There is a piece of news: “Famous actor Morgan Freeman passed away in Los Angeles at 8am local time on Tuesday, February 19th due to heart disease.” There are comments for the news: “Sincerely hope it’s spreading rumors” “Really? The list of last wishes is also funny.” “I was deceived! Rumors!” “I can’t believe it!” “I will continue to watch your movies.” “Is it fake news? Has someone refuted the rumor?” “Morgan Freeman is said to be died again?” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 The news mentions specific details like the time, cause of death, and location, which lends an air of credibility. However, there is a mix of reactions in the comments, with some expressing doubt and questioning the authenticity of the news. Some comments express skepticism or hope that it’s a rumor, indicating potential disbelief in the news. The comment “Morgan Freeman is said to be died again?” directly questions the veracity of the news, indicating a potential conflict or doubt. Therefore, the answer is: A. Fake.
Prompt of Step 3 There is a piece of news: “Famous actor Morgan Freeman passed away in Los Angeles at 8am local time on Tuesday, February 19th due to heart disease.” There are comments for the news: “Fake news!” “I’sorry to hear that…” “Why do so many people like to say that others have died ” “Just reposted it and realized it’s fake news!!!” “It has been identified as fake news, please don’t believe it” “RIP.” “This has been confirmed as a rumor.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 The presence of comments like "Fake news!" and "It has been identified as fake news, please don’t believe it" strongly suggests a lack of credibility. The comment expressing realization that it’s fake after reposting also adds to the skepticism surrounding the news. The majority of comments explicitly label the news as fake and provide reasons for their skepticism. Phrases like "confirmed as a rumor" and "identified as fake news" indicate clear rebuttals and conflicts with the authenticity of the news. There is an absence of supportive comments or credible sources affirming the news. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 18: 微博数据集中的样本“3547977337564354”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “Today, Shanghai Metro Line 10 experienced a malfunction, and the two trains successfully merged, which is truly terrifying.” There are comments for the news: “What’s wrong with Line 10 again,….” “When? I also took Line 10 today” “What? How was this achieved?” “Really?” “It’s not surprising anymore” “This is really funny.” “I think it’s the same as the last picture. Is it true?” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news itself seems unusual, as it claims that two trains on Shanghai Metro Line 10 successfully merged due to a malfunction. This scenario raises questions about the technical feasibility and safety of such an event. The comments express a mix of skepticism, concern, and curiosity. Some users question the reliability of Line 10, while others seek more information about the incident. The range of reactions suggests a level of uncertainty among the audience. Without additional details or official sources confirming the incident, it’s difficult to verify the authenticity of the news. Therefore, the answer is: A. Fake.
Prompt of Step 2 There is a piece of news: “Today, Shanghai Metro Line 10 experienced a malfunction, and the two trains successfully merged, which is truly terrifying.” There are comments for the news: “Don’t use the word terror indiscriminately, it’s just that the train ahead is malfunctioning. The train behind will push forward for rescue, connect the malfunctioning train, and push it to the designated location to restore the main line operation as soon as possible! This is just a commonly used train rescue method in the subway, there is no such thing as horror.” “Terrifying… Marketing Weibo attracts attention from the public” “I haven’t seen it either…” “Scared me…” “No knowledge, it’s really scary!” “Whoever spreads rumors is terrifying!” “Iron Brave Train Warrior, transform!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 The comment that explains the situation mentions that the merging of trains is a commonly used rescue method in the subway when there is a malfunction. This comment provides a plausible explanation for the incident, suggesting that it might not be as terrifying as the initial news headline implies. Other comments express skepticism or confusion, with some users suggesting that the use of the word “terrifying” might be for marketing purposes or to attract public attention. The explanation provided in one of the comments aligns with common subway rescue procedures, making the incident more plausible and less alarming. The skepticism expressed in some comments, coupled with the suggestion of marketing tactics, raises questions about the accuracy and sensationalism of the initial news report. Therefore, the answer is: A. Fake.
Prompt of Step 3 There is a piece of news: “Today, Shanghai Metro Line 10 experienced a malfunction, and the two trains successfully merged, which is truly terrifying.” There are comments for the news: “It’s too scary. It seems like we can’t sit on Line 10.” “Recently, the subway has been broken down frequently.” “I personally believe that facing this kind of marketing and Weibo popularization, which takes spreading rumors as its own responsibility, is futile no matter how much it is.” “Please allow me to scold you. Such a simple matter is like a car malfunctioning and being towed back for repair. I’m amazed at your understanding.” “This is a very normal rescue method, don’t use the word terror in the title party.” “Really?” “Rescue! make a big fuss!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 Comments express a mix of concern, skepticism, and criticism toward the news. Some users find the news terrifying and express reluctance to use Line 10, while others criticize the use of the word “terror” and suggest that it’s a normal rescue method. A few comments provide a more rational perspective, likening the situation to a car malfunction being towed for repair, and expressing amazement at the mis- understanding. The comments present a diversity of opinions, with some users expressing genuine concern and others debunking the severity of the situation. Rational explanations in the comments suggest that the incident might be a routine and safe subway rescue operation, contrary to the alarming tone of the news. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 19: 微博数据集中的样本“3465261540071546”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “NHL expansion – four teams added, Quebec City, Toronto, Seattle, and Las Vegas $1.4b in expansion fees” There are comments for the news: “Short term thinking… uggh” “Phoenix is not moving and sadly neither is Miami” “Hopefully this also helps Seattle get an NBA team back.” “Is this done deal? Or is this just speculation? Thank you” “Has 3-4 yrs to build up talent. Can be handled if KHL stops taking guys.” “Seattle has one Stanley Cup already. Woot! Lets get another.” “YES!!! They need a facility though!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The writing style of the news seems plausible, mentioning specific details such as the addition of four NHL teams (Quebec City, Toronto, Seattle, and Las Vegas) and the associated $1.4 billion in expansion fees. The information appears coherent and consistent, suggesting a level of credibility. However, without additional sources or context, it’s challenging to make a definitive judgment. The comments reflect a mix of opinions and concerns. Some express excitement about the expansion, while others express skepticism or raise questions about specific locations like Phoenix and Miami. The comment “Has 3-4 yrs to build up talent. Can be handled if KHL stops taking guys”implies considerations about player talent development and potential challenges related to the KHL (Kontinental Hockey League). There is no direct rebuttal or conflict in the comments that would immediately discredit the news. Therefore, the answer is: B. Real.
Prompt of Step 2 There is a piece of news: “NHL expansion – four teams added, Quebec City, Toronto, Seattle, and Las Vegas $1.4b in expansion fees” There are comments for the news: “you think the Seattle City Council will go with an NHL 1st option for the Sodo Arena? I doubt it but I hope so” “I don’t take him seriously. Also implies that Seattle can’t have NHL AND NBA… gotta be one or the other.” “if this is true then wouls arena be built prior to sea getting a team?” “Bettman said himself a few weeks ago that Seattle not being considered right now. If they build an arena? Sure. But not before that.” “Seriously?” “Toronto has a team…..” “there will be an imbalance in schedule and conferences/divisions. There’s no way.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 The news mentions the expansion of four NHL teams (Quebec City, Toronto, Seattle, and Las Vegas), along with the associated $1.4 billion in expansion fees. The comments express skepticism and doubts about the news. One comment questions whether the Seattle City Council would prioritize an NHL option for the Sodo Arena, and another doubts the credibility of the source. The comment “Toronto has a team…..” points out a factual error in the news, as Toronto already had an NHL team before the mentioned expansion. These comments indicate a level of suspicion and raise valid concerns about the accuracy of the news. Therefore, the answer is: A. Fake.
Prompt of Step 3 There is a piece of news: “NHL expansion – four teams added, Quebec City, Toronto, Seattle, and Las Vegas $1.4b in expansion fees” There are comments for the news: “I understand. But the Maple Leafs have been there for ages. Who will not stay on as a fan for them?” “Seriously, i think we can skip Toronto. One is enough” “I don’t know why they’d send another team to Toronto. The broke the first one. Awful organization.” “Doubt it” “why? we don’t have the market size.” “source? Hallucination?” “You’re saying this and the #NHL is saying no. Which one of you is lying?” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 The news mentions the expansion of four NHL teams (Quebec City, Toronto, Seattle, and Las Vegas) with $1.4 billion in expansion fees. The comments express skepticism and doubts about the news. Multiple comments question the inclusion of Toronto in the expansion, citing the Maple Leafs’ long-standing presence in the NHL. There are also negative opinions about sending another team to Toronto, with comments criticizing the organization or expressing doubts about the market size. The comment “You’re saying this and the NHL is saying no. Which one of you is lying?” directly challenges the credibility of the news source. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 20: Twitter 数据集中的示例“504448356584677377”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “Months ago, Rupert Murdoch bought National Geographic and fired most of writers. This just hit the stands.” There are comments for the news: “Hmm. Doesn’t look like the kind of cover one wd expect from the NG. And I don’t mean that in a good way.” “is there anything he doesn’t ruin?!” “this is something sad to learn of. NatGeo was one of the last.” “This isn’t true, it’s an old cover that’s doing the rounds” “noooooooo! That’s so dumb.” “I cancelled my NG subscription years ago. The writing then wasn’t much better.” “Murdock keeping his stupid sheep in line with ANYTHING but Truth! If there is a hell he should rot there!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the writing style and common sense knowledge, it’s important to note that the news lacks specific details such as the source, date, and specific actions taken by Rupert Murdoch. The language used in the news is somewhat informal and lacks the journalistic tone often found in credible news sources. Additionally, the absence of concrete information about the event raises concerns about its credibility. Analyzing the comments, there is a mix of reactions. Some express skepticism, questioning the authenticity of the news, while others react emotionally with disappointment and anger. One comment even suggests that the news might be based on an old cover that is circulating. The varied reactions indicate a lack of consensus among readers. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 21: Twitter 数据集中的示例“676120162018451456”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “Mike Brown murder trial witness, #DeAndreJoshua fatally shot and set on fire after verdict.” There are comments for the news: “This is ridiculous. It’s time we take a stand” “Your article doesn’t say anything about him being a witness. Can you link to that fact?” “The article doesn’t mention he was a witness, got a source? Let the speculation begin.” “Speechless.” “Google his name and read some articles. Few witnesses gave testimony for fear of this happening.” “ I had initially done so, however was unable to locate anything substantial.. Only speculation.” “I’m sure the witnesses were told to keep quiet to avoid situations like this. Not sure if any will be confirmed” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news report mentions a shocking incident involving the murder trial witness, DeAndre Joshua, being fatally shot and set on fire after the verdict. The writing style doesn’t provide specific details about the incident or cite any credible sources. The information lacks concrete evidence and seems sensational- ized, which raises concerns about its credibility. The comments highlight skepticism and a lack of con- crete evidence or sources. Some comments express skepticism and request sources to confirm the claim that DeAndre Joshua was a witness. One comment mentions that few witnesses gave testimony for fear of such incidents, indicating a potential environment of intimidation. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 22: Twitter 数据集中的示例“538412662858588160”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “Jay Z and Beyoncé want to buy rights to the Confederate flag” There are comments for the news: “this isn’t true. lol. do better, Complex.” “huh???” “lies!!” “wut? RT Jay Z and Beyoncé want to buy rights to the Confederate” “When you got it RT Jay and Beyoncé want to buy rights to the Confederate flag” “idek wut dis mean RT Jay Z and Beyoncé want to buy rights to the Confederate flag” “Whatever man. The stock exchange just halted trading…screw that flag” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The writing style of the comments suggests skepticism and confusion regarding the news. Phrases like “lol,” “huh???” and “wut?” indicate a lack of belief or understanding of the information. Additionally, some comments express disbelief or question the authenticity of the news, such as “lies!!” and “do better, Complex.” Overall, the tone of the comments leans towards skepticism. The comments contain a mix of reactions, with some expressing doubt or disbelief, while others seem to accept or support the news. It’s essential to note that comments alone may not provide a definitive answer to the authenticity of the news. However, the skepticism expressed in some comments suggests that there might be conflicts or differing opinions among the audience. Based on the above analysis, the news appears to be more likely fake or at least questionable. The comments reflect a degree of doubt and confusion, pointing towards a lack of credibility. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 23: Twitter 数据集中的示例“618805892222468096”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “The Panjin Public Security Bureau is investigating itself again. Villager Wang Shujie was shot and killed by police officer Zhang Yan, but the investigation found that the police shot legally! Li Hongna was gang raped by the director of Panjin Police Station and local ruffians. Faced with 30000 reposts from netizens, the Panjin Public Security Bureau couldn’t sit still and responded to netizens by posting Weibo messages online, urging everyone to be patient and wait for the investigation results. After so many days, the girl’s two distress messages were deleted. Where did you lock her up?” There are comments for the news: “Repost” “Seeking the truth.” “No wonder I always feel so familiar when I hear the place name Panjin while watching a movie today!” “It is only natural that Wang Shujie was killed. He first illegally injured the construction workers, then attacked the police with a knife and attempted to attack even after the police fired three shots into the sky. Li Hongna’s case has been investigated for a long time, and it was a false accusation. All the lies she said were unfounded. Li is a drug addict and has been falsely accused by the police on Weibo.” “angry” “angry” “Panjin Public Security is in chaos.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news contains conflicting information and seems to be presented in a sensationalized manner. The tone of the news raises suspicions, and the use of phrases like “investigating itself again” and “couldn’t sit still” may indicate bias. Additionally, the comments show a range of opinions, including both support and skepticism. Without further evidence or reliable sources, it is challenging to determine the credibility of the news. However, the news seems to lack specific details and relies heavily on emotional language, which could be indicative of a less credible source. The comments reflect a mix of opinions, with some expressing skepticism and others defending the actions of the Panjin Public Security Bureau. The conflicting narratives in the comments suggest a lack of consensus among the readers. Some comments question the legitimacy of the claims, while others show support for the authorities. There is no clear consensus among the commenters. Therefore, the answer is: A. Fake.
Final Output The final estimation of the veracity of the news is: The news is fake.
表 24: 微博数据集中的样本“3507985437364629”。 LeRuD 正确地将其预测为假新闻。
Prompt of Step 1 There is a piece of news: “The Paris attack has sounded a global anti-terrorism alert.” There are comments for the news: “Shocked, just heard!” “The cause of terrorist attacks is not due to the general public, but they bear the tragic consequences” “Only wish for world peace!” “Terrorism will always be the common enemy of all humanity, and any violent attacks against civilians should be condemned and countered. Mourning for Paris.” “The road to counter-terrorism is still long, and there is still much we need to do.” “Rest in peace for the deceased.” “Civilian people are innocent.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the information provided, it’s challenging to directly estimate the credibility of the news. The statement “The Paris attack has sounded a global anti-terrorism alert” is quite general and lacks specific details about the incident. Credibility assessment would require more information, such as the news source, additional context, and corroboration from reliable sources. Analyzing the comments, there is no evidence of rebuttals or conflicts regarding the authenticity of the news. The comments express shock, sympathy, condemnation of terrorism, and a desire for world peace. The comments are more focused on expressing emotions and opinions rather than disputing the news itself. So the answer is undetermined.
Prompt of Step 2 There is a piece of news: “The Paris attack has sounded a global anti-terrorism alert.” There are comments for the news: “Watching the report and listening to gunshots, my throat and eyes tightened, my eyes became sore, and everyone’s eyes were filled with helplessness, filled with a sense of terror everywhere. Counter terrorism is urgent and cannot be ignored.” “The world needs to work together to strike.” “Let’s pray.” “Every country should attach importance to terrorist attacks and take them as a lesson.” “Pray that this romantic city will no longer be harmed.” “The Paris attack has sounded a global anti-terrorism alert.” “May the deceased rest in peace! World peace!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 Based on the provided information, the news seems to be credible and reflects a serious incident – a terrorist attack in Paris. The writing style is coherent and consistent with typical news reports on such events. The comments express shock, sadness, and a call for global unity against terrorism, which aligns with the expected public response to such incidents. There are no apparent rebuttals or conflicts in the comments; rather, they convey a shared sentiment of concern, grief for the victims, and a collective call for action against terrorism. The comments do not raise doubts about the authenticity of the news but rather emphasize the need for global cooperation and solidarity. Therefore, the answer is: B. Real.
Prompt of Step 3 There is a piece of news: “The Paris attack has sounded a global anti-terrorism alert.” There are comments for the news: “May everyone in the world have a beautiful heart, and may there be no more disasters in the world.” “Is it not good to get along well with others?” “The casualties are continuing to expand, and the most heartbreaking thing is that countries are still debating countermeasures.” “Blessing Paris, may the world be at peace.” “I still don’t want to go out. It feels unsafe everywhere abroad.” “No matter what, innocent civilians should not be harmed.” “Why are these terrorists so terrifying? May the world be peaceful, please let go of the kind and innocent people.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 Based on the provided information, the news appears to be credible as it describes a serious event – a terrorist attack in Paris. The writing style is consistent with typical news reporting on such incidents. Analyzing the comments, there are no evident rebuttals or conflicts. Instead, the comments express concerns, sympathy, and a desire for peace. They highlight the impact of the attack on innocent civilians and call for measures to ensure safety. These sentiments align with the expected public response to a tragic event like a terrorist attack. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 25: 微博数据集中的样本“3909084371985775”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “27 years later, she found a firefighter who sacrificed her life to save her.” There are comments for the news: “They are all good examples.” “Moved!” “Give a thumbs up to the firefighting hero, and also give a thumbs up to the girl’s gratitude!” “A life-saving benefactor. Although it is his responsibility, I also want to express my gratitude.” “Repost” “Positive energy!” “Diligence and understanding gratitude are both good virtues!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the information provided, it’s challenging to definitively determine the credibility of the news without additional context or specific details about the source. However, there are some positive comments expressing emotions like being moved, giving thumbs up, and acknowledging the heroism of the firefigh- ter and the gratitude of the girl. These comments generally support the authenticity of the news. Without any specific details or evidence to suggest otherwise, it’s reasonable to lean towards a more positive assessment. Therefore, considering the writing style and the overall positive sentiment in the comments, the news appears to be more likely real than fake. Therefore, the answer is: B. Real.
Prompt of Step 2 There is a piece of news: “27 years later, she found a firefighter who sacrificed her life to save her.” There are comments for the news: “Good people must have good karma!” “With a grateful heart.” “A good girl who knows how to be grateful, also has a good family education.” “This spirit is worthy of encouragement, and a grateful heart will never be erased. The transmission and promotion of positive energy will not stop.” “Being a firefighter for 27 years is not easy!” “Sincere gratitude and gratitude in the heart are sometimes more precious than money!” “A grateful person will have a good life.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 Based on the information provided, the writing style and comments generally seem positive and suppor- tive of the news. The comments express sentiments of gratitude, encouragement, and recognition of the challenges faced by firefighters. There are no apparent signs of rebuttals or conflicts in the comments. Given the positive nature of both the news and the comments, and without specific details suggesting otherwise, it appears more likely that the news is real. Therefore, the answer is: B. Real.
Prompt of Step 3 There is a piece of news: “27 years later, she found a firefighter who sacrificed her life to save her.” There are comments for the news: “A servant of the people, a good firefighter who does not seek anything in return; For 27 years, I will never forget to search for benefactors. My parents have taught me well, and I am grateful for the people and things around me!” “So warm, Uncle Firefighter looks so kind.” “Those who show gratitude are all good people.” “The heroic product of sacrificing oneself to save babies, the true emotions in the world are deeply moving!” “Sun Bingjing visited the fire brigade, and the retired squad leader inferred that he must have rich experience in rescuing people in a dangerous fire scene. Eventually, he found the 27 year old squad leader, Lu Yuming.” “Never forget the kindness of saving lives.” “A good person lives a peaceful life.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 Based on the information provided, the writing style and comments continue to convey positive sentiments, expressing gratitude, admiration for the firefighter, and recognition of the importance of showing appreciation. There is a specific mention of Sun Bingjing visiting the fire brigade and finding the 27-year-old squad leader, Lu Yuming, which adds detail and specificity to the story. There are no apparent signs of rebuttals or conflicts in the comments, and the tone remains supportive and positive. Considering the positive writing style, lack of conflicts in comments, and the specific details mentioned, the news appears more likely to be real. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 26: 微博数据集中的样本“3912024620676243”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “The Chinese Embassy in Mali has confirmed that three employees of China Railway Construction Corporation (CRCC) who went to Mali to negotiate a project with the Ministry of Transportation have died in the attack. They are Zhou Tianxiang, General Manager of China Railway Construction International Group, Wang Xuanshang, Deputy General Manager of China Railway Construction International Group, and Chang Xuehui, General Manager of China Railway Construction International Group West Africa Company.” There are comments for the news: “Pray” “There are risks overseas.” “One of them met a year ago while eating in Cote, and their lives were so rushed. They were all people who struggled hard in Africa.” “I’m too sad, I hope my compatriots can rest in peace soon.” “The three Chinese citizens who died in the Mali hotel attack are employees of China Railway Construction Corporation!” “The deceased rest in peace.” “The risk of enterprises going global is increasing, and our responsibility to protect citizens overseas is becoming heavier. Our overseas military strike capabilities need to be greatly strengthened.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news appears to be written in a straightforward and factual manner, which adds to its credibility. The details about the individuals involved, their positions, and the circumstances of the incident provide a sense of specificity. The news itself doesn’t seem to conflict with common knowledge. Attacks and security risks can occur in various regions, including those with ongoing geopolitical tensions or internal conflicts. The comments express sympathy, sorrow, and concern for the individuals involved. There are no apparent rebuttals or conflicts among the comments, which suggests a collective acknowledgment of the tragic event. Therefore, the answer is: B. Real.
Prompt of Step 2 There is a piece of news: “The Chinese Embassy in Mali has confirmed that three employees of China Railway Construction Corporation (CRCC) who went to Mali to negotiate a project with the Ministry of Transportation have died in the attack. They are Zhou Tianxiang, General Manager of China Railway Construction International Group, Wang Xuanshang, Deputy General Manager of China Railway Construction International Group, and Chang Xuehui, General Manager of China Railway Construction International Group West Africa Company.” There are comments for the news: “Oh my, work is so dangerous.” “I used to think these things were so far away from me. Since my brother went to Africa. Every time I see such news, my heart beats tightly, hoping for world peace!!!!” “Three elites, three families.” “Sadness, it’s all corporate executives, keep going!” “The reason why we cannot see darkness is because someone is doing their best to block it from where we cannot see. All warriors who fight against terrorist forces, regardless of nationality, faction, or stance, should be remembered by us!” “I saw such heartbreaking news early in the morning!” “Mourning for the deceased! Perhaps the child loses the father, the lover loses the husband, and the white haired person gives the black haired person away.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 The news is presented in a straightforward and factual manner. The details about the victims and the circumstances of the incident are specific, contributing to the credibility of the news. The idea that work in certain regions, especially those with geopolitical tensions or security concerns, can be dangerous ali- gns with common knowledge. The news about individuals associated with a construction corporation facing risks during overseas negotiations is plausible. The comments express empathy, concern, and sadness for the victims and their families. There are no apparent rebuttals or conflicts in the comments, suggesting a shared sentiment among the commenters regarding the tragic event. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 27: 微博数据集中的样本“3911539469801102”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “Man armed with knife on Swiss train starts fire - seven people hospitalised with knife wounds and burns, police say.” There are comments for the news: “Mental Illness ?” “They put on my wall and the nearby Antonia and neighboring Edvaldo one boma H and obvious that I wiretaps giant recorders” “hope is not wat am thinking” “To Germany he was working well with the H-bomb was the one who was present in the second war with his amateur radio.” “please not another Machete!” “ i don’t even care anymore, terrorists is now a daily thing.” “horrible attack.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the provided information, it’s challenging to directly assess the credibility of the news as there are no specific details about the source, and the content is limited to a brief summary. However, the news itself doesn’t inherently raise red flags that would suggest it’s fake. News about violent incidents can occur, and it is not uncommon for such events to be reported. Analyzing the comments, it seems there is a mix of reactions ranging from speculation about mental illness, conspiracy theories about wiretaps, concerns about terrorism becoming a daily occurrence, and references to historical events. These comments do not provide concrete information about the incident itself but rather express various opinions and thoug- hts from different perspectives. Therefore, the answer is: B. Real.
Prompt of Step 2 There is a piece of news: “Man armed with knife on Swiss train starts fire - seven people hospitalised with knife wounds and burns, police say.” There are comments for the news: “there are bad people in every ethical community” “bloody hell, this is awful :( ” “hope is not wat am thinking” “worst killers are mentally ill or have problems in their personal environment” “yes i am afraid, there have been 28,000 Islamic terrorist attacks around the world since 2001” “maybe take a look back in history the crusades were almost the same thing what Isis is doing now conquer land in the name of” “Sorry!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 2 Based on the provided information, the news about a man armed with a knife on a Swiss train starting a fire, resulting in seven people being hospitalized with knife wounds and burns, appears credible. The report is specific and details the incident, which is consistent with the style of legitimate news reporting. Analyzing the comments, there is a mix of reactions expressing concern, empathy, and discussions about the potential motives or background of the attacker. While there is a mention of Islamic terrorist attacks globally, it doesn’t directly rebut or conflict with the reported incident on the Swiss train. The comments mostly reflect individual opinions and reactions to the news rather than providing contradictory informa- tion. Therefore, the answer is: B. Real.
Prompt of Step 3 There is a piece of news: “Man armed with knife on Swiss train starts fire - seven people hospitalised with knife wounds and burns, police say.” There are comments for the news: “learn to read dude. Maybe in the future, read before commenting…” “Maybe another follower of the religion of peace that’s invaded Europe.” “hope is not wat am thinking” “I agree that most of these people do pretend to follow this religion but they could also say they follow the” “ Source please. I mean actual sources contributing unhealthy-ness to con/chem trails.” “Pathetic” “how very inconvenient” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 3 Based on the provided information, the news about a man armed with a knife on a Swiss train starting a fire, resulting in seven people being hospitalized with knife wounds and burns, appears credible. The report is specific and details the incident, which is consistent with the style of legitimate news reporting. Analyzing the comments, there is some indication of conflicting opinions or skepticism. Comments suggesting a connection to a specific religion and questioning the authenticity by asking for sources may imply some level of doubt or disagreement. However, it’s important to note that comments alone may not provide concrete evidence to dispute the credibility of the news. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 28: Twitter 数据集中的示例“764497123530375169”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “NASA astronaut captures a beautiful sunrise from aboard the International Space Station” There are comments for the news: “heavenly sigh….” “Beautiful” “Wheres the UFO pictures?” “this is cool as hell” “Nature is beauty” “Thanks to #NASA to capture such a scene !” “woah it looks like the universal studios logo, the old one from 1990’s?” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news itself seems plausible, as it describes a NASA astronaut capturing a sunrise from the International Space Station, which aligns with the activities and capabilities of NASA. The language used is neutral and descriptive, without any sensational or unrealistic claims. Additionally, the idea of astronauts capturing images from the ISS is a common and well-established practice. The comments are generally positive and express awe and appreciation for the captured sunrise. There doesn’t seem to be any overt disagreement or conflicting opinions. However, one comment humorously asks about UFO pictures, but this seems more like a playful remark rather than a serious rebuttal. No significant rebuttals or conflicts found; comments are in line with the nature of the news. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 29: Twitter 数据集中的示例“693119705469587456”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “French police arrest woman who was with man carrying weapons and ammunition at Disneyland Paris hotel.” There are comments for the news: “good. Paris doesn’t play. I do not blame them.” “honey grab some guns, ammo and the Quran and we’re off to Disneyland” “thank you.” “it does matter- a lot!” “The man was armed with the koran, which was the greater danger…” “And was packing a fellow ostensibly of some ostensibly specific genealogy?” “people never read anymore. They put he was carrying a Qur’an in the subheadline in the article. Two seconds people.” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the provided information, the news appears credible as it reports a specific event involving the arrest of a woman with a man carrying weapons and ammunition at a Disneyland Paris hotel. The writing style is straightforward and consistent with typical news reporting on law enforcement activities. Analyzing the comments, there are different perspectives expressed. Some comments express support for the arrest and appreciation for the vigilance in handling potential threats, while others inject humor or skepticism into the situation. The variety of responses is not uncommon in online discussions, especially regarding sensitive topics like security incidents. There are no clear rebuttals or conflicts within the comments that challenge the authenticity of the news. Instead, the comments reflect a range of opinions and reactions to the reported event. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表 30: Twitter 数据集中的示例“692764799315398656”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “Wow. Anthony Johnson’s knockout win is the 3rd-quickest finish in light heavyweight history (0:13).” There are comments for the news: “WWE is still fake” “that’s a bad man !!” “his name is rumble” “Johnson is a true professional. Very polite and respectful; kudos to you” “The real hands of stone.” “worked for it. Well deserved” “I can’t believe there’s been TWO others quicker than that!! TWO!!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 Based on the information provided, it is challenging to definitively determine the credibility of the news. The news itself states a specific fact about Anthony Johnson’s knockout win being the 3rd-quickest finish in light heavyweight history (0:13). However, without additional context or sources, it’s difficult to verify the accuracy of this claim. Analyzing the comments doesn’t seem to reveal any direct rebuttals or conflicts regarding the authenticity of the news. Most of the comments express admiration for Anthony Johnson’s performance, acknowledge his nickname “Rumble,” and commend him for being a true professional. Some express surprise at the speed of the knockout but without disputing the information. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表31: Twitter 数据集中的示例“767203096472719364”。 LeRuD 正确地将其预测为真实新闻。
Prompt of Step 1 There is a piece of news: “A 3-year-old daughter broke a cup and her parents drove away: the cup was broken and the child compensated for it.” There are comments for the news: “Such parents are not worthy of having children!” “Such parents are not worthy of having children!” “Can this constitute the crime of abandonment?” “My god!” “I want to say that the little girl is very sad! Very sad” “Oh my goodness… How much psychological shadow does this have to cause for children.” “Sending such a child back to their parents will lead to more hardships in the future, so be careful not to abuse them!” …… You need to do: (1) Based on the writing style and the commonsense knowledge, estimate the credibility of the news. (2) Based on the comments, analyze whether there are any rebuttals or conflicts, and then accordingly verify the authenticity of the news. Based on above results, please choose the answer from the following options: A. Fake, B. Real.
Output of Step 1 The news appears to be written in a straightforward manner, describing a situation where a 3-year-old broke a cup, and her parents drove away while she compensated for it. However, without additional details or context, it’s challenging to determine the full credibility of the news. The emotional language used in the comments may suggest a strong reaction from the audience, but it doesn’t necessarily provide evidence regarding the authenticity of the incident. The comments express strong negative sentiments towards the parents, with phrases like “not worthy of having children” and concerns about potential psychological impact on the child. However, there are no comments presenting alternative viewpoints or challenging the authenticity of the news. The comments mostly focus on expressing shock, sympathy for the child, and concern for her well-being. Therefore, the answer is: B. Real.
Final Output The final estimation of the veracity of the news is: The news is real.
表32: 微博数据集中的样本“3914951724449159”。 LeRuD 正确地将其预测为真实新闻。