Page 340 - 《软件学报》2025年第12期

余建兴 等: 基于常识推理问答的多模态题文不符检测 5721

4
(Guangdong Artificial Intelligence and Digital Economy Laboratory (Guangzhou), Guangzhou 510330, China)
Abstract: This study investigates the task of clickbait detection in social media posts. These posts often employ deceptive headlines or
thumbnails to mislead readers into clicking on irrelevant or undesirable content, thus enabling widespread dissemination and generating
commercial benefits such as increased clicks. To evade detection, malicious creators frequently disguise clickbait posts as legitimate ones,
using techniques such as adding irrelevant or misleading content to deceive the detector. Detecting such posts requires a detailed analysis
and complex multi-step reasoning using commonsense knowledge to identify inconsistencies. However, existing methods typically treat a
post as a simple text span and feed it into a neural network for classification, neglecting the analysis of inherent false details, which leads
to misjudgments. Moreover, these black-box models lack explainability. To address this issue, a new question-guided detector is proposed,
which systematically analyzes the details through a doubt-then-verify approach to uncover potential inconsistencies and falsehoods.
Specifically, a multi-modal retrieval-augmented technique is used to extract detailed clues from the content of the post, followed by
questioning each clue. To ensure thorough verification of facts and their complex relationships, both simple matching questions and deep
commonsense reasoning questions with varying levels of complexity are employed. Each question yields a plausible answer from the post,
but the answer may be fabricated or inaccurate. Therefore, an open-domain QA model is utilized for cross-verification, leveraging external
knowledge to derive a more reliable answer. When discrepancies are found between answers, the post is likely to contain false content.
This inconsistency serves as a valuable feature and can be combined with other multi-modal features indicative of clickbait, improving the
discriminative power of the detection model. By breaking down the complex clickbait detection task into a series of question-guided
verification steps, inconspicuous inconsistencies can be identified to explain the underlying reasons for clickbait. Extensive experiments on
three popular datasets demonstrate the effectiveness of the proposed approach.
Key words: clickbait detection; commonsense reasoning; question generation

随着社交媒体的快速发展, 推文已成为人们获取信息和分享感悟的重要渠道. 优质的推文可以捕获大量的用
户点击量和阅读量, 这会给推文的创作者和社交平台带来盈利. 为了获得竞争优势, 恶意创作者通常使用不正当的
方式在推文的标题、封面图和正文中添加一些诱人或欺骗性的内容, 以吸引读者的关注. 当读者受好奇心驱使 [1]
而点开相关的链接, 却发现内含大量如夸张、低俗等低质内容, 进而产生被骗、失望和厌恶的感觉. 受利益的驱
使, 这种题文不符的推文在社交平台上日渐猖獗. 这不但损害了用户体验, 而且还会给社交平台甚至社会都带来严
重的负面影响, 譬如传播错误信息、误导公众舆论、侵蚀用户信任和社交平台的声誉等 [2] . 因此, 研究如何有效地
检测这种题文不符的推文具有重要的商业价值和广泛的应用前景 [3] .
为了检测题文不符推文, 社交平台采取了很多的方法, 其中一种方式是人工审核. 然而, 这种方式成本高、耗
时大, 难以应对大量题文不符推文的快速传播. 因此, 机器自动检测成为当前的重点研究方向. 根据判别依据的来
源, 机器检测方法可以归纳为两种类型. 第 1 种是基于社交行为 [4] , 即通过学习推文在网络上传播的特性来判别,
如题文不符推文通常有高点击量和高分享次数, 但由于质量差, 对应的阅读时间却很短. 但这种行为特征依赖于用
户的反馈数据, 存在延迟和遗漏等问题, 且检测时效性弱. 另一种方法是分析推文的内容, 根据题文不符推文常见
的欺骗性词语、句法、情感、写作风格 [5] , 甚至是标点符号 [6] 等语言特征来判断. 此外, 题文不符推文的各个组成
部分之间通常存在不一致, 例如标题与正文无关, 所描述的内容虚假并与实际情况相矛盾等. 为了分析内容的质
量, 早期方法依赖于人工构建的规则, 成本高且可扩展性差. 后期使用数据驱动的神经网络模型, 通过学习内容与
标签之间的关联来预测. 虽然可用性变强了, 但这种黑箱般的神经模型难以解释决策过程 [7] . 可解释性对于建立可
信人工智能至关重要 [8] . 此外, 仅基于简单关联, 恶意创作者很容易利用一些技巧来伪装, 例如添加一些无害的无
关内容以制造假关联来欺骗检测模型. 此外, 恶意创作者还会引入多种模态的虚假内容, 如使用视觉模态的与正文
无关封面图来诱惑读者点击. 这比纯文本内容更难判断, 因为需要分析各种模态内容细节的真实性和相关性.
所谓质疑是发现真理的重要途径. 针对这些复杂的题文不符推文, 我们发现人们往往通过质疑来分析出细节
中的不一致, 即对推文中的各个可疑点进行发问和深度验证, 并结合推文字面的含义和常识推理来发现矛盾点, 实
现去伪存真. 如图 1 所示, 推文带有一条吸引人的标题, 恶意创作者在正文中注入了不少虚假内容, 让其跟真实内
容混杂在一起, 力求让其伪装成合法的, 从而引起检测模型漏判和误判. 图 1 中的红框标记了诱骗的内容, 下划线
和箭头表示多步的推理过程. 针对题文不符推文, 人们往往通过质疑来揣摩细节和发现矛盾. 例如发问“由在基尔
大学实习的 15 岁男孩取得的名为 WASP-142b 的新发现是什么?” 人们发现从推文字面中得到答案 1, 但依赖外部