Robust Misinformation Detection by Visiting Potential Commonsense Conflict

1 Introduction

Over the past decades, many social media platforms e.g., Twitter and Weibo, become the mainstream avenue to share information among human beings in daily life. Unfortunately, these platforms eventually afford convenience for the dissemination of various misinformation such as fake news and rumors [Vosoughi et al., 2018; van der Linden, 2022]. To reduce the negative effect of misinformation, how to detect them effectively and efficiently becomes the primary task in this endeavor. Accordingly, the emergent topic of problem is how to measure and express them for given articles. Accordingly, we propose a novel plug-and-play augmentation method for the MD task, namely Misinformation Detection with Potential Commonsense Conflict (MD-PCC). Specifically, we propose to measure the commonsense conflicts of articles by the difference between the extracted commonsense triplet and the golden triplet inferred by the wellestablished commonsense reasoning tool [Bosselut et al., 2019; Hwang et al., 2021], and use those triplets to specify a predefined commonsense template as commonsense expressions to express the potential commonsense conflicts. For each article, we integrate it with its corresponding specific commonsense expression to form an augmented one, named commonsense-augmented article. Given those augmented articles, one can build effective detectors by any existing MD methods and backbones.

∗Corresponding authors

For empirical evaluations, we employ 4 public benchmark datasets GossipCop [Shu et al., 2020], Weibo [Sheng et al., 2022], PolitiFact [Shu et al., 2020] and Snopes [Popat et al., 2017]. Additionally, we further collect a new Commonsenseoriented Misinformation benchmark datasets, named CoMis, whose all fake articles are caused by commonsense conflict. We integrate MD-PCC with various existing MD backbones and compare them across public benchmark datasets and CoMis. Empirical results demonstrate that MD-PCC can consistently outperform the existing MD baselines. The source code and data of MD-PCC are released in the repository https://github.com/wangbing1416/MD-PCC.

The primary contributions of this paper can be summarized as the following three-folds:

• We propose a plug-and-play augmentation MD method, named MD-PCC, by expressing the potential commonsense conflict.
• We collect a new commonsense-oriented misinformation dataset, named CoMis, whose all fake articles are caused by commonsense conflict.
• We conduct experiments across both public benchmark datasets and CoMis, and empirical results indicate the effectiveness of MD-PCC.

2 ProposedMD-PCC Method

In this section, we briefly review the task definition of MD and prevalent commonsense reasoning methods. We then describe the proposed method MD-PCC in more detail.

2.1 Preliminaries

Task formulation of MD. Commonly, the basic goal of MD is to induce a detector Fθ(·) over a given training dataset D, and use Fθ(·) to distinguish whether any unseen article is real or fake. We formally describe the dataset of N training samples as D = {xi, yi}Ni=1 , where each sample is composed of a raw article xi and its corresponding veracity label yi ∈ {0, 1}, i.e., 0/1 indicating fake/real. With any specific detector Fθ(·), it can be trained by optimizing the following objective with respect to θ:

where ℓ(·, ·) denotes the binary cross-entropy loss commonly.

Commonsense reasoning. Generally speaking, current commonsense reasoning methods aim to train a generative language model referring to the relation triplet (s, r, o), where s and o are the subject and object, respectively, and r is the relation between them. Given any subjectrelation pair (s, r), a commonsense reasoning method can accurately predict the corresponding object o. Typically, the methods are trained across the commonsense-oriented dataset ATOMIC20 20 [Hwang et al., 2021], which comprises a substantial collection of relation triplets. The typical commonsense-oriented relations of ATOMIC20 20 include {xNeed, xAttr, xReact, xEffect, xWant, xIntent, oEffect, oReact, oWant, isAfter, HasSubEvent, HinderedBy}, representing the relations between specific events or human actions. Beyond these ones, these methods can also be generalized to a large knowledge base ConceptNet [Speer et al., 2017], so as to capture the relations between entities including {MadeOf, AtLocation, isA, Partof, HasA, UsedFor}. To make notation simple, we use R to denote the set of all those relations captured by the commonsense reasoning methods.

2.2 Overview of MD-PCC

Basically, our MD-PCC is a plug-and-play augmentation method for the MD task. We take inspiration from the assumption that fake articles are more likely to involve commonsense conflict. Accordingly, we design a commonsense template to express the potential commonsense conflict measured by prevalent commensense reasoning methods and specify it for each original article as the augmentation. To be specific, the commonsense template is designed as

where (s, r, o) indicates the representative commonsense triplet extracted from the article, ˆo is the golded object corresponding to (s, r) generated by commonsense reasoning methods, and Γ(r) denotes the original expression of r, e.g., "is made of " is the original expression of MadeOf. We suppose that an article involves a commonsense conflict if o ̸= ˆo, otherwise o = ˆo. Accordingly, we define that c will be specified by the adversative conjunction word "However" when o ̸= ˆo; and by contrast, it will be specified by "And", and the text segment "instead of" ⊕ o will be excluded.

With this commonsense template, for each article xi, we form its corresponding commonsense expression ei by specifying (si, ri, oi), ˆoi and ci with three stages: commonsense triplet extraction, golden object generation, and commonsense expression construction, respectively. Accordingly, we concatenate xi and ei as a commonsenseaugmented article ˆxi. Given all commonsense-augmented samples {ˆxi, yi}Ni=1 , we can formulate the following objective with respect to any specific detector Fθ(·):

For clarity, the overall framework of MD-PCC is depicted in Fig. 1. In the following subsections, we will describe the three stages of generating commonsense expressions.

2.3 Commonsense Triplet Extraction

In this stage, for each article xi, we extract a certain number of relevant commonsense triplets {(sγi , rγ i , oγi )}|Ri| γ=1 . To achieve this, we first screen all relations of R to extract all corresponding triplets {(sγi , rγ i , oγi )}|R| γ=1 fromxi and then filter out the meaningless ones from them.

Specifically, we first extract {(sγi , rγ i , oγi )}|R| γ=1 by prompting an existing LLM with the In-Context Learning (ICL) method [Brown et al., 2020; Min et al., 2022]. We design natural language queries {T γ }|R| γ=1 for relations {rγ i }|R| γ=1 , e.g., "Extract entity1 and entity2 from the text where entity1 is made of entity2. Text:" for the relation MadeOf. Accordingly, the formulation of in-context examples {I γ k }Kk=1 for the relation rγ is delineated as:

We collect K labeled examples for each relation to facilitate ICL. Then, we input both in-context examples and query T γ ⊕xi into a triplet extractor GΦ(·) specified by a pre-trained T5 model [Raffel et al., 2020] to generate sγi and oγi :

Because an article does not always contain all relations of R, we filter out the meaningless ones. We design a filtering method based on the conditional generation logits. It follows the spirit that generative models always output lower probabilities for its generated uncertain word tokens, which can be evaluated by perplexity [Jurafsky, 2000;

Lee et al., 2021]. Specifically, we define the text generated by GΦ(·) in Eq. (4) as tγi = {tγi1 , tγi2 , · · · , tγiL}, where L indicates the length of the text. And we remove the triplet (sγi , rγ i , oγi ), if∑L

where ϵ is a controllable hyper-parameter. After filtering, the set of commonsense relations for xi is refined as Ri ∈ R.

2.4 Golden Object Generation

Given {(sγi , rγ i )}|Ri| γ=1 , we generate their golden objects {ˆoγi }|Ri| γ=1 , which are aligned with real-world commonsense knowledge. To be specific, we feed each (sγi , rγ i ) into the prevalent commonsense tool GΠ(·) [Bosselut et al., 2019] to generate its golden object ˆoγi as

We specially explain that because the prevalent commonsense reasoning tool has been pre-trained on a large-scale commonsense dataset ATOMIC20 20 , we treat ˆoγi as the ground-truth knowledge of (sγi , rγ i ), i.e., the corresponding golden object.

2.5 Commonsense Expression Construction

In this stage, we construct commonsense expression ei by filling the commonsense template in Sec. 2.2 based on {(sγi , rγ i , oγi )}|Ri| γ=1 and {ˆoγi }|Ri| γ=1 . Specifically, we first compute conflict scores {cγi }|Ri| γ=1 for each pair of oγi and ˆoγi . We

Algorithm 1 Training summary of MD-PCC.

Input: Training dataset D = {xi, yi}Ni=1; pre-trained language model GΦ(·); commonsense reasoning tool GΠ(·); commonsense relations R; query templates {T γ }|R| γ=1 . Output: detection model Fθ(·); expressions {ei}Ni=1.

1: for i = 1, 2, · · · ,N do 2: Ci ← [ ], 3: for rγ i in R do 4: extract sγi and oγi with GΦ(·) in Eq. (4), 5: if Eq. (5) is not satisfied then 6: ̂oγi ← GΠ (sγi , rγ i ), 7: calculate cγi with Eq. (6), Ci ← cγi . 8: end if 9: end for 10: select {si, ri, oi} and ̂oi with max(Ci), 11: construct ei with Eq. (7). 12: end for 13: train Fθ(·) with L in Eq. (2).

take inspiration from BARTSCORE [Yuan et al., 2021], and present a new evaluation metric to compute the commonsense conflict score cγi during the process that we input sγi and rγ i into the commonsense reasoning tool GΠ(·) with Eq. (5). The specific metric is as follows:

where L denotes the length of the generated ˆoγij .

Then, we select the highest conflict score ci from the set of {cγi }|Ri| γ=1 , and denote its corresponding representative commonsense triplet and golden object as {si, ri, oi} and ˆoi, respectively. Accordingly, we fill them into the commonsense template to obtain the expression as follows:

where Γ(·) is the original expression for each relation, e.g., "is made of " for the relation MadeOf. When ci ≥ µ, we argue that the article xi exists the commonsense conflict; otherwise, there is not. In summary, the training summary of MD-PCC is presented in Alg. 1.

3 Datasets

To evaluate the performance of MD-PCC, we conduct experiments by employing four public MD datasets GossipCop [Shu et al., 2020], Weibo [Sheng et al., 2022], PolitiFact [Shu et al., 2020] and Snopes [Popat et al., 2017]. Additionally, we also collect a new Chinese MD dataset, referred to as CoMis, wherein all fake articles can be verified by leveraging commonsense conflict. We describe their details in the following section. For clarity, their statistics are shown in Table 2.

3.1 Prevalent MDDatasets

We evaluate the method with the following four MD datasets:

• GossipCop and PolitiFact are English MD datasets sourced from FakeNewsNet [Shu et al., 2020]. We divide GossipCop based on [Zhu et al., 2022], which includes articles posted between 2000 and 2017 for training, with the test set consisting of articles from 2018. For PolitiFact, we adhere to its original dataset division.
• Weibo [Sheng et al., 2022] is sourced from a Chinese social media platform, and we split articles published from 2010 to 2017 allocated for training and those from 2018 used for testing.
• Snopes [Popat et al., 2017] is gathered from a wellknown fact-checking website snopes.com. We split the dataset according to its original paper.

3.2 Our Collected CoMis

We collect a new commonsense-oriented MD dataset CoMis with the effort of human annotators. Table 2 provides the statistics of our newly constructed dataset CoMis. The dataset contains a total of 1,580 pieces of data entries, covering diverse domains. The domain most extensively represented in the dataset pertains to food safety.

Data source. Our MD data is sourced from two distinct channels: pre-existing MD datasets and external websites.

First, we select suitable data items from pre-existing datasets dedicated to fake news and rumor detection, e.g., Weibo-16, Weibo-20, and Weibo-COVID19. Specifically, Weibo-16 [Ma et al., 2016] comprises posts spanning from December 2010 to April 2014, and many duplications are meticulously filtered by [Zhang et al., 2021]; Weibo-20 [Zhang et al., 2021] extends the temporal scope of Weibo-16, encompassing data from April 2014 to November 2018, and its labels are verified through NewsVerify 1 ; Weibo-COVID19 [Lin et al., 2022] is collected during the surge of the COVID19 pandemic, so all articles within this dataset are exclusively centered on COVID-19 topics.

To ensure completeness and timeliness, we also manually collect commonsense-oriented samples from two external websites. First, Food Rumor 2 is a Chinese rumor-refuting platform, which serves as a repository for misinformation and its corresponding verification, with a predominant focus on topics related to food safety, health science, and similar domains. Then, Science Facts 3 is another Chinese platform that specializes in disseminating science popularization content, covering subjects, e.g., food safety and biological science.

1 https://www.newsverify.com/

2 http://www.xinhuanet.com/food/sppy/

3 https://piyao.kepuchina.cn/

Annotation and post-process. The annotators are instructed to select and post-process the data items that can be verified using commonsense from the aforementioned data sources. For the data from pre-existing MD datasets, we preserve their veracity labels while systematically filtering any special symbols and website links from their content. For the data sourced from external websites, we collect fake claims in the rumor-refuting channels of these websites, and real claims from their science popularization channels. Meanwhile, we maintain a consistent average claim length of approximately 50, aligning with the standards set by existing datasets.

4 Experimental Results

In this section, we aim to empirically evaluate our proposed method MD-PCC, and answer the following questions:

• Q1: Can the proposedMD-PCC consistently improve the performance of misinformation detectors?
• Q2: Is MD-PCC sensitive to its hyper-parameters and primary components?
• Q3: Can the generated expression e expresses the commonsense conflict of the article?

4.1 Experimental Settings

Baselines. We evaluate our plug-and-play method MDPCC across five prevalent MD approaches, including EANN [Wang et al., 2018], BERT [Devlin et al., 2019], BERTEMO [Zhang et al., 2021], the SOTA MD model CED [Wu et al., 2023], and DM-INTER [Wang et al., 2024a].

Implementation Details. In our experiments, we employ pre-trained language models FlanT5Large 4 [Chung et al., 2024] and mT5Large 5 [Xue et al., 2021] to extract commonsense triplets for the English and Chinese MD datasets, respectively. To generate golden objects, we use COMETATOMIC20 20 6 [Hwang et al., 2021] for English datasets and comet-atomic-zh 7 for Chinese datasets Weibo and CoMis.

4 https://huggingface.co/google/flan-t5-large.

5 https://huggingface.co/google/mt5-large.

During the training stage, we use an Adam optimizer with a learning rate of 7 × 10-5 for the BERT model in baseline methods. For the other modules such as the linear classifier, we use a learning rate of 1×10-4 , and the batch size is consistently fixed to 64. We also fix some other manual parameters empirically, such as K, ϵ, and µ to 5, 0.8, and 0.6, respectively. To avoid overfitting of detectors, we adopt an early stop strategy. This means that the training stage will stop when no better Macro F1 value appears for 10 epochs.

4.2 Main Results (Q1)

To answer Q1, Tables 3 and 4 report the performance outcomes of our method MD-PCC on two benchmark datasets and our constructed dataset, respectively. To mitigate the influence of randomness, we repeat each experiment five times using five different seeds {1, 2, 3, 4, 5}. The standard deviations of the five replicates are also illustrated in Tables 3 and 4. Overall, our MD-PCC method, which functions as a plug-in approach, can significantly and consistently improve the performance of the baseline models across all evaluation metrics. For example, on the Weibo dataset, our MD-PCC improves the overall F1 and fake news F1 scores by 1.91 and 2.54, respectively, compared to the current state-of-the-art MD method CED. Additionally, on GossipCop, it achieves improvements of 1.46 and 3.19 in macro F1 and precision scores. When we compare different MD datasets, we observe that MD-PCC performs better on CoMis than on the other Chinese dataset Weibo across most evaluation metrics. Specifically, when compared to the BERT baseline, MD-PCC improves its macro F1 and precision scores by 1.16 and 0.74 on Weibo, while it shows more significant improvements of 2.85 and 2.52 on CoMis. These results highlight the effectiveness of MD-PCC in incorporating commonsense knowledge to enhance the detection of knowledge-rich misinformation.

6 https://github.com/allenai/comet-atomic-2020.

7 https://huggingface.co/svjack/comet-atomic-zh.

4.3 Ablative Study (Q2)

To investigate Q2, we implement ablative experiments to assess the effectiveness of key components in MD-PCC. Specifically, we conduct experiments on Weibo and GossipCop, and present three ablative versions of CED + MD-PCC as follows:

• MD-PCC w/o ICL: the version without ICL (K = 0) in the commonsense triplet extraction stage;
• MD-PCC w/o c: the version without conjunction words c, e.g., "However", in commonsense expressions;
• MD-PCC w/o o: the version without "instead of " ⊕ o in commonsense expressions.

The ablative results are presented in Table 5. Generally, each ablative version exhibits a decreasing trend compared to MD-PCC, illustrating the contribution of each component in our model. The overall performance ranking of these ablative versions is w/o o >w/o c >w/o ICL. This ordering indicates that: (1) the direct impact of commonsense triplet extraction on the model's performance is significant, and in-context learning consistently enhances the extraction; (2) conjunction words c is more important than o in commonsense expressions. This is because misinformation detectors can effectively learn the pattern between conjunctions and veracity labels, e.g., the pattern between "However" and Fake.

4.4 Case Study (Q3)

The goal of MD-PCC is to construct commonsense expressions that express the potential commonsense conflict. Therefore, we provide some representative cases in Table ?? to evaluate the generated expressions. Specifically, we select two cases from CoMis and translate them into English versions. We observe that (1) MD-PCC extracts commonsense triples accurately, even from relatively complex articles, e.g., the second case; (2) MD-PCC can assign a higher conflict Article: Meat floss is made of cotton. This was discovered by my niece's mother-in-law. Moms, please pay attention. Expression: However, meat floss is made of meatloaf instead of cotton.

score to the triplet that does exist the commonsense conflict; (3) our presented filtering method in Sec. 2.3 can indeed filter out commonsense relations that do not exist in the article, e.g., AtLocation in the first case.

5 RelatedWorks

In this section, we briefly review the related literature about misinformation detection and commonsense reasoning.

5.1 Misinformation Detection

Misinformation, e.g., fake news and rumors, has had a detrimental impact on society [Vosoughi et al., 2018; Zhang et al., 2023]. As a result, it has become increasingly important to identify and detect misinformation, which is referred to as misinformation detection. Specifically, most cuttingedge MD techniques focus on detecting misinformation based on its textual and multimodal content [Ying et al., 2023; Wang et al., 2024b], using advanced deep learning models [Ma et al., 2016; Shu et al., 2020; Wang et al., 2024c; Xiao et al., 2024]. These models often incorporate external features like knowledge bases [Dun et al., 2021], emotional signals [Zhang et al., 2021; Jiang et al., 2024], and user feedback [Ma et al., 2016; Lin et al., 2023]. Meanwhile, some recent works have also explored strategies to leverage pre-trained large models for MD [Hu et al., 2024; Chen and Shu, 2024; Nan et al., 2024; Wan et al., 2024].

In this study, we integrate commonsense knowledge into MD models. Prior to our work, certain research efforts have aimed to leverage knowledge graphs for the enhancement of MD models. These endeavors have primarily involved learning knowledge embeddings [Dun et al., 2021; Sun et al., 2022] or retrieving entity descriptions [Hu et al., 2021; Jiang et al., 2022]. In contrast to these approaches, our incorporation of commonsense knowledge aligns more closely with human reasoning and reactions. Meanwhile, we employ generative models for data augmentation explicitly, which obviates the need for extensive retrieval from large knowledge bases and reduces computational complexity.

5.2 Commonsense Bases and Reasoning

Commonsense knowledge bases, such as ConceptNet [Speer et al., 2017], ATOMIC20 20 [Hwang et al., 2021], offer a valuable resource for direct reasoning with commonsense knowledge and have found applications in various academic topics, e.g., machine translation [Liu et al., 2023], question answering [Wang et al., 2023; Chen et al., 2023] and sarcasm detection [Min et al., 2023]. Recently, especially within the context of Large Language Models (LLMs), the utilization of commonsense reasoning with LLMs has garnered significant attention [Liu et al., 2023; Shen, 2024; Wang et al., 2024d]. These works frequently treat commonsense knowledge as supplementary information or assess the presence of commonsense knowledge within LLMs.

6 Conclusion

In this paper, we aim to enhance MD models by uncovering commonsense conflicts. To achieve this goal, we propose a novel MD method named MD-PCC, designed to generate commonsense expressions for each article, explicitly expressing commonsense conflict existing inherent in articles, and leverage it to augment original articles. Specifically, the expression is constructed through a commonsense triplet extracted from the original article, the corresponding golden object, and a conjunction word. To obtain these components, we first prompt the pre-trained language model with in-context examples to extract triplets and filter out irrelevant triplets. Then, the commonsense tool is employed to generate their corresponding golden objects. Finally, a new metric is designed to measure the commonsense conflict, and the conjunction word is determined using this metric. Additionally, we also collect a new commonsense-oriented MD dataset, and extensive experimental results on the datasets are conducted and prove the effectiveness of our proposed MD-PCC.

Acknowledgements

We acknowledge support for this project from the National Key R&D Program of China (No.2021ZD0112501, No.2021ZD0112502), the National Natural Science Foundation of China (No.62276113), China Postdoctoral Science Foundation (No.2022M721321).