SEN-CTD: semantic enhancement network with content-title discrepancy for fake news detection

• Date: 04 November 2024
• Pages: 603-620
• DOI: https://doi.org/10.1108/IJWIS-04-2024-0116
• Published date: 04 November 2024
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet
• Author: Jiaqi Fang,Kun Ma,Yanfang Qiu,Ke Ji,Zhenxiang Chen,Bo Yang

SEN-CTD: semantic enhancement

network with content-title discrepancy

for fake news detection

Jiaqi Fang,Kun Ma,Yan fa ng Qiu,Ke Ji and Zhenxiang Chen

School of Information Science and Engineering, University of Jinan,

Jinan, China, and

Bo Yang

School of Information Science and Engineering, University of Jinan, Jinan, China

and Quancheng Laboratory, Jinan, China

Abstract

Purpose –The discrepancy between the contentof an article and its title is a key characteristic of fake news.

Current methods for detectingfake news often ignore the signif‌icant difference in length between the content

and its title. In addition, relying solely on textual discrepancies between the title and content to distinguish

between real and fake news has proven ineffective. The purpose of this paper is todevelop a new approach

called semantic enhancement network with content–title discrepancy (SEN–CTD), which enhances the

accuracyof fake news detection.

Design/methodology/approach –The SEN–CTD framework is composed of two primary modules: the

SEN and the content–title comparison network(CTCN). The SEN is designed to enrich the representation of

news titles by integratingexternal information and position informationto capture the context. Meanwhile, the

CTCN focuses on assessingthe consistency between the content of news articles and their corresponding titles

examiningboth emotional tones and semantic attributes.

Findings –The SEN–CTD model performs well on the GossipCop, PolitiFact and RealNews data sets,

achieving accuraciesof 80.28%, 86.88% and 84.96%, respectively.These results highlight its effectivenessin

accuratelydetecting fake news across different types of content.

Originality/value –The SEN is specif‌ically designedto improve the representation of extremely short texts,

enhancing the depth and accuracy of analyses for brief content. The CTCN is tailored to examine the

consistency betweennews titles and their corresponding content, ensuring a thorough comparativeevaluation

of both emotionaland semantic discrepancies.

Keywords Fake news detection, Heterogeneousinformation network, Discrepancy between content and title,

Attention mechanism

Paper type Research paper

1. Introduction

Fake news, always fabricated by making some minor changes to the correct statement, is highly

deceptive and indistinguishable. The widespread of fake news in diverse domains, such as politics

(Allcott and Gentzkow, 2017) and public health (Salman Bin Naeem and Bhatti, 2020), has posed

This work was supported by the Natural Science Foundation of Shandong Province

(ZR2022LZH016), the National Natural Science Foundation of China (72471103), the Shandong

Provincial Key R&D Program of China (2021CXGC010103) and the Shandong Provincial Teaching

Research Project of Graduate Education (SDYAL2022102and SDYJG21034).

Declarations: The authors declare that they have no known competing f‌inancial interests or

personal relationships that could have appeared to inf‌luence the work reported in this paper.

International

Journal of Web

Information

Systems

603

Received18 April2024

Revised 22 June 2024

24 July 2024

29 August2024

Accepted29 August 2024

InternationalJournal of Web

InformationSystems

Vol.20 No. 6, 2024

pp. 603-620

© Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-04-2024-0116

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1744-0084.htm

a huge threat to Web security and human society. The spread of fake news is largely driven by

misleading article titles, which can be attributed to two main factors (Shrestha and Spezzano,

2021). First, many Internet users only read titles and ignore the full content. Second, some

We-mediaeditors create sensational titles to attract attention and increase click-through rates.

Some existing fake news detection methods, such as bidirectional long short-term memory

(Bi-LSTM), and convolutional neural networks (CNN) ( Jamal Abdul et al.,2021;Huosong

et al.,2023), rely heavily on sequential features. Although these methods can capture local word

sequences well, they struggle with the problem of long-distance and nonconsecutive word

interactions (Zhang et al.,2020). More recent methods based on transformer models (Ben et al.,

2021;Shaina and Chen, 2022;Zhiwei et al.,2023) excel at capturing sequential context within

local word sequences, which is crucial for text classif‌ication (Ding et al.,2020). However, they

fail in modeling the distinct interaction patterns between real and fake news sentences (Vaibhav

et al.,2019). Multimodal learning methods have also integrated features from both images and

text (Qichao et al.,2023;Chuanming et al.,2022). Yet, these approaches often fail to extract the

high-order complementary information from multimodal context (Qian et al.,2021).

The relationship between article content and titles is another key factor in fake news

detection. For example, Peter et al. (Ting et al., 2019) analyzed title–content alignment to

identify fake news. Similarly, Guo et al. (2023) incorporated text similarity, multimodal

information and authorsentiment analysis in their approach.

However, these methods face two main challenges. First, news titles are typically fewer

than 10 words long (Gu et al., 2021), brief‌ly to capture less comprehensive semantic

information than longer texts. Second, the signif‌icant length disparity between news titles

and content (Milojević, 2017) leads to textual misalignments, making it diff‌icult to rely

solely on textual differencesto distinguish real news from fake news.

To address these challenges, we propose a semantic enhancement network with content–title

discrepancy (SEN–CTD). Our model uses two differenttypes of information. One of them is the

textual information extracted from the news it self. Another type is the discrepancy information

captured from news titles and content. Specif‌ically, the contribution of SEN–CTD can be divided

into two parts: (1) SEN, which is used to extract rich semantic information from news headlines

and content and (2) content–title comparison network (CTCN), which is used to capture the

discrepancy of news titles and content. Our contributions are summarized as follows:

Semantic enhancement network:The SEN involves two designs: supplementing text with

external information and fully using position information. By leveraging heterogeneous

information networks(HIN) to integrate multisource data, the def‌iciencies in text content can

be addressed. In addition, the introduction of word position information through

bidirectional encoders enhances the contextual representation of the text. By combining

these two designs, we f‌ind that the incorporation of position information mitigates the

challenge of syntactic loss in graph-basedtext modeling.

Content–title comparison network: The CTCN is proposed to evaluate the alignment

between news content and titles. By examiningtheir emotional and semantic features, CTCN

assesses the consistency from both discrepancy and similarity perspectives. It is found that

the incorporation of emotional comparison enhances the ability to discern between real and

fake news articles. In addition, the simultaneous consideration of both discrepancy and

similarity facilitatesa more comprehensive understanding of their consistency.

2. Related work

2.1 Fake news detection

Fake news detection methods can be divided into content-based and social-context-based

methods (Shu et al.,2017). Social-context-based methods analyze how news spreads on

IJWIS

20,6

604