Online News Veriﬁcation: An AI-Based Platform for Assessing and

Visualizing the Reliability of Online News Articles

Anne Grohnert, Simon Burkard, Michael John, Christian Giertz, Stefan Klose, Andreas Billig,

Amelie Schmidt-Colberg and Maryam Abdullahi

Fraunhofer FOKUS, Kaiserin-Augusta-Allee 31, 10589 Berlin, Germany

ﬁ

Keywords:

Online News Veriﬁcation, Web Information Analysis, Fake News Detection, Disinformation, Reliability.

Abstract:

Assessing the reliability of online news articles poses a signiﬁcant challenge for users. This paper presents a

novel digital platform that enables users to analyze German-language news articles based on various reliability-

related aspects, including opinion strength, sentiment, and article dissemination. Unlike many existing ap-

proaches focused solely on detecting fake news, this platform emphasizes the comparative analysis and vi-

sualization of relevant reliability indicators across articles from different publishers. The paper provides a

comprehensive overview of the current state-of-art describing various existing approaches for the detection

and presentation of disinformational online content before presenting the technical system architecture and

user interfaces of the designed platform. A concluding user evaluation reveals some limitations and oppor-

tunities for further developments, but showed generally positive feedback on the platform’s diverse analysis

criteria and visual presentation to support users in assessing the credibility of news articles. Potential future

applications range from evaluating article neutrality to verifying citations in academic contexts.

1 INTRODUCTION

Assessing the credibility and reliability of online con-

tent is often difﬁcult for users. Valuable support in

this task could be provided by an intelligent system

that assesses the trustworthiness of online content and

offers assistance in interpreting the characteristics of

online articles with regard to their credibility. This

applies not only to the assessment of user-generated

posts on social media platforms but also to the analy-

sis of the credibility of journalistic articles from rep-

utable as well as less reputable news publishers. A

tool for analyzing the reliability of online news arti-

cles could, for example, help in assessing the objectiv-

ity of online articles and the citability of journalistic

sources, or assist in analyzing the dissemination paths

of certain online content.

Existing research focuses mostly on the develop-

ment of efﬁcient methods and algorithms for the au-

tomated detection and classiﬁcation of disinformative

content (fake news), especially in the context of so-

cial media, but less on the development of systems

that analyze and compare news articles with respect

to various reliability features visualizing also the re-

sults of the aspectoriented analysis.

This paper presents a novel digital platform

on which users can examine individual German-

language news articles with respect to various aspects

of disinformation to gain a structured, uniform, and

personalized overview of news articles from differ-

ent news publishers. The paper begins with a de-

tailed overview of existing methods and the current

state of research in the ﬁeld of disinformation detec-

tion. Afterwards, the technical system architecture of

the developed platform is presented and the compo-

nents and interfaces involved are described, e.g., for

AI-based assessment of article trustworthiness, deter-

mination of opinion strength, and visualization of ar-

ticle dissemination paths. A concluding summary of

a user evaluation highlights potentials and opportuni-

ties for further development.

2 STATE OF THE ART

Nowadays exist several approaches to detect fake

news. These are divided into manual checks, auto-

mated procedures and systems that combine several

approaches.

412

Grohnert, A., Burkard, S., John, M., Giertz, C., Klose, S., Billig, A., Schmidt-Colberg, A. and Abdullahi, M.

Online News Veriﬁcation: An AI-Based Platform for Assessing and Visualizing the Reliability of Online News Articles.

DOI: 10.5220/0013778100003985

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 412-419

ISBN: 978-989-758-772-6; ISSN: 2184-3252

2.1 Manual Disinformation Detection

Various fact checking organizations rely on manual

human-based analysis of news content, such as Cor-

rectiv (CORRECTIV, 2025) or Faktenfuchs (BR24,

2025) in German-language media, PolitiFact (Poyn-

ter, 2025) as US fact-checking platform or Full Fact

(FullFact, 2025) in the UK. They aim to identify,

verify and correct misinformation and disinformation

in the public sphere helping to shape public opin-

ion by providing fact-based information. They verify

claimed facts or political statements circulating in the

media, social networks and publish explanatory arti-

cles why the provided information is false, misleading

or correct. Many of these organizations work accord-

ing to a transparent code of conduct (e.g. the Inter-

national Fact-Checking Network - IFCN of Pointer

(IFCN, 2025)). They disclose how they work, who

funds them and on what basis judgments are made.

2.2 Approaches for Automatic

Detection of Disinformation

Besides manual fact checking various (scientiﬁc) ap-

proaches exist at present for the automatic detection

of disinformation. Table 1 shows the various ap-

proaches, the targeted detection focus and a few ex-

amples of state-of-the-art existing systems.

Fact Checking. In the ﬁeld of automated fact-

checking, statements are veriﬁed by comparing them

with facts or knowledge bases. Typical methods in-

clude claim detection (identifying veriﬁable claims),

claim matching (comparing with known facts, e.g.,

from PolitiFact, Snopes, Wikipedia, Wikidata) and

natural language inference (NLI), assessing whether

a text supports, contradicts or is neutral towards a

claim. For example, dEFEND is an explainable fake

news detection system that analyzes both the news

content for particularly check-worthy sentences and

processes user comments to ﬁlter out opinions and in-

dications of false information (e.g. skepticism, ref-

erences to fact checks) (Shu et al., 2019). By analyz-

ing news content and comments together, relevant key

sentences and comments are identiﬁed that serve as

an explanation. As a result, the model not only pro-

vides a classiﬁcation (fake vs. real), but also shows

the speciﬁc sentences and comments that inﬂuenced

its decision.

Source Credibility. Source analysis (source credi-

bility) involves assessing the reliability of the origi-

nal source (website, author, domain). Among other

characteristics, the domain reputation (e.g. .edu vs.

.xyz), but also the history of the source (e.g. ex-

istence of false statements in the past) are consid-

ered. Browser plugins such as NewsGuard (News-

Guard Technologies, 2025) or TrustyTweet (Hartwig

and Reuter, 2019) can be used directly for source

credibility analysis without having to leave the web-

site. NewsGuard evaluate news sites according to

nine publicly accessible criteria that are based on jour-

nalistic standards and source transparency (e.g. clear

separation of opinion and news, regular publications

and transparency about provider, publisher and own-

ership structure). TrustyTweet analyzes and evaluates

Twitter/X posts for credibility through source veriﬁ-

cation and the use of AI. Warnings or trust ratings are

displayed directly below the posts.

Style Analysis. The aim of style analysis (writing

style, linguistic features) is to identify disinformation

through characteristic linguistic patterns. Character-

istics such as the use of excessive adjectives, superla-

tives and exclamation marks as well as low objectiv-

ity and high subjectivity or the use of sensationalistic

language are taken into account. The FakeFinder app

recognizes fake news from the Twitter livestream and

warns users in real time. It is based on the small open-

source NLP language model ALBERT from Google

AI (Tian et al., 2020). Grover is an AI model for

generating and recognizing fake news (Zellers et al.,

2019). It was trained using a transformer model on

a data set of real journalistic articles. Grover anal-

ysis the style, structure, metadata and linguistic fea-

tures and classiﬁes whether the text probably origi-

nates from a human journalist, from Grover itself or

from another AI model.

Sentiment Analysis. Sentiment analysis and emotion

recognition are about recognizing emotionally con-

tent that is typical for propaganda or disinformation.

The dominant emotions here are fear, anger or out-

rage and the language used is often polarizing. Deep-

Moji is a neural language model that was developed

to recognize emotions and moods in texts based on

the emojis used (Felbo et al., 2017). Around 1 bil-

lion tweets on Twitter with emojis as labels served as

training data and 64 of the most frequently used emo-

jis served as target classes for the modeling.

Social Signals. In the area of social context analy-

sis (social signals), the behavior of dissemination and

interactions in social networks is examined. Mod-

els such as Propagation Tree Analysis, Graph Neu-

ral Networks (GNNs) or Temporal Pattern Modeling

(e.g. LSTM, Transformers) are used. Botometer (for-

merly: BotOrNot) is an online tool designed to detect

social media bots on Twitter/X (Yang et al., 2022). It

evaluates how probable it is that an account is an au-

tomated Bot.

Hoaxy is a web platform and visualization tool

that shows how fake news and fact checks spread in

Online News Veriﬁcation: An AI-Based Platform for Assessing and Visualizing the Reliability of Online News Articles

413

Table 1: Different approaches for disinformation detection.

Approach Detection Focus Example System(s)

(1) Fact Checking Veracity of Claims dEFEND

(2) Source Credibility Trustworthiness of the Source NewsGuard, TrustyTweet

(3) Style Analysis Linguistic Patterns Grover, FakeFinder

(4) Sentiment Analysis Emotional Manipulation DeepMoji

(5) Social Signals Dissemination Patterns Botometer, Hoaxy

(6) Multimodal Inconsistency between Image and Text VisualBERT, VilBERT

(7) Knowledge Graphs Logical Consistency and World Knowledge DeFacto, Kauwa-Kaate

social networks, especially on Twitter/X (Hui et al.,

2018). It was developed by Indiana University as part

of the OSoMe (Observatory on Social Media) project

by the same researchers behind Botometer.

Multimodal. Multimodal approaches combine the

analysis of visual and linguistic content, motivated by

the fact that many fake news stories do not only con-

sist of textual content, but are also image or video-

based. Methods such as image forensics (e.g. re-

verse image search, manipulation detection) and mul-

timodal models such as VisualBERT (Li et al., 2019)

or VilBERT (Lu et al., 2019) are used to compare

whether there is a contradiction between image con-

tent and text content. Both are multimodal trans-

former models designed to process text and image

data together and being able to perform tasks such as

a text-image coherence check or a multimodal senti-

ment analysis. The main difference between the two

models lies in the architecture, particularly how im-

age and text are combined.

Knowledge Graphs. Approaches for checking log-

ical contradictions by comparison with structured

knowledge databases (e.g. DBpedia, Wikidata) build

knowledge graphs and check for logical consistency.

Techniques include Triple Extraction (subject, pred-

icate, object), Graph Reasoning or Knowledge Base

Completion Consistency Check. DeFacto (Deep Fact

Validation) was developed with the aim of automat-

ically checking whether a claimed fact is true or

false by comparing it with known facts from semantic

knowledge databases (Linked Open Data, e.g. DBpe-

dia) (Gerber et al., 2015). It provides additional evi-

dence in text form, e.g. from Wikipedia paragraphs,

for transparency and traceability. The research group

continues to work on these topics and is expanding its

model to include the validation of facts that are lim-

ited in time. The model thus takes into account points

in time and validity periods and achieves greater accu-

racy (Qudus et al., 2023). The Kauwa-Kaate-System

supports querying based on text as well as on im-

ages and video and can be accessed via WhatsApp

or browser (Bagade et al., 2020). The input is com-

pared with content collected from fact checking sites.

In addition, a comparison is made with articles from

established, trustworthy news sources.

2.3 Systems with Combined Approaches

The detection and evaluation of disinformation has

long been studied in different ways with varying re-

sults. Many of the approaches presented above for

evaluating an article have proven to be successful

in their ﬁeld. However, they only analyze a certain

aspect of the article (e.g. sentiment, text style or

source). It should be highlighted that some of the

tools mentioned above are freely available and can be

used and combined in external systems. It has been

shown that the combination of different methods and

approaches has led to more successful results. Also,

the reliability of such tools depends to a large degree

on the explainability of the rating results. Some in-

novative tools already combine a few techniques, e.g.

XFake (Vosoughi et al., 2018), BRENDA (Botnevik

et al., 2020) and Fake Tweet Buster (Saez-Trumper,

2014). Table 2 shows the combined approaches in the

three systems in comparison to our platform.

The XFake system analyzes both attributes of the

article (e.g. author, source) and statements. Three

frameworks were developed for this purpose, with

one designed for attribute analysis (corresponds to

approach (2) Source Credibility), the second for the

semantic analysis of statements (corresponds to ap-

proach (4) Sentiment Analysis) and the third for the

linguistic analysis of statements (corresponds to ap-

proach (3) Style Analysis). In addition to the mere

classiﬁcation (fake vs. credible), XFake provides rea-

sons about decisive factors (e.g. tense language, in-

consistent sources), as well as relevant supporting ex-

amples and visualizations to facilitate interpretation.

XFake was trained and evaluated on a dataset of thou-

sands of political news claims veriﬁed by PolitiFact.

The XFake tool considers already many analysis cri-

teria (style, semantic and source analysis), but does

not take into account the dissemination of an arti-

cle. BRENDA recognizes and evaluates news sources

(corresponds to approach (2) Source Credibility) and

content (corresponds to approach (1) Fact Checking).

They help users to distinguish trustworthy informa-

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

414

Table 2: Systems combining different approaches.

Approach XFake Brenda FTP Our tool

(1) x x

(2) x x x x

(3) x x

(4) x x

(5) x x

(6) x

(7)

tion from disinformation through browser extensions

or in social networks. It uses a tested deep neural

network architecture in the background to automati-

cally check facts and present the result with evidence

to users by using an AI-classiﬁer to assess whether a

claim is likely to be true or false following a prelimi-

nary claim detection. At the same time, sources or ev-

idence (e.g. fact check websites) are presented to con-

ﬁrm the claim. The Fake Tweet Buster (FTB) web ap-

plication identiﬁes tweets with fake images and users

who regularly upload and/or spread fake information

on Twitter/X by combining reverse image search (cor-

responds to approach (6) Multimodal), user analysis

(corresponds to approach (2) Source Credibility) and

crowd-sourcing (corresponds to approach (5) Social

Signals) to detect malicious users. The FTB takes

into account the dissemination of an article, but con-

centrates on a limited area of application: identifying

tweets with fake content.

In our work, the focus is not on the mere reliabil-

ity classiﬁcation of an article. The tool developed by

us is primarily intended to support the user in forming

their own opinion on the trustworthiness of a news ar-

ticle. We follow an integrative approach by analyzing

the article using different methods and presenting the

results visually with suitable UI elements. In this way,

we incorporate several of the aforementioned criteria

(2, 3, 4, 5 and partly 1 claim detection) in a meaning-

ful combination to provide the user with a compre-

hensive evaluation of the article. To the best of our

knowledge, we are not yet aware of the existence of

any other such instrument.

3 SYSTEM ARCHITECTURE

The system architecture of the developed platform

with the technical system components involved and

their interfaces to each other is outlined in Figure 1. A

central controller coordinates all processes, addresses

all required analysis components, and forwards data

to the web interfaces. The controller also serves as an

interpretation layer preparing the results of the indi-

vidual services in such a way that they can be visually

presented in an understandable form. The workﬂows

from article analysis to the visual presentation of the

results can be summarized as follows:

Web Scanner. The web service continuously

searches for online sources on a topic deﬁned by

keywords and prepares the content found in a struc-

tured manner. The created article corpora can then

be searched for a speciﬁc article URL. On request,

all article features belonging to a URL are delivered

to the central controller. In addition, a separate text

extraction component is used to ensure an optimized

and complete extraction of the entire body text of the

article. For the prototype implementation, this text

extraction component is limited to a list of predeﬁned

online sources (news publishers). Selected article fea-

tures are prepared as part of the interpretation layer to

be directly used in the visual presentation of the anal-

ysis results. This includes general article information

(e.g. title, publication date, URL) as well as infor-

mation on the dissemination of the article (referenced

and referencing other articles).

Reliability-Rating. The component for analyzing ar-

ticle credibility assesses whether an article is more

similar to online articles from credible sources or to

articles from non-credible sources. For this evalu-

ation, different feature domains are analyzed (text

style, structure of the article website and article dis-

tribution network). The assessment of article credi-

bility is based on an AI transformer model, which is

pre-trained on the above-mentioned feature domains

using extensive training data, including the FANG-

COVID (Mattern et al., 2021) dataset. The credibility

of the article sources in the training data is estimated

using the NewsGuard rating (NewsGuard Technolo-

gies, 2025). The classiﬁcation results of the feature

domains are then comparatively weighted based on

evaluated quality factors (f1-scores) and sent back to

the controller as an overall assessment.

Claim Detection and Sentiment Analysis. A further

analysis component breaks down the supplied article

text into individual sentences. The individual sen-

tences are further analyzed with regard to their sen-

timent (positive/negative) as well as whether the sen-

tence represents a statement. The sentiment analy-

sis is based on an AI transformer model pre-trained

for the German language used to detect sentiment

depending on the overall context of the sentence.

The claim detection component uses a deep learning

model that also uses AI transformer architectures. For

the prototype implementation, the method was opti-

mized for the detection of statements in the context

of COVID-19 and MPox with the help of a manually

annotated training data set. The individual results are

also sent back to the controller at sentence level.

Online News Veriﬁcation: An AI-Based Platform for Assessing and Visualizing the Reliability of Online News Articles

415

Figure 1: Technical system architecture of the implemented platform for assessing and visualizing online news reliability.

Linguistic Analysis. The extracted article text is also

used for linguistic analyses. On the one hand, the text

is evaluated in terms of lexical complexity (MTLD

(McCarthy and Jarvis, 2010)) and readability (Flesch

Reading Ease (Amstad, 1978)). On the other hand,

a classiﬁcation determines whether the text is more

similar to an objective news article or to a subjective

opinion article. The classiﬁcation is based on the use

of certain function words (pronouns, adverbs, etc.).

With the help of a Support Vector Machine (SVM),

which was trained with a corpus of 600 labeled news

and opinion articles from 27 different publishers, sim-

ilarity values of the text to opinion and news articles

are calculated.

Database. The analysis results delivered to the con-

troller are stored in a database for persistent storage.

If an article URL to be analyzed already has analy-

sis results stored, these are loaded directly from the

database.

User Interface. The processed article features and

analysis results are ﬁnally transferred to the user in-

terface component (web frontend) visually presenting

the results. Analogously, user actions on the web in-

terface (e.g. entering a new article URL) are received

by the controller and processed accordingly.

Finally, a FHIR store (Fast Healthcare Interoper-

ability Resources) was designed as a backend compo-

nent to store questionnaire-based result data as part of

the evaluation (see chapter 5).

4 USER INTERFACE DESIGN

The graphical interfaces for presenting the analysis

results were created through iterative UI design pro-

cesses including multiple prototyping, testing and re-

ﬁnement steps. The ﬁnal interfaces created are brieﬂy

presented and described below.

The mode of article input can be selected on the

initial page. Either an existing URL in the database

or the title and text for the article analysis can be

entered. By clicking the button ‘article analysis’, a

query is sent to the technical components (controller

and database) and the article will be analysed on the

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

416

ﬂy. Once the analysis has been completed, the article

results are displayed in the user interface. In the sec-

tion ’General Information’ you ﬁnd basic information

(title, publication date, URL, source) on the analysed

article, if available (Figure 2, upper part).

The ’Credibility’ section displays the results

whether the article is more similar to online articles

from credible sources or more similar to articles from

non-credible sources (Figure 2, lower part). For this

assessment, the three feature areas ‘Text style of the

article’, ‘Distribution network’ and ‘Website struc-

ture of the article source’ are analysed separately.

The individual classiﬁcation results of each category

are then weighted comparatively and an overall sum-

marised score on a scale of up to 100 is given. The

external NewsGuard rating for the trustworthiness of

the article source can also be viewed by clicking on

‘Comparison with external source’.

Figure 2: Visualization of general article information and

results of reliability analysis.

In the ‘Language and Style’ section, the current

article is analysed in terms of its readability, its com-

plexity and its positive or negative sentiment. The re-

sults are presented on a scale from 1 (low) to 5 (high).

A radar chart shows the correlation of these three text

style areas (Figure 3, upper part). Here, the current

article can also be compared with three different com-

parison corpora. The comparison corpora are subsets

of the entire built-up article repository being formed

from the training data set (FANG-COVID) or from

articles previously analysed by the platform.

In the section ‘Classiﬁcation of Opinion vs. News

Text’ (Figure 3, lower part), a linguistic analysis us-

ing function words is carried out to classify to which

extent this article (red dot) is more likely to be classi-

ﬁed as a news article (green dots) or an opinion article

(blue dots). The visualisation - reduced to two dimen-

sions - illustrates the function word-based features of

Figure 3: Visualization of ’Language and Style’ analysis as

well as results of the news and opinion article classiﬁcation.

all texts on which the classiﬁcation is based.

In the section ‘Article Dissemination’, the articles

that refer to the analysed article (in-links) are listed

in a bar chart chronologically (Figure 4). Also, the

articles to which the analysed article refers (out-link)

are listed. The visualisation also illustrates the media

types (news portals, social media, other websites) of

the in- and out-links indicating the media segments in

which the article is distributed. If there are no further

links in the article, the graphic remains empty.

Figure 4: Presentation of article dissemination.

The section ‘COVID-19 or MPox claims (Claim

Detection)’ identiﬁes and highlights the claims or

statements relating to COVID-19 or MPox in the anal-

ysed article at sentence level (Figure 5). An addi-

Online News Veriﬁcation: An AI-Based Platform for Assessing and Visualizing the Reliability of Online News Articles

417

tional sentiment analysis colours individual sentences

that convey a positive or negative sentiment (polarity)

green or red. Beige stands for a neutral sentiment.

Figure 5: Visual output of components for claim Detection

and sentiment analysis.

5 EVALUATION

An evaluation was conducted to ﬁnd out how users

perceive the tool by a quantitative and a qualitative

section. The key ﬁndings are described below.

The application of the VERITAS demonstrator

resulted in a predominantly positive overall evalua-

tion. Even though the signiﬁcance or usefulness of

some displayed analysis criteria appeared less impor-

tant according to the initial impressions of the partici-

pants, the tool generated a positive overall impression

in terms of its presentation character – also with re-

gard to the diversity of the displayed analysis criteria.

The diverse analysis criteria presented are interesting,

each with different potential: (1) The highly complex

analysis of individual stylistic aspects at article level

can be used to reveal potentially implicit structures

(e.g., positive/negative sentiment) and to offer differ-

ent approaches for interpreting the texts (opinionated

vs. neutral). (2) The analysis of the article’s context

(e.g., dissemination) or of a fact seems to be an impor-

tant criterion. This involves not only the direct linking

of articles/posts to one another, but also the adoption

and manipulation of a topic/fact from one article in

other articles or posts (content similarity between ar-

ticles/posts). (3) Certain stylistic feature spaces such

as readability and complexity are considered less rel-

evant for the assessment of the disinformative nature

of an article.

A key ﬁnding was that it is difﬁcult or impos-

sible to clearly assess individual articles or excerpts

from articles as credible (“not fake”) or non-credible

(“fake”). Automatically checking the truthfulness or

reliability of content is very complex; even manually

evaluating the truthfulness of individual statements is

labor-intensive. Therefore, a tool like the VERITAS

portal should rather be used to analyze and compare

articles from different perspectives.

6 LIMITATIONS AND OUTLOOK

In terms of the variety of analysis criteria presented,

the platform made a positive overall impression on

users, although the signiﬁcance and usefulness of in-

dividual selected parameters are not easy to under-

stand at a ﬁrst glance. However it turned out, that

complex representations of analysis results, even if

the single analysis parameters can be seen as easy to

understand, lead to uncertainty, especially if a cor-

relation of these parameters is visualised (e.g. with

the radar chart). The work presented here provides a

valuable starting point, but it would beneﬁt from sim-

plifying the interface and the generalization of claim

detection through the addition of further training data.

Certain areas of analysis, such as claim detection, are

limited in their focus to the topics of COVID-19 and

MPox.

The following ﬁndings can be concluded from the

work on the project: It seems difﬁcult to categorise

an article as credible (non-fake) or not credible (fake)

solely based on its textual characteristics. This is due

to the fact that even less reputable portals and me-

dia often publish serious, factually neutral informa-

tion within the article. According to the current state

of the art, the truth or factuality of an article can only

be checked in comparison with external sources that

are as objective as possible. This requires the man-

ual labelling of facts or objective statements by edi-

tors or fact checkers in order to ensure that the ground

truth can be used for the further automated veriﬁca-

tion of the article’s content. Automatically checking

the truthfulness of article content is very complex, and

even manually assessing the truthfulness of individ-

ual statements is time-consuming. A tool such as the

presented platform should therefore not be used to as-

sess the fake character (factuality) of individual arti-

cles, but merely to analyse articles from different per-

spectives and make them comparable. A good way

of recognising to which extent an article has opinion-

forming or manipulative intentions is by analysing the

strength of opinion, sentiment and the credibility of

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

418

the source. Although the results of the analyses car-

ried out for this purpose should not be presented in de-

tail to not confuse users, the further technical develop-

ment of a component for the evaluation of the strength

of opinion of an article would be recommendable.

For more valid results, a large, unbiased new train-

ing data set with manually labeled manipulative and

non-manipulative articles would therefore have to be

set up in further developments. Analysing article

sources (e.g. news publishers) and the article distri-

bution network (especially the distribution in various

social networks) is also very important and should

be expanded in follow-up projects. In particular, the

dissemination analysis should extract facts within the

article text and link them to ofﬁcial announcements

and legislative texts similar to the Knowledge-Graph

based approach (e.g. press releases, legal texts, if nec-

essary also complete original quotations) in order to

show the provenance of the information and any alien-

ating editing strategies.

In the future, we see further potential applica-

tions on the presented platform like searching reliable

sources on specialised topics and checking the cita-

tion of references for scientiﬁc papers, as well as a

‘neutrality analysis’ of one’s own texts.

REFERENCES

Amstad, T. (1978). Wie verst

andlich sind unsere Zeitungen?

Studenten-Schreib-Service.

Bagade, A., Pale, A., Sheth, S., Agarwal, M., Chakrabarti,

S., Chebrolu, K., and Sudarshan, S. (2020). The

kauwa-kaate fake news detection system. In Proceed-

ings of the 7th ACM IKDD CoDS and 25th COMAD,

pages 302–306.

Botnevik, B., Sakariassen, E., and Setty, V. (2020). Brenda:

Browser extension for fake news detection. In Pro-

ceedings of the 43rd international ACM SIGIR con-

ference on research and development in information

retrieval, pages 2117–2120.

BR24 (2025). Faktenfuchs Faktenchecks. https://www.br.

de/nachrichten/faktenfuchs-faktencheck.

CORRECTIV (2025). Faktenforum. https://www.

faktenforum.org/.

Felbo, B., Mislove, A., Søgaard, A., Rahwan, I., and

Lehmann, S. (2017). Using millions of emoji occur-

rences to learn any-domain representations for detect-

ing sentiment, emotion and sarcasm. arXiv preprint

arXiv:1708.00524.

FullFact (2025). Full Fact. https://fullfact.org/.

Gerber, D., Esteves, D., Lehmann, J., B

uhmann, L., Us-

beck, R., Ngomo, A.-C. N., and Speck, R. (2015). De-

facto—temporal and multilingual deep fact validation.

Journal of Web Semantics, 35:85–101.

Hartwig, K. and Reuter, C. (2019). Trustytweet: An

indicator-based browser-plugin to assist users in deal-

ing with fake news on twitter.

Hui, P.-M., Shao, C., Flammini, A., Menczer, F., and

Ciampaglia, G. L. (2018). The hoaxy misinformation

and fact-checking diffusion network. In Proceedings

of the International AAAI Conference on Web and So-

cial Media, volume 12.

IFCN (2025). International Fact-Checking Network - Poyn-

ter. https://www.poynter.org/ifcn/.

Li, L. H., Yatskar, M., Yin, D., Hsieh, C.-J., and Chang,

K.-W. (2019). Visualbert: A simple and performant

baseline for vision and language. arXiv preprint

arXiv:1908.03557.

Lu, J., Batra, D., Parikh, D., and Lee, S. (2019). Vil-

bert: Pretraining task-agnostic visiolinguistic repre-

sentations for vision-and-language tasks. Advances in

neural information processing systems, 32.

Mattern, J., Qiao, Y., Kerz, E., Wiechmann, D., and

Strohmaier, M. (2021). Fang-covid: A new large-scale

benchmark dataset for fake news detection in german.

In Proceedings of the fourth workshop on fact extrac-

tion and veriﬁcation (fever), pages 78–91.

McCarthy, P. M. and Jarvis, S. (2010). Mtld, vocd-d, and

hd-d: A validation study of sophisticated approaches

to lexical diversity assessment. Behavior research

methods, 42(2):381–392.

NewsGuard Technologies (2025). News reliability

ratings. https://www.newsguardtech.com/solutions/

news-reliability-ratings/.

Poynter (2025). Politifact. https://www.politifact.com/.

Qudus, U., R

oder, M., Kirrane, S., and Ngomo, A.-C. N.

(2023). Temporalfc: a temporal fact checking ap-

proach over knowledge graphs. In International Se-

mantic Web Conference, pages 465–483. Springer.

Saez-Trumper, D. (2014). Fake tweet buster: a webtool

to identify users promoting fake news ontwitter. In

Proceedings of the 25th ACM conference on Hypertext

and social media, pages 316–317.

Shu, K., Cui, L., Wang, S., Lee, D., and Liu, H. (2019).

defend: Explainable fake news detection. In Proceed-

ings of the 25th ACM SIGKDD international confer-

ence on knowledge discovery & data mining, pages

395–405.

Tian, L., Zhang, X., and Peng, M. (2020). Fakeﬁnder: twit-

ter fake news detection on mobile. In Companion Pro-

ceedings of the Web Conference 2020, pages 79–80.

Vosoughi, S., Roy, D., and Aral, S. (2018). The spread of

true and false news online. science, 359(6380):1146–

1151.

Yang, K.-C., Ferrara, E., and Menczer, F. (2022). Botome-

ter 101: Social bot practicum for computational social

scientists. Journal of computational social science,

5(2):1511–1528.

Zellers, R., Holtzman, A., Rashkin, H., Bisk, Y., Farhadi,

A., Roesner, F., and Choi, Y. (2019). Defending

against neural fake news. Advances in neural infor-

mation processing systems, 32.

Online News Veriﬁcation: An AI-Based Platform for Assessing and Visualizing the Reliability of Online News Articles

419