The proliferation of hoaxes on the Internet and toxic messages (with ofensive or very negative content and high dissemination rates) constitute a current problem on the road to truthful and respectful information. Due to the enormous influence that social media have in the generation of opinion and as a channel of information for society, the eforts that various companies, organisations and institutions are making to detect and counteract the high volume of disinformation circulating on the networks are important. This project deals with the implementation of a proof of concept of a system for analysing the fairness of messages published through social media, built on the basis of various methods and algorithms from human language technologies. These methods and algorithms are the result of research that the participating groups have been working on for the last few years and are promising solutions for the determination of diferent levels of quality of publications in two fundamental aspects: their veracity and their toxicity. To address the proof of concept, activities aimed at the definition and integration of these technologies and their evaluation by stakeholders are proposed. This will make it possible to establish the responsiveness of these technologies to the needs of society and industry, as well as their viability to work towards higher levels of technological maturity.
lished in a digital media and with a journalistic format to detect the main parts of a news item written under with the novelty of not only characterizing the news, but the journalistic style of the inverted pyramid (the most also indicating which parts of the news text are or are relevant information is presented at the beginning of not trustworthy. The research from which this project is the news item) and within each part, the tool will detect derived is based on the fact that misinformation is cur- the relevant information by marking the 5W1H. Once rently going viral at high speed, intermingling true and the parts of the news and the essential information are false information to confuse the user of such information. identified, the tool will categorize those elements into a Given this premise, the proposed technology will be able reliability value based on a language model previously generated with a training corpus. Furthermore, depending on the reliability of the elements, the whole news item will also take an overall reliability value.
On the other hand, and given that the toxic and/or constructive content of a news item, as well as its comments can be indicators of whether or not a news item is false, a module (SocialTox) will be developed to automatically determine the degree of toxicity of a publication. To do this, an analysis of the language and virality of the content will be performed. In previous studies, it has been identified that negative content tends to be more viral than positive content, which could also be an important feature to determine that a news item is false, given the rapid spread of misinformation.
Both modules will be integrated in a common architecture, whose interface will be a backend with APIs.
Through calls to these API endpoints it will be possible to interact with the system taking, for example, as input a text (from a news item, a comment or comment thread together with an associated news item, a post in some social network, including or not also the comment thread...), a professional report will be returned indicating the degree of veracity, as well as the toxicity of the document (see Figure 1). framework for both the SocialTrust module (UA) and the SocialTox module (UJA). • Objective 3: Integrate in a single cloud service the algorithms and models that compose a system for measuring the honesty (trustworthiness and toxicity) of a message published in digital media (UJA, UA). • Objective 4: Define the evaluation framework of the proposed service by identifying the stakeholders, the necessary data and evaluation metrics, which enable the validation of the solution according to the needs of these groups (UJA, UA). • Objective 5: Perform an analysis of the tests carried out and obtain a final evaluation. This is a primary objective, as it represents the expected goal of any proof of concept (UJA, UA). • Objective 6: Determine the aspects that can be highlighted for subsequent technological maturation processes. The aim is to identify improvements that facilitate the evolution of the product to higher TRL levels (UJA, UA).
productive and social sector and, in some specific cases, could be a support tool for various sectors (journalism, 2. Goals education, administration, etc.), since its application is to support the detection of reliable information, as well As the origin of this PoC, we will start from an experi- as toxic content in the text. In its current state, it is able mental system generated in the origin project that uses a to give with high precision a value of veracity not only very limited set of training data, but that obtained very of the news itself, but also of the various information satisfactory and competitive results. Thus, the main ob- contained in it. In this way, following the annotation jective of the PoC is to have a tool that applies diferent structure already successfully defined in its current state, language models for the detection of the degree of confi- after an enrichment of the generated datasets with vedence and toxicity of the information. This tool will allow, racity and toxicity features and after a training process in future projects, its applicability to the value chain of with a much larger dataset, it will quickly allow to obdiferent sectors. With this proof of concept we intend tain an evidence report of the veracity and toxicity of to validate its incorporation in a real way in productive the information. The initial trainings of the tool have sectors such as journalism, politics, public sector, among been in Spanish and therefore currently that would be others. To achieve this general objective it will be neces- its language of application, but being a tool that is not sary to address the following specific objectives. These language dependent, in the future, with a training set in objectives are shared in a symbiotic way by the two re- other languages it could be applied to any other language. search teams, being the UJA team in the toxicity aspects focusing on the implementation and development of the 3. Scientific and Technical Impact SocialTox module, while the UA team will focus on the trustworthiness aspects oriented to the implementation and development of the SocialTrust module: One of the objectives of this project is the integration of advances achieved in real-world environments. Dif• Objective 1: Build a dataset large enough to gen- ferent types of collaborations with external entities will erate the most powerful language models adapted be explored to assess the possibilities of transferring the to the domain and language possible for the de- generated products and their impact. In essence, the velopment of both the SocialTrust module (UA) proposed proof of concept has real-world scope, and its and the SocialTox module (UJA). application in various scenarios, such as digital media • Objective 2: Train the tool with the dataset. As a or monitoring information on social networks, will conresult of this training, new language models will stitute a disruptive technology that enhances access to be obtained and adjusted to obtain an optimized information.
Natural language is the primary means of interaction 1. As a licensing product: Any consulting comin human societies. The rapid development of the In- pany providing content management solutions ternet in terms of volume and diversity of information, for news (news portals or online journalistic edicoupled with user access to this data, poses a significant tions) can benefit from marketing the tool as part challenge for retrieving and analyzing factual and sub- of their business solution. In this case, the tool jective content for specific purposes and representing provides a quality filter ensuring the reliability of them to gain knowledge. These new types of texts are content. highly subjective, and their automated treatment requires 2. As a SaaS (Software as a Service) product: Online specific methods from Human Language Technologies advertising agencies marketing their advertising (HLT). portals can benefit from having a quality seal that
Another factor to consider is that digital media has guarantees the prestige of the advertising space. created an ecosystem of spaces where content is created In this case, marketing would be done per service and consumed at increasing quantity and speed. How- provided. ever, this ubiquitous environment of interaction among members of current society harbors certain omnipresent 4.2. Description of the Project’s Proximity content that negatively afects the quality and freedom of information. Digital media has become a space where to the Market or Target End Users misinformation, hate speech, or abusive behaviors pro- The issue of misinformation is not exclusive to a particliferate, among other contents that can directly harm in- ular sector; rather, it afects virtually all productive and dividual users and society as a whole. Thus, this project social sectors. Sectors such as politics, public administrawill provide the modeling of digital content behavior and tion, healthcare, industry, brands, advertising, and culthe availability of a tool that, by applying diferent lan- ture have ample examples of detrimental consequences guage models, returns the reliability and toxicity level of misinformation caused by fake news when they enter of information. It will contribute, on one hand, to the their sphere of action. However, the primary afected detection, mitigation, and prevention of harmful digital sector and the origin of problems for other sectors is the content, towards cleansing social media on the Internet, journalistic industry. The inclusion of misinformation and on the other hand, to characterizing beneficial and in a media outlet’s editing, often caused by echoing untrustworthy content, thus contributing to ensuring a re- reliable sources, afects the credibility of the outlet and spectful, safe, and reliable communication environment. its editorial, seriously impacting its advertising capacity
For all these reasons, the project is expected to have due to the discrediting that may occur to a commercial
scientific-technical impact (both nationally and interna- brand appearing alongside news proven to be unreliable
tionally) in various fields, such as the creation and use of or toxic. The possible inclusion of misinformation in
advanced resources or the development, implementation, traditionally considered reliable media by their audience
and integration of specific methods and tools. All these poses a significant risk that the publishing industry tries
developments will have a strong impact on the scientific to combat by dedicating considerable time to verifying
community and society. Medium-term transferable re- sources and information. However, in the current state
sults are also expected, working with real practical cases. of this sector, time works against the business, as news
loses its value within a few hours of being published, and
4. Social and Economic Impact a media outlet that delays publication becomes outdated.
The traditional industry believed that news could have a
The project has a clear ethical and social orientation, lifespan of at least 24 hours, the time it took to edit a new
proposing the study of automated measures to com- newspaper, but the digitization of news and its
availabilbat toxicity in digital communications. Given the so- ity on the Internet dramatically reduces the news cycle.
cietal concern generated by certain contents, leading to For this reason, the industry needs automated tools to
the adoption of political and legal measures for their detect unreliable or toxic information, reducing the
veritreatment, our project will decisively contribute in the ifcation time for news by professional journalists while
medium and long term to the construction of safer digital increasing precision in determining content reliability.
environments, more beneficial for everyone. Under this premise, this proposal of PoC is
primarily framed in the journalistic industry sector.
According to the report "The Newspaper Publishing Industry"
4.1. Description of Impact and Social [
Benefits dustry has considerably altered its value chain due to The product can be commercialized from two diferent digitization and Internet growth, leading to a shift in perspectives: business from print journalism to online. In this new scenario, advertising goes from representing 43% of revenue sources (in 2010) to being the main source of revenue today. However, according to this report, online advertising is cheaper, and the audience is more fragmented and pays less attention to advertising than in the case of "written" advertising. All of this leads media outlets to the need for greater eforts in finding and retaining advertisers. This is one of the determining aspects to care for the article’s quality and thus ensure an advertiser a reliable space for their brand using tools like the one proposed in this PoC.
According to this source, the diferent actors participating in this sector, and therefore potential clients of our PoC, include especially the journalistic sector: print and broadcast media, and online media: web news, portals and news aggregators, user-generated news (blogs), social networks, and platforms.
The Figure 2 illustrates the value chain of the journalistic industry, as per the referenced report, highlighting the involvement of new actors. In this scenario, users take on a more participatory role, not only as assessors and judges of information but also as content creators. The proposed tool gains significance in this new context by acting as a filter for information whose origin is beyond the editor’s control, providing assurance for advertisers seeking to associate their product with a prestigious space.
Alternatively, another productive sector that could benefit from the tool is the food industry. Social networks and other media portals daily accumulate a vast collection of articles analyzing the merits and drawbacks related to the consumption of both generic and commercialized food products. These articles, widely circulated, create a network of influencers with a high number of followers who entrust their diet to the judgment of these analysts, causing radical changes in consumption habits.
This landscape is further influenced by commercial wars and other market interests [2]. Specifically, the II Study on Health Hoaxes [3], conducted by SaludsinBulos and Doctoralia, determines that hoaxes about food constitute 57% of false beliefs detected by doctors in consultations.
The proposed tool will allow the evaluation of the reliability of articles published about commercial products, generating efective reliability and toxicity reports against negative articles with the brand. It can also enhance the dissemination of quality articles, enabling companies to strategically manage communication with potential clients, where silence is not an option.
cialTOX (PDC2022-133146-C21), SocialTRUST (PDC2022-133146-C22) funded by MCIN/AEI/ 10.13039/501100011033 and by the “European Union NextGenerationEU/PRTR" competitiveness analysis of the media and content industries, 2012. [2] I. Lorenzo, Cómo afrontar la desinformación en la alimentación: Organizaciones sectoriales, empresas e instituciones se enfrentan al reto de las fake news, con estrategias de educación tecnológica y comunicación reputacional, Distribución y consumo 29 (2019) 62–67. [3] A.-A. de Investigadores en eSalud, Ii estudio sobre bulos en salud. encuesta a profesionales de la salud de españa, 2019.