The rapid proliferation of large language models (LLMs) has raised growing concerns about distinguishing between human-written and AI-generated text. This work addresses the task of detecting AI-generated content by evaluating the latent similarity between a given input text and an alternative response generated for the same prompt, either known or inferred. Accordingly, CLAID (Contrastive Learning for AI Detection) is proposed as a Siamese Neural Network architecture utilizing BERT encoders and contrastive loss to capture semantic similarity between text pairs. Unlike prior approaches that rely on explicit classification or domain-specific features, our method focuses on modeling pairwise similarity, enabling a flexible and model-agnostic detection framework. To evaluate the generalization capabilities of the system, a comprehensive multi-domain and multi-model benchmark comprising three diverse datasets (i.e., HC3, DAIGT, and OUTFOX), encompassing a wide range of text genres, prompt structures, and generative models, has been constructed. Experimental results show that the proposed model achieves near-perfect classification accuracy across both single-domain and mixed-domain scenarios, demonstrating strong robustness to domain shifts, prompt variability, and authorship ambiguity. The model also exhibits strong data efficiency, attaining high performance with minimal supervision.

Social media platforms have become arenas for political discourse, where political influencers and news media organizations actively employ rhetorical strategies to shape public opinion. However, how these strategies differ between actors and adapt to audience expectations remains underexplored. This study investigates the nature of digital persuasion and its impact on the audience engagement in political discourse on the X platform (formerly Twitter). Specifically we analyze how persuasive strategies employed by political influencers and news media accounts differ during a U.S. election event, and how types of persuasion correlate with the engagement the audiences. Our approach encompasses an analysis of the Aristotles persuasion framework that consists of three appeals: Ethos (credibility), Pathos (emotion appeal), and logos (logic and reasoning), along with sentiment, and social network analysis. The findings identify that political influencers tend to employ a hybrid strategy that combines ethos and logos to maximize user engagement and exhibit high linguistic homophily with their ego networks. In contrast, news media accounts predominantly rely on ethos-driven appeals and show limited rhetorical alignment with their audiences. These insights highlight how rhetorical adaptation and audience alignment differ between influencers and institutional actors, offering a new lens for understanding digital political communication.

The spread of misinformation and manipulative narratives is a significant challenge in today’s information landscape, shaping public opinion across diverse communities. Traditional framing detection methods are constrained by their post-hoc nature, limiting their ability to anticipate emerging framing tactics in real-time. To overcome these limitations, this work proposes a framework that combines Retrieval-Augmented Generation (RAG) models with Federated Prompt Tuning to generate and iteratively refine community-aware prompts. These prompts simulate how narratives derived from events, news, or factual statements might be reframed across different communities, enabling proactive identification of potentially manipulative content. Additionally, the framework exemplifies the exploitation of the framework to quantify the susceptibility of narratives to framing and support early warning systems for disinformation. Preliminary experiments conducted on a real-world dataset, including diverse community profiles, reveal the framework’s potential to combat the spread of disinformation and assess framing vulnerabilities in a timely and scalable manner.

In the evolving landscape of artificial intelligence, ensuring the robustness and explainability of machine learning models is valuable. This study presents an innovative method based on the Rough Set Theory and Principles of Justified Granularity to enhance the explainability of text-based classifiers, specifically in style-based news bias classification. The method helps understand why a classifier can be deceived with an Adversarial Attack. It leverages two levels of insight. The first level is independent of the specific classifier and consists of generating rules from a boundary region built with Rough Sets Theory starting from train data. The second level considers the behavior of a specific machine learning model in classifying manipulated observations and, starting from the classification results, constructs information granules of true positives and false negatives. These granules are representative of observations that deceived a classifier. By comparing boundary rules with information granules, it is possible to acquire actionable knowledge that is useful for making decisions on making a machine learning model more resilient. Results are evaluated with real data containing biased news. The success rate of adversarial examples generated using LLM to test classifiers on borderline cases, where minor textual changes cause false negatives, ranges from 45% to 68%.

The widespread utilization of social media for information consumption has significantly exacerbated the problem of information disorder. Recognizing the difficulty people face in discerning the truth, automated assistance is urgently needed. Current state-of-the-art approaches often involve fine-tuning existing models with contributions from domain knowledge bases. The black-box nature and interpretability issues of deep neural networks have increased interest in hybrid approaches, giving rise to deep neural fuzzy systems (DNFSs). This article presents the Hybrid Fact-Checking Framework leveraging a DNFS tailored to address the uncertainty inherent in fact verification tasks and enhance the reliability of model responses. The DNFS integrates a large language model with an adaptive neuro fuzzy inference system for automated fact verification. The framework utilizes relevant evidence from open-world and closed-world sources, leveraging deep language models and employing few-shot prompting without additional training to generate and justify verdicts. Including fuzzy rules and considering the trustworthiness and relevance of retrieved evidence enhances response reliability, thereby improving overall effectiveness and outcome interpretability. Experimental validations have been conducted on three publicly available datasets ranging in different domains of expertise: Climate-FEVER, SciFact, and FEVER. The results demonstrate that the proposed framework ensures better outcomes, transparency, and mindful decision-making.

Abstract The spreading of machine learning (ML) and deep learning (DL) methods in different and critical application domains, like medicine and healthcare, introduces many opportunities but raises risks and opens ethical issues, mainly attaining to the lack of transparency. This contribution deals with the lack of transparency of ML and DL models focusing on the lack of trust in predictions and decisions generated. In this sense, this paper establishes a measure, namely Congruity, to provide information about the reliability of ML/DL model results. Congruity is defined by the lattice extracted through the formal concept analysis built on the training data. It measures how much the incoming data items are close to the ones used at the training stage of the ML and DL models. The general idea is that the reliability of trained model results is highly correlated with the similarity of input data and the training set. The objective of the paper is to demonstrate the correlation between the Congruity and the well-known Accuracy of the whole ML/DL model. Experimental results reveal that the value of correlation between Congruity and Accuracy of ML model is greater than 80% by varying ML models.

In online learning, the dropout phenomenon is a relevant issue to address with practical solutions. Several data sets stimulate original, and resolutive data analysis approaches, demonstrating the importance of the dropout phenomenon. This study proposes a novel approach to predicting massive online open course (MOOC) students at risk of dropout stressing the need to consider the temporal dimension in the data log. The proposal aims to build a data-driven decision support system able to identify students at risk of dropout based on the conceptualization of such students' behavior and its evolution along the time dimension. The primary theoretical model behind the proposed method is the formal concept analysis, and its temporal extension (i.e., temporal concept analysis) for analyzing timestamped data and carrying out a timed lattice. The main result of the paper is a method to extract behavioral patterns of MOOC students at risk of dropout. Such patterns are defined as Time-based Behavior Rules extracted from the aforementioned timed lattice obtained through the preprocessing of MOOC platform log files. The resulting rule set can be easily integrated for implementing educational DSS, as shown in the last part of the paper. The conducted experiments reveal promising results in terms of F-score and students' monitoring time.

This special issue entitled "cognitive machie intelligence for Cyber-Physical Systems" addresses under-researched and controversial topics on new emerging themes of cyber-physical systems (CPS).

Summary Social media is assuming a crucial role in purchasing decisions, and most companies are using social media for marketing. Making sense of the unstructured information content shared by users along the time is an emerging challenge. We advise that the dynamic nature with which trends and user's interests evolve along the timeline requires the revising of well‐assessed methods to address, for instance, recommendation provisioning, information retrieval, and so on. Time‐awareness is crucial to more effectively estimate user's interests in the future to better address social media marketing attempting to increase the traction, for instance, posting the right message at the right time. This work defines time‐aware collaborative filtering for estimating users' interest along the time in Twitter. It uses text analysis services to semantically annotate tweets' content and to track concepts considering post frequencies along the time. A model‐based approach implementing K‐Nearest Neighbors is used to estimate user's similarity representing their profile by sampling user's interest with three different techniques: Vectorial Representation, Symbolic Aggregation Approximation, and Median. We show the experimental results comparing these techniques and performing model training in different time windows. The proposed approach is used to address some social media marketing questions.

With the emergence of new technology-supported learning environments (e.g., MOOCs, mobile edu games), efficient and effective tutoring mechanisms remain relevant beyond traditional intelligent tutoring systems. This paper provides an approach to build and adapt a tutoring model by using both artificial neural networks and reinforcement learning. The underlying idea is that tutoring rules can be, firstly, learned by observing human tutors' behavior and, then, adapted, at run-time, by observing how each learner reacts within a learning environment at different states of the learning process. The Zone of Proximal Development has been adopted as the underlying theory to evaluate efficacy and efficiency of the learning experience.

With the rapid development ot biomedical sciences, a growing amount of papers reporting new scientific findings are published and indexed in different unstructured biomedical data sources. In order to really appreciate and effectively benefit from the availability of this amount of data there is an urgent need to support the deployment of intelligent information services, such as: temporal trends and group detection, expert finding, review experts, link prediction, and so on. This need is even more stressed if we analyze dissemination activity of emerging scientific communities that are working on specific research topics in the field of biomedical science. Motivated by the fact that nanotechnologies are one of the key enabling technologies nowadays, in this paper we instantiate and contextualize the Time Aware Knowledge Extraction (TAKE) methodology, introduced in previous work, as a tool to analyze the activities of the nano-safety scientific community coordinated by the EU NanoSafety Cluster (NSC). This methodology enables us to extract timed association rules. To validate and give evidence of the goodness of these rules, a summary of the so obtained results of the analysis is provided identifying distinguishing features, and detecting emerging collaboration among the NSC's Working Groups and their members over the timeline.

Formal Concept Analysis (FCA) and its fuzzy extension have been widely used to arrange data into a lattice that is an effective data structure useful to address several aims, such as: data mining, ontology learning and merging, and so on. In literature it is possible to distinguish two main approaches to address fuzzy FCA implementation: the one-sided threshold and the fuzzy closure one. This work focuses on a specific definition of one-sided threshold algorithm and fuzzy closure one. Specifically, it shows that these methods can be unified, since the one-sided threshold approach can be seen as a specialization of the fuzzy closure. The lattice generated using one-sided fuzzy threshold approach is a substructure of the lattice generated using the fuzzy closure approach. In addition, an experimentation has been performed on both implementations of the fuzzy FCA, one-sided threshold and fuzzy closure. In particular, the results are compared in terms of running time and number of extracted fuzzy concepts by varying the t-norm function Łukasiewicz, Gödel, and Product.

Nowadays, many research works are moving toward the definition of models for human decision support systems within business process executions. Existing solutions, in general, do not take into account the context in which such processes run but they provide rigid models that could erroneously support decision-making activities when a different context needs to be considered. This work focuses on the definition of a framework to support and trace human decision-making activities, in business processes, when more heterogeneous decision-makers have to find a consensus to select one alternative among the others. One class of such processes is that of Innovation Processes. In particular, the main result described here is a Context-aware Fuzzy Linguistic Consensus Model, based on Fuzzy Logic, Semantic Web technologies and Reinforcement Learning, that considers heterogeneous decision makers with different levels of influence (assigned by considering their past decisions) in the context where the decision activity takes place.

Social media and networks are used by millions of people to share with their friends across the world: tastes, opinions, ideas, etc. The volume and the speed at which these data are produced make it a challenging task to discover meaningful patterns in the data. Nevertheless, very interesting business goals could be achieved collecting these data and performing analytics on social media data streams, such as: addressing marketing strategies, targeting advertisements, and so forth. We emphasize that there is a need to investigate and define suitable knowledge mining approaches to go beyond explicitly available metadata by analyzing unstructured data to provide intelligent analytics services. Specifically, in this paper we provide first results on applying OLAP analysis to multidimensional Tweet streams.

Microblog service has attracted much attention in big data analysis. Twitter statistics remark that the average number of tweets per day is greater than 100 million and many thousands of them happen every minute. It's an urgent challenge to face with such a large amount of collected tweets. This work defines an online query-focused Twitter summarization framework. It crawls and semantically indexes tweets exploiting wikification service. When a user's query is submitted, the system filters out the most relevant tweets. Then, a summarization algorithm browses the knowledge structure extracted by performing a Fuzzy Formal Concept Analysis on the given filtered tweets. Experimental results reveal good performances.

With the proliferation of different heterogeneous biomedical data sources and with the growing amount of their content available over the Web, there is, on one side, the need to support mashing and data integration and, on the other side, the more urgent need to relate literature and research results that are often enclosed in unstructured textual documents. Nowadays, ontologies have been used as a common access knowledge layer playing a crucial role to support categorized access to the information resources. Moreover, manual construction of a domain-specific ontology and content categorization is a labor intensive and a time-consuming process. This work focuses on the development of a novel biomedical ontology-driven multi-facets visualization to support categorized access to heterogeneous and unstructured biomedical data sources (e.g., PubMed, WikiGenes). Specifically, the framework relies on: knowledge extraction methodology, to automatically extract ontology exploiting the Fuzzy Formal Concept Analysis theory; and ontology matching strategy to find relation between extracted ontology and the available ones in the field of biomedicine (e.g., Ontology of Gene and Genomes, Gene Ontology, Protein Ontology). The evaluation will be shown in terms of Precision and Recall by using biomedical ontology concepts as input query to the multi-facets visualization engine.

Situation awareness computing employs sensor networks to collect large amounts of heterogeneous data in different and complex environments. The rapid development and deployment of sensor technology stress the problem related to the availability of too much and heterogeneous data. Last trend emphasizes the semantic annotation of acquired sensor data. Semantic sensor data provides machine understandable contextual information. In particular, the availability of semantic sensor data allows situation awareness in several application domains. This paper introduces a swarm-based approach to semantic web reasoning in order to identify situations. On one hand, fuzzy control has been employed in order to face with uncertainty of happening situations. On the other hand, Situation Theory has been used in order to model situation awareness. A multi agent swarm architecture enables to monitor complex environments by using spatially distributed autonomous sensors. An application scenario for bank intrusion detection has been described.

Context aware computing as well as wearable and ubiquitous computing often attain with pattern recognition on incoming sensor data. Recognizing more (useful) contexts requires more information about the context, and thus more sensors and better recognition algorithms. In order to enable logic inference on incoming data, the proposed work assumes that incoming data are represented by means of semantic languages (e.g., RDF, OWL, etc.). Nevertheless, in a context aware computing purely logic-based reasoning on context may not be enough. So, the work introduces soft computing techniques to approximate context recognition. Specifically, this paper introduces an approach to context analysis and recognition that relies on f-SPARQL[1] tool, that is a flexible extension of SPARQL. In particular, in this work a JAVA implementation of f-SPARQL and the integrated support for fuzzy clustering and classification are discussed. This tool is exploited in the architecture that foresees some task oriented agents in order to achieve context analysis and recognition in order to identify critical situations. Finally, a simple application scenario and preliminary experimental results have been described.

Assessing employees' competences to properly support Competence-based Management processes (e.g. Career Development, Workforce Planning, etc.) in Organizations is a complex task. Difficulties concern with both the right assessment methodology and the most effective tools. Moreover, the assessment process is time-consuming both for assessors and assessees and often it is performed at the wrong time with considerable costs for external resources. This work proposes a novel approach, based on semantic technologies, to enhance competence assessment in Organizations by analysing content produced, tasks completed and professional relationships established by employees in their day by day activities at the workplace.

Context aware computing attains environments monitoring by means of sensors in order to provide relevant information or services according to the identified context. Nowadays, ad hoc wireless sensor networks for medical purposes are playing an increasing role within healthcare. Specifically, Body Sensor Networks (BSN) and Wireless Sensor Network, are being designed for prophylactic and follow-up monitoring of patients e.g., at home, at hospital, and so on. This work defines a framework aimed at proactively providing personalized healthcare services by performing sensor data analysis in order to recognize specific user's context. In particular, the approach is strongly based on the synergy between semantic formalisms and soft computing techniques. Semantic Web formalisms are exploited to model healthcare services and context. Soft computing techniques are applied in order to support activity of unsupervised context analysis and semantic service matchmaking. Specifically, Fuzzy Logic enable us to automatically characterize the context and to consequently find the set of healthcare services among the available ones that approximately meet the user's context. Experimental results shows performance in terms of services matchmaking.

In the area of ambient intelligence there is a need to address user needs according with context features. Recently, the synergy between context aware computing and collaborative filtering is leading to enhance recommender systems with capabilities always nearer to user needs. Specifically, in the domain of tourism it is useful to proactively suggest right sets of attractive locations, events and so on. This work defines a context aware recommender system aimed at suggesting pertinent points of interest (POIs) to tourists. In particular, the approach is strongly based on the synergy between soft computing and data mining techniques. The general framework integrates user profiles, history of social networking and POIs data. Then by defining collaborative filtering approach on the history meaningful POIs are extracted. Indeed, soft computing techniques are mainly applied in order to support activity of unsupervised users and POIs classification. On the other hand, data mining techniques are exploited in order to extract rules able to associate user profile and context features with an eligible set of recommendable POIs. Experimental results show performance in terms of recommendations accuracy.

Applying best available evidences to clinical decision making requires medical research sharing and (re)using. Recently, computer assisted medical decision making is taking advantage of Semantic Web technologies. In particular, the power of ontologies allows to share medical research and to provide suitable support to the physician's practices. This paper describes a system, named ODINO (Ontological Disease kNOwledge), aimed at supporting medical decision making through semantic based modeling of medical knowledge base. The system defines an ontology model able to represent relations between medical disease and its symptomatology in a qualitative manner by using fuzzy labels. Medical knowledge is defined according with physician experts members of INMP (National Institute for Health Migration and Poverty). The main aim of ODINO is to provide an effective user interface by using ontologies and controlled vocabularies and by allowing faceted search of diseases. In particular, this work mashes the capabilities of Description Logic reasoners and information retrieval techniques in order to answer to physician's requests. Some experimental results are given in the field of dermatological diseases.

Situation awareness is crucial factor in decision-making. It involves monitoring and identification of relationships among objects in collaborative dynamic environments. In the domain of Airport Security one of the main needs is to support the security operator to manage in real-time risk scenarios in the airside. This work relies on a cognitive approach to model the awareness ontology and introduces an agent-based architecture to address the problem. In particular, in order to model situation awareness the work instantiates the generic Situation Theory Ontology(STO) in the specific domain of airport security. Furthermore, some task-oriented agents allow to distribute the information in order to achieve better performances.

Nowadays, in enterprise environments there is a wide and consolidated utilization of software for the human resource management providing functionalities like organizational management, personnel development, training event management, etc. that lay upon a competencies repository mostly populated through expensive and inefficient data entry activities. The new trends in Web 2.0 see a paradigm namely Enterprise 2.0, for supporting business activities within organizations. Web 2.0 is mainly exploited to sustain collaboration, information exchange and knowledge sharing. This work introduces an agent-based framework for the dynamic refinement of employees' competencies profiles by analyzing and monitoring collaborative activities executed through Enterprise 2.0 tools (e.g. corporate blogs, enterprise wikis, etc.). A fuzzy extension of Formal Concept Analysis model supports the elicitation of implicit knowledge and the content structuring into a conceptual representation. The resulting concept-based organization of initial user-generated content will be exploited to provide automatic hints to human resources (HR) managers in order to support them in making safer decisions that involve employees' competencies.

The current “semantic” generation of Web strongly lies in sharing knowledge rather than linkages among digital resources. The Semantic Web represents an effective infrastructure based on ontologies, languages and tools to enhance visibility of knowledge on the net. The imprecise nature of knowledge often requires fuzzy techniques to coherently represent the imprecise and uncertain information of the real world. It is indubitable that many decision making problems within business, industrial and web applications are solved by fuzzy approaches, especially by exploiting fuzzy control. In order to integrate fuzzy knowledge in the Semantic Web, appropriate formal schemas are introduced for describing fuzzy data types and uncertainty information. In particular, this paper presents an OWL-based upper ontology, called OWL-FC (Ontology Web Language for Fuzzy Control) which provides a set of ontological constructs for defining semantic specification of Fuzzy Control. The OWL-FC ontology represents a straightforward contribute to support automation in discovery, usage and interoperability among a large number of fuzzy controls; built-in markups for the fuzzy controls enable the natural integration in description logics-based reasoners and guarantee the specification of fuzzy concepts which do not depend on the application domain.

The huge growth of data on the Web and the requirement of semantic content analysis make the knowledge management and data mining very difficult activities. The knowledge elicitation, codification, and storage need not trivial techniques to improve formal information structuring on the Internet. Ontologies provide conceptualization and processing knowledge, sharing of consolidate understanding, reusing of domain knowledge codification for many Web applications. Manual construction of a domain-specific ontology is an intensive and time-consuming process, which requires an accurate domain expertise, because of structural and logical difficulties in the definition of concepts, as well as conceivable relationships. At the same time, the ontology visualization process requires similar endeavors to support ontology management, exploration, and browsing. This work describes an automatic method for ontology design from the content analysis of Web resources. The approach exploits a fuzzy extension of formal concept analysis model for structuring the elicited knowledge, viz. concepts and relations embedded in the resources content. Final result is an effective ontology visualization through a navigable, facet-based view of the built ontology across the extracted concepts and their own population. Furthermore, the approach proposes a simple labeling of ontology concepts through a sketched and intuitive process. © 2010 Wiley Periodicals, Inc.

In recent years, the success of Semantic Web is strongly related to the diffusion of numerous distributed ontologies enabling shared machine readable contents. Ontologies vary in size, semantic, application domain, but often do not foresee the representation and manipulation of uncertain information. Here we describe an approach for automatic fuzzy ontology elicitation by the analysis of web resources collection. The approach exploits a fuzzy extension of Formal Concept Analysis theory and defines a methodological process to generate an OWL-based representation of concepts, properties and individuals. A simple case study in the Web domain validates the applicability and the flexibility of this approach.

Nowadays, the emphasis on Web 2.0 is specially focused on user generated content, data sharing and collaboration activities. Protocols like RSS (Really Simple Syndication) allow users to get structured web information in a simple way, display changes in summary form and stay updated about news headlines of interest. In the e-Learning domain, RSS feeds meet demand for didactic activities from learners and teachers viewpoints, enabling them to become aware of new blog posts in educational blogging scenarios, to keep track of new shared media, etc. This paper presents an approach to enrich personalized e-learning experiences with user-generated content, through the RSS-feeds fruition. The synergic exploitation of Knowledge Modeling and Formal Concept Analysis techniques enables the definition and design of a system for supporting learners in the didactic activities. An agent-based layer supervises the extraction and filtering of RSS feeds whose topics are specific of a given educational domain. Then, during the execution of a specific learning path, the agents suggest the most appropriate feeds with respect to the subjects in which the students are currently engaged in.

This paper describes a system for supporting the user in the discovery of semantic Web services, taking into account personal requirements. Goal is to model an ad-hoc service request by filtering semantic specifications rather than the exploitation of strict syntax formats. Adaptive agent-based techniques help the user to compose his Web service request, exploiting the semantic annotation of the browsed Web resources. This annotation reflects concepts or ontological terms that are relevant for the user services request formulation. Once the request is formulated, the system returns the list of semantic Web services that match the query input and output concepts.

The evolution of the Semantic Web promises infrastructures for the semantic interoperability of Web Services. Hindrances in the service discovery, composition and execution are often of syntactic nature: the difficulty in the interpretation of inputs, outputs or other nontrivial statements does not favor to find eligible advertised services which appropriately meet the consumer's demand. This paper deals with the semantic matchmaking focusing on the ontology mismatch problem: concepts appearing in the services description are compared at semantic level in order to profit by the semantic similarity existing among entity classes (i.e. concepts). The approach is based on a multi-agent architecture and exploits fuzzy techniques to represent the multi-granular capabilities of a web service. The semantic similarity among concepts supports the clustering of the advertised services and improves the quality of the retrieved results, given a request.