Critical data, including reports of injured persons and infrastructure damages is shared on social media after disasters. Nevertheless, during the disaster, detecting valuable information from a huge quantity of social media posts is regarded as more challenging. Therefore, a Convolutional Neural Network-based Transfer Learning+Deep CNN_Taylor Archery Algorithm (CNN-based TL+DCNN_TAA) is devised for event detection using images and texts from social media. Firstly, event detection is performed based on the input image using a Deep Convolutional Neural Network (DCNN), which is tuned by a Taylor-based Archery Algorithm (TAA). Here, TAA is formulated by the incorporation of the Taylor series and Archery Algorithm (AA). Simultaneously, input text data is tokenized, and then feature extraction is executed. Next, sentiment classification for event detection is performed by Random Multimodel Deep Learning (RMDL) with Convolutional Neural Network (CNN)-based Transfer Learning (TL). Alternatively, keyword extraction is carried out from the input text. Next, the keyword graph is generated by considering the keywords as nodes, and then keyword graph clustering is executed. Lastly, event detection is performed, and majority voting is utilized for combining all the outputs. Furthermore, the experimental outcomes of CNN-based TL+DCNN_TAA measured a maximum precision, recall, F1-score, and accuracy of 90.754%, 93.766%, 92.235%, and 91.787%.
The datasets used as input for AI models are often orders of magnitude smaller than raw big data files after summarization and feature engineering. Therefore, it is feasible to compute models locally on a modest computer or in the cloud on a powerful virtual server. From a machine learning theory perspective, linear regression is a fundamental predictive model and is used as a building block of neural networks when combined with non-linear activation functions. Historically, database systems have integrated highly optimized algorithms to store, update, and analyze large-scale data, but with somewhat limited mathematical capabilities. As a consequence, AI analytics have progressively migrated outside database systems, mostly to the Python ecosystem, leaving many statistical packages behind. Motivated by this gap, and drawing on core database system principles, we introduce a generic parallel algorithm that accelerates AI model computations by integrating two fundamental database optimizations: block-based I/O and data partitioning for parallel processing. This is programmed in Python via a Single-File Multi-Pointer (SFMP) system architecture and multiprocessing. We theoretically and experimentally show this generic algorithm can be used to accelerate and scale linear regression beyond main memory limits with two complementary, well-established solutions: data summarization with sufficient statistics and stochastic gradient descent. We present an extensive experimental evaluation validating accuracy and showing significant I/O acceleration. We also analyze parallel speedup trends and determine optimal parameter settings with large datasets of diverse dimensionality. Finally, we characterize analytic computations for which our parallel I/O algorithm is a good fit.
Successful human resource management facilitates business process improvement. One of the key decisions is concerned with work assignment, which ensures a good match between human resources and process tasks. To do so, organizations need to be able to continuously review work assignment and understand its relation to resource performance—a challenging task due to the lack of up-to-date information and the complexity of organizational data. Recent advances in process mining make it possible for organizations to efficiently conduct organizational analytics using event logs, a form of data recording details of business process execution and human resource performance in reality. Yet, it remains an under-explored question how to apply process mining to analyze historical work assignment decisions and support future ones. In this paper, we investigate this research gap following the Design Science Research Methodology. We develop a process mining approach that can systematically discover human resource groupings using trace clustering and resource cluster analysis, infer relations between resource groupings and resource performance in process execution using statistical tests, and derive insights for future work assignment decisions. The proposed approach is implemented and evaluated through experiments on real-life datasets. Furthermore, a survey with fifty participants is conducted to validate the usefulness of the approach. This study contributes a novel process mining approach from the resource perspective and provides organizations with an efficient, data-driven method to review and improve work assignment decisions.
Augmented reality (AR) is a technology that embeds digital information onto real-world objects through the use of visual overlays, employing devices such as smartphones, tablets, or head-mounted displays. The goal of AR is to enrich human comprehension and interaction with the physical environment. The creation of AR software applications is complex and traditionally requires advanced coding skills. To simplify AR development, we propose ARWFML, a domain-specific visual modeling language for designing AR scenarios as executable AR applications, enabling users to define augmentations and AR workflows graphically. The research process described in this paper was conducted following the Design Science Research (DSR) paradigm and was organized into four design cycles. First, the modeling language was designed at a conceptual level, followed by the initial implementation of the language on the two-dimensional (2D) metamodeling platform ADOxx. Since 2D modeling tools are inadequate for three-dimensional (3D) modeling, MM-AR, a new web-based 2D and 3D modeling environment for augmented reality applications, was developed in the third design cycle. Additionally, the usefulness of the new modeling environment was demonstrated through use cases of the new modeling language on the platform, and the environment was evaluated across multiple facets, highlighting the potential of the modeling language for AR development. Finally, in the fourth design cycle, the execution of ARWFML models was addressed, proving that ARWFML models are executable as AR applications.
Process compliance ensures that processes adhere to requirements imposed on them. Failure to comply with legal and industry regulations can result in substantial fines, legal consequences, and reputational damage. Compliance-by-design approaches thus aim to ensure that every process execution is compliant, which is essential in domains where violations are unacceptable. However, existing approaches fail to achieve the dream of compliance-by-design due to abstraction and implementation gaps between regulatory documents, designed process models, and executed processes as well as performance limitations of model checking. In particular, data, time, and resource requirements are often annotated to process models, but not enforced during execution. This work aims to achieve more substantial compliance-by-design assurances at an execution-level abstraction using structural verification of executable process trees (process tree verification, PTV), and to identify in which cases structural verification cannot ensure compliance. We provide formal results for the comparison of structural and behavioral verification regarding computational complexity and trustworthiness of verification results. The formal results are complemented through a prototypical implementation that enables a comparative evaluation of PTV with two model-checking approaches and a quantitative evaluation using a custom data set.
In the digital era, the vast volume of unstructured textual data is growing across various domains. However, manually extracting the knowledge or generating question-answer pairs from these texts is difficult, and traditional rule-based or keyword-matching systems used in the context struggled with informal or multilingual text and failed to handle complex document structures, leading to error-prone outcomes. To mitigate these challenges, this research proposes an automated question-answering system termed the Pyramid Attention-enabled Deep Recurrent Bidirectional Language Model (PA-DRBiLM). The PA-DRBiLM architecture leverages the strengths of the Recurrent Neural Network (RNN) and the Bi-directional Long Short-term Memory (BiLSTM) to effectively understand the semantic relationships between the sequences, thereby leading to the generation of accurate and relevant question-answer pairs. Additionally, the Pyramid Attention enhances the model’s performance by a precise focus on relevant textual regions. The Multimetric RoBERTa Heterogeneous Knowledge Graph-based Feature Extraction (MRHKG) helps to extract deep contextual and semantic features crucial for high-quality question generation. The experimental outcomes on the SQuAD dataset for the context-question achieved a BLEU score of 0.946, CIDEr score of 0.946, and ROUGE score of 0.948, while for the question-answer system, a 0.950 BLEU score, 0.947 CIDEr score, and 0.943 ROUGE score were attained by the proposed PA-DRBiLM, which highlights the superiority against other models in generating relevant question-answer pairs.
As modern software-intensive systems grow in size, complexity and autonomy there is a demand for them to incorporate services that explain their actions. Such explanation services allow stakeholders, especially end-users, to develop trust that a system complies with its objectives and restrictions. Before such explanation services can be designed, their requirements need to be analyzed, in terms of both the kinds of stakeholder questions such services can entertain and the content of the answers they respond with. We propose a framework for goal-oriented modeling and analysis of explanation requirements. Explanation requirements are captured as stakeholder goals and are subsequently analyzed into specific explanation tasks and explanation interaction templates. By further extending the models with temporal and causal constraints, the resulting conceptual model can be used to generate answers to certain families of stakeholder questions. Formalization of the model allows both simulations of the explanation service and, under restrictive assumptions, a working prototype thereof. An empirical evaluation using two separate system cases shows that study participants, who role-play as explainees, find actions included in system logs presented to them worthy of explanations by various degrees, and also evaluate explanations rendered using our approach as useful with respect to their assumed explainee role.
Investor sentiment embedded in financial news and social media strongly influences market dynamics, but automatically identifying sentiment tied to specific financial aspects remains a challenging task. This is largely due to the scarcity of annotated datasets and the complexity of hierarchical aspect structures in financial text. This work addresses the identification of aspect categories and sentiments from social media posts related to finance and investing, and introduces a new model based on Hierarchy-Aware Global Model (HiAGM) and RoBERTa, FinRoBERTa-TRC2, to tackle this problem. FinRoBERTa-TRC2 is obtained via domain-adaptive pre-training on a subset of the Thomson Reuters Text Research Collection 2 and fine-tuned on the Financial Opinion Mining and Question Answering dataset. The model first predicts a hierarchical aspect category, which is used to predict the associated aspect sentiment score. Our approach follows a three-stage pipeline: domain-adaptive pre-training of RoBERTa on financial corpora, hierarchical aspect categorization through HiAGM, and final fine-tuning for sentiment prediction on FiQA. We find that domain-adaptive pre-training, in combination with an extensive hyperparameter search, leads to state-of-the-art performance. For aspect sentiment prediction, we find that this approach reduces the mean squared error by 5.6% compared to the current state-of-the-art on FiQA.
We propose a few-shot clustering framework for analysing COVID-19 vaccine hesitancy on Twitter/X. A keyword network is constructed from processed vaccine-related posts, and labels are propagated from a small set of annotated seed nodes using a novel damping mechanism able to exploit the statistical properties of the constructed network, combining local graph structure and global distance to seed nodes. Unlike standard propagation-based approaches that rely on a single global diffusion parameter, our method introduces a node-adaptive damping mechanism that combines local graph structure and global distance to the seed set, allowing the propagation strength to vary across nodes according to their structural role in the network. In contrast to feature-based graph neural baselines, the framework operates directly on the keyword network under an extremely low-label regime, without requiring external node attributes or pretrained embeddings. Beyond predictive performance, our methodology supports a semantic analysis operated through entropy-based and information-theoretic diagnostics on the full network. The proposed approach outperformed classical label-propagation algorithms when tested on multiple configurations and 20-trial evaluation regimes. The resulting clusters reveal predominantly negative sentiment highlighting themes of mistrust, safety concerns, and scepticism towards health authorities, alongside access- and convenience-related issues. Overall, the framework advances online public health surveillance, offering a scalable and interpretable tool for public health communication and policy design.
With the rapid advancement of social media and artificial intelligence, fake news emerges as a critical societal challenge, contributing to widespread fake news and societal polarization. To address these challenges, multimodal approaches are considered promising solutions that integrate diverse data modalities to capture complementary information. This study provides a comprehensive examination of the evolution of fake news detection models, emphasizing the distinctive potential of multimodal techniques. We propose a novel taxonomy for categorizing multimodal approaches, which is structured around three critical dimensions: feature extraction, fusion methods, and learning objectives. Our review begins with a detailed comparison of key datasets, highlighting their characteristics, limitations, and applicability to multimodal tasks. We then conduct an in-depth analysis of multimodal detection techniques, exploring their methodologies and performance. In addition, we explore emerging technologies such as generative AI for robust detection and possible ethical considerations, including privacy risks and model biases. To address challenges such as cross-platform generalizability and real-world scalability, we propose several future directions and opportunities for developing effective multimodal fake news detection systems. Through this comprehensive review, we aim to contribute to the ongoing efforts to combat fake news by providing valuable insights and actionable guidance for researchers and practitioners in the field.
Recent papers have introduced novel approaches to explain why a Predictive Process Monitoring (PPM) model for outcome-oriented predictions provides incorrect outputs. Moreover, they have shown how to exploit the explanations obtained using state-of-the-art post-hoc explainers to identify, in a semi-automated way, the most common features that induce a predictor to make mistakes and, in turn, to mitigate their impact and improve the accuracy of the predictive model. This work starts from the assumption that frequent control-flow patterns in event logs may represent important features that characterize, and therefore explain, a certain prediction. Therefore, in this paper, we (i) employ a novel encoding able to leverage Declare constraints in PPM and compare the effectiveness of this encoding with PPM state-of-the-art encodings, in particular for the task of outcome-oriented predictions; (ii) introduce a completely automated pipeline for the identification of the most common features inducing a predictor to make mistakes; and (iii) show the effectiveness of the proposed pipeline in increasing the accuracy of the predictive model by validating it on different real-life datasets.
Relational databases (RDBs) remain the backbone of data storage for most applications, yet lack the semantic expressiveness needed for advanced data integration and reasoning in modern environments. Ontologies address this gap by providing formal and machine-readable representations of domain knowledge. Although various approaches exist for transforming RDBs into ontologies, rule-based mapping methods have gained attention due to their systematic nature. However, existing approaches exhibit three critical limitations: (1) inability to handle junction tables with more than two foreign keys, (2) absence of formal validation methods, and (3) inadequate preservation of database integrity constraints. This paper presents a comprehensive solution for transforming RDBs into ontologies based on physical schemas using a schema-driven approach. Our contributions include: (1) a systematic comparison of existing rule-based mapping approaches identifying specific coverage gaps, (2) five novel or corrected mapping rules that extend coverage for complex database patterns including n-ary relationships and self-references, (3) formal representation of all mapping rules using first-order logic enabling transparent understanding and automated validation, and (4) a Z3-based validation pipeline that verifies rule satisfiability and consistency, with concrete experimental results across three domains. We validate our approach through an end-to-end implementation evaluated on database schemas from university management, healthcare, and e-commerce domains, producing OWL ontologies that preserve 15 categories of database constraints compared to 3 for W3C Direct Mapping. Z3 verification confirms logical consistency across all 11 encoded rules, 55 pairwise combinations, and 3 schema-specific instantiations with up to 118 ground facts This work bridges the gap between RDBs and semantic web technologies, providing a formally grounded foundation for schema-driven database-to-ontology transformation.
Audio-based sentiment and emotion analysis is an increasingly important area of investigation due to its broad applications in health, business, education, and human-computer interactions. Audio input data can be represented as transcribed texts, acoustic features, spectrograms, and multimodal representations for audio-based sentiment and emotion analysis. This paper surveys various techniques and algorithms used for audio-based sentiment and emotion analysis, exhibiting the benefits and limitations of different data representations. Additionally, we identify representative datasets for audio-based sentiment and emotion analysis. The primary contributions of this paper include a structured review of data representations for audio-based sentiment and emotion analysis, as well as an exploration of the suitability of these audio data representations for different tasks. By presenting a comprehensive overview of the state-of-the-art data representation methods in audio-based sentiment and emotion analysis, this paper aims to facilitate advancements in the field and inspire new research directions. This work offers a valuable resource for researchers, practitioners, and professionals interested in understanding and harnessing the power of audio-based sentiment and emotion analysis.
Large Language Models (LLMs) are increasingly being integrated into knowledge-based applications, leveraging their language understanding capabilities to construct and interact with Knowledge Graphs (KGs) in meaningful ways. However, effective utilization of LLMs in knowledge-based applications requires more than just language processing; it requires an approach that enables structured/linked data handling, semantic querying, and interaction. This paper presents a novel approach that utilizes LLMs with Retrieval-Augmented Generation (RAG) and AI agents to support the development of KG-based semantic applications. Specifically, the approach is implemented with a custom framework, namely LLM4KGen, and the use of LangChain and LangGraph frameworks, implemented specifically using two versions of Gemini, 1.5 and 2.0. The first prototype implementation of the LLM4KGen framework has been evaluated within the context of a digital culture semantic application. The presented approach involves deploying specialized AI agents capable of performing targeted tasks like KG generation, semantic query handling, and framework-specific systems engineering (semi-automated code generation), enabling dynamic interactions between LLMs, developers, and knowledge engineers, within structured data environments. The proposed approach integrates LLM-generated KGs in graph databases (Neo4J) enabling semantic querying and data retrieval through AI agents. The evaluation of the presented LLM4KGen framework focuses on assessing the effectiveness of agent-based LLM interactions for KG generation and retrieval, measuring the query accuracy and efficiency of the approach.
Recommender systems are essential tools for handling the vast amount of data generated daily. Their aim is to provide personalized recommendations to users by suggesting options that are likely to be of interest. In this paper, we investigate the performance of an LSH-based Recommender System in the E-commerce domain, which utilizes Locality-Sensitive Hashing (LSH) as the main indexing technique. We present the system architecture and the algorithms to build the LSH index (including MinHash construction) and to retrieve recommendations, and we formalize the user-item relation used to construct the index. Furthermore, we evaluate the impact on space utilization, processing time, and quality of recommendations generated, while varying the number of users used to construct the index and the signature length, using a real dataset from a medium-sized E-commerce platform in Brazil. Our findings reveal that the index produced by the LSH-based Recommender System is compact and can be easily stored in main memory, improving query performance. The index can also be created rapidly (in a few seconds), allowing for frequent updates to account for changes in the list of products and prices. Furthermore, the index enables fast retrieval of recommended products (in a few milliseconds) for E-commerce platforms with up to 1000 products.
In a world where information is exchanged at an increasing pace, knowledge becomes quickly outdated. Formal constructs that capture human knowledge, such as knowledge graphs and ontologies, need to be updated and evaluated to stay relevant and functioning. However, evaluating knowledge models is labour-intensive and prone to errors. This study addresses the challenge of automatically evaluating changes in existing knowledge graphs. We introduce syntactic and semantic metrics tailored for change evaluation. The metrics are implemented and validated through experiments on knowledge graphs across various domains. In these experiments, real-world changes are simulated by removing concepts and introducing faulty ones before evaluation with the syntactic and semantic metrics. The hypothesis is that such changes decrease the quality aspects of the knowledge graph: removing concepts influences syntactic qualities such as the structure of the model, while adding faulty concepts affects semantic qualities like model consistency. The validation results support this hypothesis, demonstrating that the introduced metrics effectively reflect the intended performance differences when making changes to the graph. Additionally, the experiments show that the size and domain specialisation of a knowledge graph influence how well the metrics detect changes. Overall, this study proposes a novel set of evaluation metrics and provides evidence of their effectiveness for assessing modifications to knowledge graphs across different domains. These metrics can help developers detect errors, highlight unintended side effects, and flag other quality changes that might otherwise go unnoticed.
With the rapid development of semantics-driven information systems and large language models (LLMs), constructing reliable knowledge graphs from multimodal data remains challenging due to semantic misalignment and feature inconsistency between modalities. This paper presents a novel text-oriented cross-modal extraction strategy. The proposed approach, named ToME, leverages a diagonal tagging mechanism combined with hierarchical attention fusion, allowing for unified modeling of both entities and relations through a structurally transparent design. Building upon pre-trained language models and multimodal representations, ToME adopts a concise structural framework that emphasizes strategy innovation and semantic adaptability. Experimental evaluation on the MNRE dataset verifies the effectiveness and stability of the proposed structure. This study provides a strategy-level perspective toward extensible and interpretable multimodal knowledge graph construction.
