Applying Large Language Models (LLMs) to whole-page semantic element detection in e-commerce is prohibitively expensive: typical product pages contain hundreds to thousands of DOM elements, and LLM costs scale linearly with input tokens. We propose a hybrid approach that uses a Graph Neural Network (GraphSAGE) as a pre-filter to reduce the candidate set before LLM processing. Our preliminary results on the Klarna Product Page Dataset show approximately 90
Open Government Data (OGD) publishing is often manual and locally heterogeneous, yielding inconsistent dataset quality and weak traceability of pre-publication decisions. SODPUB is a BPMN-orchestrated workflow-as-a-service that targets interoperability-oriented improvement before portal upload. Data providers submit CSV datasets via REST; the pipeline applies checks derived from W3C guidance, normalises tabular structure, generates CSV on the Web metadata, and produces DCAT-compatible catalog descriptions. Each proposed change is explained and versioned, and the workflow exports a machine-readable process trace aligned with PROV-O so that decisions remain auditable. The current implementation provides an executable BPMN skeleton, initial rule catalogue, and RDF trace generation; REST endpoints and portal connectors are under active development. Planned evaluation measures include acceptance rates of proposed fixes by portal staff, reduction of detectable structural and metadata issues, time per publication, and perceived transparency. The expected outcome is a reusable reference workflow that improves quality while preserving publisher release responsibility.
We introduce SmartScrape, a neuro-symbolic framework for extracting structured information from the web. The task addressed is to extract records that follow a predefined schema from semi-structured sources, such as product web pages. The system integrates a Graph Neural Network (GNN) that scores content elements with an Integer Linear Programming (ILP) constraint solver, which guarantees schema consistency, uniqueness, and additional logical and presentation constraints. We illustrate the method on a basic (product, price) extraction scenario. The implemented system outputs visual proof-carrying justifications, supports drift detection using the stability metric σ (P) , and we also include an ablation study contrasting ILP-based and greedy reasoning modes. On a set of 100 pages, SmartScrape attains 97
Various reports reveal that Web Accessibility remains one of the major concerns that still requires additional awareness. In addition to providing development support in IDEs for creating Web pages, post-development support can also be provided by using Web augmentation techniques. However, the scripts for augmenting Web pages to improve such aspects must currently be developed manually and may require continuous adaptation to the evolving Web sites. In this work, we tackle these challenges by integrating LLMs into the Web augmentation process to improve accessibility concerns on the client side. In particular, our approach first performs an LLM-based accessibility evaluation of Web pages, and based on this feedback, the LLM is instrumented to resolve reported issues.
The quality of life of aging adults has improved over time through the incorporation of technological innovations in their homes, such as voice assistants or monitoring systems. However, current systems do not detect certain patterns that indicate dementia or mild cognitive decline due to limitations in the level of detail of the collected data or in the interpretability of the information already available. In this context, IoT and IoMT devices can be used to collect information that allows automatic detection of possible patterns or alterations. Integration of collected data with artificial intelligence models can improve the interpretability of the data and the detection of behavior patterns that indicate cognitive decline. This thesis aims to design an architecture that uses data from IoT/IoMT devices to identify behavior patterns using artificial intelligence models, with the aim of preventing dementia or mild cognitive decline in aging people through early detection and monitoring strategies.
While digital accessibility is a fundamental human right, modern web systems often fail to meet legal standards due to inadequate developer tools and persistent barriers for users with visual impairments. This research proposes an Accessibility by Design framework that integrates accessibility considerations into both the development and consumption phases of web systems. The framework is composed of two complementary components: an AI-assisted development environment that leverages Large Language Models to support developers in auditing and generating accessible web components and a web-based multimodal system that translates visual STEM data into multisensory representations through sonification. The study evaluates how integrating accessibility considerations into product design enhances usability while ensuring equal access for all users regardless of their abilities.
Evaluating Knowledge Graph Question Answering (KGQA) systems is increasingly unreliable due to static benchmarks, evolving Knowledge Graphs (KG), and memorization effects in Large Language Models (LLM). We present DynBench Generator, a web frontend for automatically generating new KGQA benchmark datasets that both preserve the original question and query complexity and are resistant to LLM memorization. In this demo, we showcase an interactive platform that enables on-demand generation and inspection of memorization-resistant question-query pair for fair and reproducible KGQA evaluation.
Predictive maintenance systems rely on machine learning (ML) to anticipate equipment failures, but domain engineers often cannot explain why a prediction was made, limiting trust and adoption. Knowledge graphs (KGs) can connect predictions with structured domain knowledge, but building industrial KGs requires semantic web expertise domain engineers lack. This proposal addresses three problems: ML predictions lack human-readable explanations, KG construction has no reproducible method with formal validation, and no web-based approach enables non-technical users to build and query KGs. The proposed Web Engineering method integrates large language model (LLM) assistance across the KG lifecycle, from ontology elicitation and data validation to natural language querying and explanation generation. The method is realised through a web portal following End-User Development (EUD) principles, enabling domain engineers to construct ontologies, ingest data, query the KG, and inspect prediction explanations. Evaluation uses CMP semiconductor manufacturing, with a second use case planned in wind turbine monitoring.
Applying machine learning to web pages is challenging due to the need to interpret HTML together with associated resources and perform rendering to obtain a meaningful visual and layout-aware representation. As a result, machine learning over web content remains comparatively underexplored. In this paper, we present a platform for visual-aware representation and machine learning over web pages based on the open-source rendering tool FitLayout. The platform provides a server capable of rendering web pages, explicitly capturing their visual and structural properties in an RDF-based representation, and persisting the rendered documents in an integrated storage. The processing pipeline is controlled via a REST API, while SPARQL queries are used to retrieve structured data suitable as input for machine learning algorithms. By explicitly modeling rendered web pages, including fine-grained layout details, the platform enables dataset sharing and supports the reproducibility of experimental results. The architecture supports the complete dataset preparation workflow, from web page collection and rendering through preprocessing and annotation of content elements to downstream learning tasks. We further provide a Python client library that integrates the platform with standard machine learning workflows. As a demonstration, we show how rendered web pages can be transformed into graph-based representations and used to train graph neural networks for recognizing key content elements, illustrating both the applicability of the approach and the reproducibility of the results.
Simulation-based research has been widely conducted to ensure the safe operation of cyber-physical systems (CPS), such as autonomous vehicles. However, most simulation-based studies focus on the functionality of the ego vehicle during the development phase of the DevOps lifecycle. As a result, simulations performed during the development phase may differ subtly from real-world conditions and cannot cover all possible scenarios. Although simulations should also be conducted during the operations phase to ensure more rigorous safety assurance, significant challenges remain due to the limited computing resources of the ego vehicle and the virtually infinite number of possible driving scenarios. To address these challenges, this study proposes a simulation method that leverages a cloud–edge computing environment by partitioning the timeline, thereby overcoming the computational limitations of the ego vehicle. In addition, the proposed approach generates a comprehensive set of semantic trajectories—including illegal behaviors of non-player character (NPC) vehicles—using an ontology.
The huge amount of academic data is challenging for research processes, but a rigorous review of prior studies remains a must for scientific progress. While Large Language Models (LLMs) are seen in several studies as potential help, it is evident that they have limitations, including the generation of fake references and misclassifications of research. Existing ways to test LLMs, such as manual checks or text comparisons, are tedious and costly. Therefore, automation support to systematically assess them for this particular context is needed. To shed more light on this area, this paper presents a methodology and a Web-based platform for assessing LLM-based literature search. The methodology starts by using human-composed systematic literature reviews (SLRs) as a benchmark. The platform provides search prompt templates, sends instantiations of them to LLMs, and then systematically validates if the generated references actually exist in authoritative databases. Subsequently, the platform compares the LLM’s findings with the set of included studies in the benchmark SLRs we use as ground truth. We specifically measure three metrics: how often LLMs generate fake references (existence check), how relevant the reported papers are compared to the ground truth (precision), and how well the LLMs have found all the important papers (recall). Ultimately, this work provides a flexible methodology, a practical platform, and reliable benchmarks to determine the current status of LLMs for academic literature search.
The increasing adoption of knowledge graphs in data-intensive applications contrasts with the limited availability of developer-friendly interfaces for managing RDF data. This paper presents an approach for automatically generating RESTful API endpoints from Shape Expressions (ShEx) schemas. Given only a set of ShEx schemas and the endpoint of a triplestore, our approach derives endpoints for creating, reading, updating, and deleting shape instances, as well as attribute-level endpoints for more fine-grained access. The generated REST API is implemented as a layer on top of the SPARQL endpoint, thereby abstracting query formulation from clients while maintaining direct access to the underlying triplestore. To facilitate documentation, client generation, and integration with existing tooling, OpenAPI specifications are automatically generated for all derived endpoints. In contrast to prior approaches, which typically rely on additional configuration, such as hand-crafted query templates or manual mappings of endpoints to SPARQL queries, the proposed approach does not require additional configuration. By exposing RDF data through conventional REST-style endpoints, our approach aims to lower the entry barrier for web developers unfamiliar with Semantic Web technologies.
The foundation of scientific research is the comprehensive analysis of existing literature. However, the exponential growth of published research leaves scientists increasingly overwhelmed, making it difficult to maintain a complete overview of the state-of-the-art or to discover hidden synergies between studies. The root of this problem lies in the traditional format of scholarly communication: crucial knowledge about applied methods, datasets, and metrics remains locked in semantically unlinked documents. While this format is optimal for human reading, it is highly inefficient for machine processing. Even modern AI research assistants frequently fail to provide complete, verifiable, and hallucination-free answers to complex research queries. To enable true machine-assisted exploration and verification, scholarly literature must be transformed from isolated documents into deeply interlinked, machine-readable structures. To achieve this, we introduce an agentic AI framework that automatically extracts key research entities, seamlessly interlinking the literature by mapping them to standard knowledge bases via unified URIs.
Sexism detection in social media memes remains challenging due to the inherent complexity of multimodal sources and the possible disagreement among different annotators. Results from the EXIST 2025 Challenge show that current systems often perform unevenly across hard and soft evaluations, where hard settings involve binary sexism prediction whereas soft ones entail probabilistic estimation of the judgment of multiple annotators. We propose a unified approach that jointly models both label types by combining soft label learning under an ensemble strategy, in which two models are trained on distinct, class-unbalanced dataset partitions, with supervised learning of hard labels on borderline cases. Experiments on the EXIST Meme Dataset demonstrate that our approach achieves an enhancement of +7.73 : This paper includes examples of hateful, explicit and sexist language presented for illustrative purposes.
Governance of AI systems is becoming a critical issue as both societal and regulatory circumstances demand requirements such as safe information handling, transparency, and explainability. This is complicated by the widespread use of third-party-operated LLMs, which limits organizations’ control over the embedding and enforcement of governance and organizational policies directly within the agent. Consequently, users are required to exercise caution or are prevented from using these agents when interacting with web applications that contain potentially sensitive information or require transparent and explainable processing. We propose a framework and architecture for embedding governance and organizational policies for data handling and explainability within web agents based on local, tool-calling small language models. This core will be expanded with governance-aware hybrid routing to remote LLMs for non-critical tasks, balancing the high capabilities of large, third-party provided systems with the safe and transparent environment established through local processing.
JavaScript (JS) has been the de facto standard for developing interactive web applications for nearly 30 years. Although it was initially a significant step forward in enabling interactive web pages, JS reaches its limit for complex web applications. Today, games, maps, spreadsheets, and image manipulation require more performance than JS can offer. This is where developers already use WebAssembly, an efficient binary format and compilation target for high-level languages. However, WebAssembly is typically used only to extend JS-based applications. With reaktor we promote WebAssembly, exemplified through Go, from its supporting role to a first-class citizen of the web. reaktor is a Go framework that adds templating to Go, enables compiling it into efficient WebAssembly and automatically generates the necessary JS code to interact with the APIs provided by the browser. Through the evaluation of reaktor, we identified inhibitors of the adoption of WebAssembly-aware high-level languages for full-fledged web development.
Energy-related Internet of Things (IoT) devices such as inverters, storage systems, heat pumps, and Electric Vehicle (EV) chargers are increasingly connected and remotely controllable, yet their integration logic is often tightly coupled to specific standards, service providers, or vendor clouds. As protocols and business processes evolve, this coupling can lead to vendor lock-ins or complex gateway-based compatibility layers, which make local fallbacks or user-defined automation difficult. At the same time, most residential energy automations are conceptually simple and follow a consistent pattern: ingest a signal (e.g., dynamic tariffs or local PV surplus), evaluate it against a policy, and actuate a device accordingly. In this paper, we explore how these automations can be enabled on resource-constrained hardware using standardized, machine-readable interfaces. To this end, we present a software architecture that realizes a consumer-side subset of the Web of Things (WoT) Scripting API on embedded devices, including consume, readProperty, writeProperty, and invokeAction. The architecture supports asynchronous I/O via an on-device JavaScript executor and exposes additional device functionality to scripts through host bindings. To quantify the limits of our approach, we evaluate practical resource costs on an ESP32 using an energy automation scenario and report on firmware footprint and memory usage. Overall, the results indicate manageable firmware and memory overhead, demonstrating the feasibility of deploying standards-based automation logic directly on constrained IoT devices.
The increasing deployment of autonomous AI agents on the web is hampered by a fundamental architectural misalignment: current web applications are designed primarily for human interaction, forcing agents to infer affordances from user interfaces through brittle, inefficient, and insecure interactions. To address this, we introduce VOIX, a web-native framework that enables websites to expose reliable, auditable, and privacy-preserving capabilities for AI agents through simple, declarative HTML elements. VOIX introduces and tags, allowing developers to explicitly define available actions and relevant state, thereby creating a clear, machine-readable contract for agent behavior. This approach shifts control to the website developer while preserving user privacy by disconnecting the conversational interactions from the website. Ultimately, this work provides a foundational web engineering mechanism for realizing the Agentic Web, enabling a future of seamless and secure human-AI collaboration on the web.
The Vietnamese stock market, as an emerging market, exhibits high sensitivity to information dynamics, which directly influence investor behavior. Manual analysis of financial news is not only time-intensive but also prone to subjectivity and cognitive bias, thereby increasing investment risk. To address these challenges, this study proposes and develops a big data–driven platform that integrates artificial intelligence (AI) within a comprehensive end-to-end architecture. The proposed system automates the entire workflow - from real-time, multi-source news acquisition and preprocessing to deep analysis and strategic insight generation. Leveraging big data technologies such as Apache Kafka and Apache Spark, the platform establishes a high-throughput, scalable data pipeline. At its core, advanced natural language processing (NLP) models execute essential tasks including news categorization, named entity recognition (NER), event extraction, and automatic linkage to relevant stock symbols. The overarching objective is to deliver a transformative technological solution that enhances the efficiency, speed, and objectivity of information processing in the Vietnamese stock market. System-level evaluations demonstrate that the proposed architecture achieves high accuracy (>94
Large language models (LLMs) create new opportunities for scalable, personalized support in web-based education, but current systems often treat grading, feedback, tutoring, and deployment as separate problems. This PhD proposes a unified methodology for trustworthy, teacher-aligned adaptive learning loops on the web, integrating auditable grading automation, instructor steering, adaptive intervention, and scalable web deployment into one connected process. The research combines controlled classroom experiments, iterative web-architecture design, and production log analysis across enterprise and higher-education settings. Work completed in the first year includes one published and three submitted conference papers, classroom experiments with 132 students, and a production chatbot deployment, while the next phase targets cross-module transfer, broader higher-education rollout, and a systematic review.
