Small language models (SLMs) have shown promises for relation extraction (RE) when extracting RDF triples guided by SHACL shapes focused on common Datatype Properties. This paper investigates how SLMs handle both Datatype and Object Properties for a complete RDF graph extraction. We show that the key bottleneck is related to long-tail distribution of rare properties. To solve this issue, we evaluate several strategies: stratified sampling, weighted loss, dataset scaling, and template-based synthetic data augmentation. We show that the best strategy to perform equally well over unbalanced target properties is to build a training set where the number of occurrences of each property exceeds a given threshold. To enable reproducibility, we publicly released our datasets, experimental results and code. Our findings offer practical guidance for training shape-aware SLMs and highlight promising directions for future work in semantic RE.

The SPARQL query language is the standard method to access knowledge graphs (KGs). However, formulating SPARQL queries is a significant challenge for non-expert users, and remains time-consuming for the experienced ones. Best practices recommend to document KGs with competency questions and example queries to contextualise the knowledge they contain and illustrate their potential applications. In practice, however, this is either not the case or the examples are provided in limited numbers. Large Language Models (LLMs) are being used in conversational agents and are proving to be an attractive solution with a wide range of applications, from simple question-answering about common knowledge to generating code in a targeted programming language. However, training and testing these models to produce high quality SPARQL queries from natural language questions requires substantial datasets of question-query pairs. In this paper, we present Q2Forge that addresses the challenge of generating new competency questions for a KG and corresponding SPARQL queries. It iteratively validates those queries with human feedback and LLM as a judge. Q2Forge is open source, generic, extensible and modular, meaning that the different modules of the application (CQ generation, query generation and query refinement) can be used separately, as an integrated pipeline, or replaced by alternative services. The result is a complete pipeline from competency question formulation to query evaluation, supporting the creation of reference question-query sets for any target KG.

In this paper we summarized the main historical steps in making the Web, its foundational principles and its evolution. First we mention some of the influences and streams of thought that interacted to bring the Web about. Then we recall that its birthplace, the CERN, had a need for a global hypertext system and at the same time was the perfect microcosm to provide a cradle for the Web. We stress how this invention required to strike a balance between the integration of and the departure from the existing and emerging paradigms of the day. We then review the pillars of the Web architecture and the features that made the Web so viral compared to competitors. Finally we survey the multiple mutations the Web underwent no sooner it was born, evolving in multiple directions. We conclude on the fact the Web is now an architecture, an artefact, a science object and a research and development object, and of which we haven’t seen the full potential yet.

Relational Graph Convolutional Networks (RGCNs) are commonly used on Knowledge Graphs (KGs) to perform black box link prediction. Several algorithms, or explanation methods, have been proposed to explain their predictions. Evaluating performance of explanation methods for link prediction is difficult without ground truth explanations. Furthermore, there can be multiple explanations for a given prediction in a KG. No dataset exists where observations have multiple ground truth explanations to compare against. Additionally, no standard scoring metrics exist to compare predicted explanations against multiple ground truth explanations. In this paper, we introduce a method, including a dataset (FrenchRoyalty-200k), to benchmark explanation methods on the task of link prediction on KGs, when there are multiple explanations to consider. We conduct a user experiment, where users score each possible ground truth explanation based on their understanding of the explanation. We propose the use of several scoring metrics, using relevance weights derived from user scores for each predicted explanation. Lastly, we benchmark this dataset on state-of-the-art explanation methods for link prediction using the proposed scoring metrics.

Electronic medical records (EMR) contain key information about the different symptomatic episodes that a patient went through. They carry a great potential in order to improve the well-being of patients and therefore represent a very valuable input for artificial intelligence approaches. However, the explicit knowledge directly available through these records remains limited, the extracted features to be used by machine learning algorithms do not contain all the implicit knowledge of medical expert. In order to evaluate the impact of domain knowledge when processing EMRs, we augment the features extracted from EMRs with ontological resources before turning them into vectors used by machine learning algorithms. We evaluate these augmentations with several machine learning algorithms to predict hospitalization. Our approach was experimented on data from the PRIMEGE PACA database that contains more than 350,000 consultations carried out by 16 general practitioners (GPs).

As the Web of Linked Open Data is growing the problem of crawling that cloud becomes increasingly important. Unlike normal Web crawlers, a Linked Data crawler performs a selection to focus on collecting linked RDF (including RDFa) data on the Web. From the perspectives of throughput and coverage, given a newly discovered and targeted URI, the key issue of Linked Data crawlers is to decide whether this URI is likely to dereference into an RDF data source and therefore it is worth downloading the representation it points to. Current solutions adopt heuristic rules to filter irrelevant URIs. Unfortunately, when the heuristics are too restrictive this hampers the coverage of crawling. In this paper, we propose and compare approaches to learn strategies for crawling Linked Data on the Web by predicting whether a newly discovered URI will lead to an RDF data source or not. We detail the features used in predicting the relevance and the methods we evaluated including a promising adaptation of FTRL-proximal online learning algorithm. We compare several options through extensive experiments including existing crawlers as baseline methods to evaluate their efficacy.

DBpedia is a huge dataset essentially extracted from the content and structure of Wikipedia. We present a new extraction producing a linked data representation of the editing history of Wikipedia pages. This supports custom querying and combining with other data providing new indicators and insights. We explain the architecture, representation and an immediate application to monitoring events.

Users in question-answer sites generate huge amounts of high quality and highly reusable information. This information can be categorized by topics but since users' interests change with time, uncovering the temporal patterns and trends in their activity is of prime interest to detect their current expertize. These temporal variations have long remained unexplored in question-answer sites while detecting them enables us to improve tasks such as: question routing, expert recommending and community life-cycle management. In this paper, we propose a generative model of such a community and its dynamics, and we perform experiments with real-world data extracted from the StackOverflow website to confirm the effectiveness of our model to study the users' behaviors and topics dynamics.

Axiom scoring is a critical task both for the automatic enrichment/learning and for the automatic validation of knowledge bases and ontologies. We designed and developed an axiom scoring heuristic based on possibility theory, which aims at overcoming some limitations of scoring heuristics based on statistical inference and taking into account the open-world assumption of the linked data on the Web. Since computing the possibilistic score can be computationally quite heavy for some candidate axioms, we propose a method based on time capping to alleviate the computation of the heuristic without giving up the precision of the scores. We evaluate our proposal by applying it to the problem of testing SubClassOf axioms against the DBpedia RDF dataset.

In many social networks, people interact based on their interests. Community detection algorithms are then useful to reveal the sub-structures of a network and help us find interest groups. Identifying these social communities can bring benefit to understanding and predicting users behaviors. However, for some kind of online community sites such as question-and-answer (Q&A) sites or forums, there is no friendship based social network structure, which means people are not aware who they are in contact with. Therefore, many traditional community detection techniques do not apply directly. In this paper, we propose an empirical approach for extracting data from Q&A sites suitable to apply community detection methods. Then we compare three kinds of community detection methods we applied on a dataset extracted from the popular Q&A site StackOverflow. We analyze and comment the results of each method.

In this paper the authors focus on context-aware adaptation for linked data on mobile. They split up the problem in two sub-questions: how to declaratively describe context at RDF presentation level, and how to overcome context imprecisions and incompleteness when selecting the proper context description at runtime. The authors answer their two-fold research question with PRISSMA, a context-aware presentation layer for Linked Data. PRISSMA extends the Fresnel vocabulary with the notion of mobile context. Besides, it includes an algorithm that determines whether the sensed context is compatible with some context declarations. The algorithm finds optimal error-tolerant subgraph isomorphisms between RDF graphs using the notion of graph edit distance and is sublinear in the number of context declarations in the system.

In the domain of Linked Open Data a need is emerging for developing automated frameworks able to generate the licensing terms associated to data coming from heterogeneous distributed sources. This paper proposes and evaluates a deontic logic semantics which allows us to define the deontic components of the licenses, i.e., permissions, obligations, and prohibitions, and generate a composite license compliant with the licensing items of the composed different licenses. Some heuristics are proposed to support the data publisher in choosing the licenses composition strategy which better suits her needs w.r.t. the data she is publishing. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Building on top of our results on semantic social network analysis, we present a community detection algorithm, SemTagP, that takes benefits of the semantic data that were captured while structuring the RDF graphs of social networks. SemTagP not only offers to detect but also to label communities by exploiting (in addition to the structure of the social graph) the tags used by people during the social tagging process as well as the semantic relations inferred between tags. Doing so, we are able to refine the partitioning of the social graph with semantic processing and to label the activity of detected communities. We tested and evaluated this algorithm on the social network built from Ph.D. theses funded by ADEME, the French Environment and Energy Management Agency. We showed how this approach allows us to detect and label communities of interest and control the precision of the labels.

The advent of tagging and folksonomies for organizing shared resources on the social Web brought promising opportunities to help communities of users capture their knowledge. However, the lack of semantics, or the spelling variations between tags lowers the potentials for browsing and exploring these data. To overcome these limitations, we propose exploiting the interactions between the users and the systems to validate or correct semantic analysis automatically applied to the tags. This process is based upon our model of the assistance of folksonomies enrichment which supports conflictual points of view. Several strategies can then be applied to propose novel browsing facilities to users.

We present our experiment in integrating semantic web services in the existing corporate semantic web server architecture we use to implement corporate memories. We rely on a semantic web search engine, to provide a semantic registry and automatically discover and invoke corporate applications wrapped into semantically annotated web services. Using rules we also demonstrate how to compose the web services with queries on the knowledge stored in the corporate memory to automatically populate the service inputs.

The semantic Web is a vision of a Web augmented with formalized knowledge annotating it. Currently, there is a huge gap between the conceptual structures underlying the semantic Web and the final rendering of a user-interface enabling an end-user to peruse or act on part of it. We describe an approach we experimented to automate part of the process of generating representations for concepts mobilized in the semantic Web. We reuse the notion of surrogate from information retrieval and we show that surrogate patterns tend to be close to the patterns of identity conditions used in ontology engineering. From this observation we propose and discuss a mechanism to derive surrogate templates from structures found in ontologies and rules.

Increasingly, application developers are looking for ways to provide users with higher levels of personalization that capture different elements of a user’s operating context, such as her location, the task that she is currently engaged in, who her colleagues are, etc. While there are many sources of contextual information, they tend to vary from one user to another and also over time. Different users may rely on different location tracking functionality provided by different cell phone operators; they may use different calendar systems, etc. In this paper, we describe work on a Semantic e-Wallet aimed at supporting automated discovery and access of personal resources, each represented as a Semantic Web Service. A key objective is to provide a Semantic Web environment for open access to a user’s contextual resources, thereby reducing the costs associated with the development and maintenance of context-aware applications. A second objective is, through Semantic Web technologies, to empower users to selectively control who has access to their contextual information and under which conditions. This work has been carried out in the context of myCampus, a context-aware environment aimed at enhancing everyday campus life. Empirical results obtained on Carnegie Mellon’s campus are encouraging. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.