International audience

When the global pandemic struck in 2020, most countries established task forces to meet a challenge that impacted governmental resources. It became apparent that data, intelligence gathering, and both modelling and predictive capabilities were required. While artificial intelligence (AI) based solutions had already begun to emerge within the public sector, the Covid-19 pandemic accelerated this process. In particular, modelling of case numbers with the development of predictive algorithms. The development of AI solutions for public sector organizations is inherently multidisciplinary. This is crucial to understanding how solutions can be developed, outputs understood, and the benefits and risks measured. Furthermore, the development of AI solutions often requires data which may not be accessible from a single location. In the case of Covid-19 modelling, data must be extracted from multiple locations to construct data assets. In this research, a collaborative approach to developing machine learning expertise for the public sector is presented. Using Covid-19 as a case study, the role of different government sectors when building data assets is examined along with the use of standard data models, and how this type of cooperation led to the development of a pipeline for data assets to underpin AI solutions for the public sector.

In sufficiently complex tasks, it is expected that as a side effect of learning to solve a problem, a neural network will learn relevant abstractions of the representation of that problem. This has been confirmed in particular in machine vision where a number of works showed that correlations could be found between the activations of specific units (neurons) in a neural network and the visual concepts (textures, colors, objects) present in the image. Here, we explore the use of self-organizing maps as a way to both visually and computationally inspect how activation vectors of whole layers of neural networks correspond to neural representations of abstract concepts such as ‘female person’ or ‘realist painter’. We experiment with multiple measures applied to those maps to assess the level of representation of a concept in a network’s layer. We show that, among the measures tested, the relative entropy of the activation map for a concept compared to the map for the whole data is a suitable candidate and can be used as part of a methodology to identify and locate the neural representation of a concept, visualize it, and understand its importance in solving the prediction task at hand.

Knowledge in mathematics (definitions, theorems, proofs, etc.) is usually expressed in a way that combines natural language and mathematical expressions (e.g. equations). Using an ontology formalism such as OWL DL is well-suited for formalizing the natural language part, but complex mathematical expressions can be better handled by symbolic computation systems. We examine this representation issue and propose an original extension of OWL DL by call formulas, i.e., formulas from which assertions can be drawn thanks to calls to external functions. Using this formalism makes it possible to classify a mathematical problem defined by its relations to instances and classes and by some mathematical expressions: if a theorem for solving this problem is represented in the knowledge base, it can be retrieved, and thus, the problem can be solved by applying this theorem. We describe an inference algorithm and discuss its properties as well as its limitations. Indeed, the proposed extension, algorithm, and implementation represent a first step towards a combined formalism for representing mathematical knowledge, with some open issues regarding the representation of more complex problems: the resolution of multiscale, multiphysics cases in physics are foreseen.

While neural networks models have shown impressive performance in many NLP tasks, lack of interpretability is often seen as a disadvantage. Individual relevance scores assigned by post-hoc explanation methods are not sufficient to show deeper systematic preferences and potential biases of the model that apply consistently across examples. In this paper we apply rule mining using knowledge graphs in combination with neural network explanation methods to uncover such systematic preferences of trained neural models and capture them in the form of conjunctive rules. We test our approach in the context of text classification tasks and show that such rules are able to explain a substantial part of the model behaviour as well as indicate potential causes of misclassifications when the model is applied outside of the initial training context.

One of the existing query recommendation strategies for unknown datasets is "by example", i.e. based on a query that the user already knows how to formulate on another dataset within a similar domain. In this paper we measure what contribution a structural analysis of the query and the datasets can bring to a recommendation strategy, to go alongside approaches that provide a semantic analysis. Here we concentrate on the case of star-shaped SPARQL queries over RDF datasets. The illustrated strategy performs a least general generalization on the given query, computes the specializations of it that are satisfiable by the target dataset, and organizes them into a graph. It then visits the graph to recommend first the reformulated queries that reflect the original query as closely as possible. This approach does not rely upon a semantic mapping between the two datasets. An implementation as part of the SQUIRE query recommendation library is discussed.

When publishing data, data licences are used to specify the actions that are permitted or prohibited, and the duties that target data consumers must comply with. However, in complex environments such as a smart city data portal, multiple data sources are constantly being combined, processed and redistributed. In such a scenario, deciding which policies apply to the output of a process based on the licences attached to its input data is a difficult, knowledge-intensive task. In this paper, we evaluate how automatic reasoning upon semantic representations of policies and of data flows could support decision making on policy propagation. We report on the results of a user study designed to assess both the accuracy and the utility of such a policy-propagation tool, in comparison to a manual approach.

The focus of this work is to exploit ontologies to make robotic systems more accessible to non-expert users, therefore supporting the deployment of robot-integrated applications. Due to the increasing number of robotic platforms available for commercial use, robotic systems are nowadays being approached by users with different backgrounds, who are often more interested in the robots' high-level capabilities than their technical architecture. Without the right expertise however, using robots is restricted to the capabilities exposed by the platform provider, i.e. they can only be used as end products rather than as development platforms. Our hypothesis is that an ontological representation of the capabilities of robots could make these capabilities more accessible, reducing the complexity of robot programming and enabling non-experts to exploit these systems to a much larger extent. To demonstrate this, an ontology abstracting the capabilities exposed by the most common robotic middleware (ROS) is integrated in a system to allow non-experts to program robots of different types and capabilities without previous knowledge either of the specific robotic platform being considered, or of the intricate systems used in its implementation. Our experiments, in which non-experts users had to configure the system in order to make robots achieve different tasks, show how the efforts required for realizing basic tasks using available robotic platforms can be sensibly reduced through our approach.

Governing the life cycle of data on the web is a challenging issue for organisations and users. Data is distributed under certain policies that determine what actions are allowed and in which circumstances. Assessing what policies propagate to the output of a process is one crucial problem. Having a description of policies and data flow steps implies a huge number of propagation rules to be specified and computed (number of policies times number of actions). In this paper we provide a method to obtain an abstraction that allows to reduce the number of rules significantly. We use the Datanode ontology, a hierarchical organisation of the possible relations between data objects, to compact the knowledge base to a set of more abstract rules. After giving a definition of Policy Propagation Rule, we show (1) a methodology to abstract policy propagation rules based on an ontology, (2) how effective this methodology is when using the Datanode ontology, (3) how this ontology can evolve in order to better represent the behaviour of policy propagation rules.

In this paper, we propose an ontology design pattern for the concept of "explanation". The motivation behind this work comes from our research, which focuses on automatically identifying explanations for data patterns. If we want to produce explanations from data agnostically from the application domain, we first need a formal definition of what an explanation is, i.e. which are its components, their roles or their interactions. We analysed and surveyed works from the disciplines grouped under the name of Cognitive Sciences, with the aim of identifying differences and commonalities in the way their researchers intend the concept of explanation. We then produced not only an ontology design pattern to model it, but also the instantiations of this in each of the analysed disciplines. Besides those contributions, the paper presents how the proposed ontology design pattern can be used to analyse the validity of the explanations produced by our, and other, frameworks.

With the rise of linked data, more and more semantically described information is being published online according to the principles and technologies of the Semantic Web (especially, RDF and SPARQL). The use of such standard technologies means that this data should be exploitable, integrable and reusable straight away. However, once a potentially interesting dataset has been discovered, significant efforts are currently required in order to understand its schema, its content, the way to query it and what it can answer. In this paper, we propose a method and a tool to automatically discover questions that can be answered by an RDF dataset. We use formal concept analysis to build a hierarchy of meaningful sets of entities from a dataset. These sets of entities represent answers, which common characteristics represent the clauses of the corresponding questions. This hierarchy can then be used as a querying interface, proposing questions of varying levels of granularity and specificity to the user. A major issue is however that thousands of questions can be included in this hierarchy. Based on an empirical analysis and using metrics inspired both from formal concept analysis and from ontology summarization, we devise an approach for identifying relevant questions to act as a starting point to the navigation in the question hierarchy.

Ontologies are conceptual models of particular domains, and domains can be modeled differently, representing different opinions, beliefs or perspectives. In other terms, ontologies may disagree with some particular pieces of information and among themselves. Assessing such agreements and disagreements is very useful in a variety of scenarios, in particular when integrating external elements of information into existing ones. In this paper, we present a set of measures to evaluate the agreement and disagreement of an ontology with a statement or with other ontologies. While our work goes beyond the naive approach of checking for logical inconsistencies, it relies on a complete formal framework based on the semantics of the considered ontologies. The experiments realized on several concrete scenarios show the validity of our approach and the usefulness of measuring agreement and disagreement in ontologies.

This demo presents the Cupboard online system for sharing and reusing ontologies linked together with alignments, and that are attached to rich metadata and reviews.

Nowadays, the increasing amount of semantic data available on the Web leads to a new stage in the potential of Semantic Web applications. However, it also introduces new issues due to the heterogeneity of the available semantic resources. One of the most remarkable is redundancy, that is, the excess of different semantic descriptions, coming from different sources, to describe the same intended meaning.