Section: New Results

Representing and Processing Imperfect Information

Participants : Abdallah Arioua, Jean-François Baget, Meghyn Bienvenu, Pierre Bisquert, Patrice Buche, Madalina Croitoru, Jérôme Fortin, Fabien Garreau, Abdelraouf Hecham, Marie-Laure Mugnier, Odile Papini, Swan Rocher, Rallou Thomopoulos, Bruno Yun.

Inconsistency-Tolerant Query Answering is one of the challenging problems that received a lot of attention in recent years as inconsistency may arise in practical applications due to several reasons: merging, integration, revision. In the context of Ontology-Based Data Access (OBDA), where the ontological knowledge is assumed to be coherent and fully reliable, inconsistency comes from the data, i.e., occurs when some assertional facts contradict some constraints imposed by the ontological knowledge. Existing works in this area have studied different inconsistency-tolerant query answering techniques, called “semantics”: some examples include Brave, IAR, ICR, AR etc..These proposals are closely related to works on querying inconsistent databases, or inference from inconsistent propositional logic knowledge bases.

The work of this year on inconsistency-tolerant query answering techniques for Ontology Based Data Access focused on (i) new results about different kinds of semantics or (ii) the user interaction with such semantics (we investigated the notion of repair based explanation or argumentation based explanation). We have investigated the interest of inconsistency-tolerant semantics in general and argumentation techniques for the agrifood chain in particular.

Inconsistency-Tolerant Semantics for Query Answering

In all approaches considered here, a knowledge base can have, in opposition to the logics studied in 6.1, several incompatible “minimal” models. Those models can correspond to possible repairs of an inconsistent knowledge base or can be the models generated by a non-monotonic logic. The questions we address here are linked to the semantics (how to define those models, how to define preferences on those models), while trying to preserve a satisfying trade-off between expressivity and computational complexity of the querying mechanism.

  • We proposed a new inconsistency-tolerant inference relation, called non-objection inference, where a query is considered as valid if it is entailed by at least one repair and it is consistent with all the other repairs. The main salient points of the newly introduced semantics is its efficiency (query answering with non-objection inference is achieved in polynomial time) and the fact that the inferences are strictly more productive than universal inference while preserving the consistency of its set of conclusions. The intuition behind is that no repair has an objection veto to the acceptance of the query. If query entailment from repairs is seen as posing a vote for, against or abstaining to a query then, in this semantics, some repairs are “voting” for a query (i.e., the query is entailed) and the rest of the repairs are not against (i.e., the query body atoms together with the atoms in the repair are consistent with the terminology) then the query is accepted without any objection. In addition, two variants of non-objection inference based on a selection of repairs (that can be against a query) are also considered.

  • We provided a dialectical characterization of the Brave and IAR semantics. We proposed an argumentation dialogue system that considers a turn taking game between a proponent and an opponent. We defined the concept of participant’s profile and depending on these profiles we were able to give necessary and sufficient conditions for the Brave and IAR semantics. We further proposed a new TPI-like dialectical proof theory (a procedure where two players exchange arguments (moves) until one of them cannot play) for universal acceptance (i.e.,AR semantics). We limit the scope of the work to finite and coherent logic-based argumentation frameworks that correspond to the OBDA instantiation we consider in practical applications.

  • We proposed a unifying framework for inconsistency-tolerant query answering within existential rule setting. In this framework, an inconsistency-tolerant semantics is seen as a pair composed of a modifier, which produces consistent subsets of the data, and an inference strategy, which evaluates queries on the selected subsets. We systematically compared the productivity and the complexity of the obtained semantics.

  • We studied the relationships between our unifying repair framework and stable model semantics. In particular, we provided a generic encoding for most semantics defined in that framework using Answer Set Programming.

Practical Applicability of Inconsistency-Tolerant Semantics and Argumentation

  • Several inconsistency-tolerant semantics have been introduced for querying inconsistent knowledge bases. In order for users to be able to understand the query results, it is crucial to be able to explain why a tuple is a (non-)answer to a query under such semantics. We defined explanations for positive and negative answers under the brave, AR and IAR semantics. We then studied the computational properties of explanations in the lightweight description logic DL-LiteR. For each type of explanation, we analyzed the data complexity of recognizing (preferred) explanations and deciding if a given assertion is relevant or necessary. We established tight connections between intractable explanation problems and variants of propositional satisfiability (SAT), enabling us to generate explanations by exploiting solvers for Boolean satisfaction and optimization problems. Finally, we empirically studied the efficiency of our explanation framework using the well-established LUBM benchmark.

  • We considered the problem of query-driven repairing of inconsistent DL-Lite knowledge bases: query answers are computed under inconsistency-tolerant semantics, and the user provides feedback about which answers are erroneous or missing. The aim is to find a set of data modifications (deletions and additions), called a repair plan, that addresses as many of the defects as possible. After formalizing this problem and introducing different notions of optimality, we investigated the computational complexity of reasoning about optimal repair plans and proposed interactive algorithms for computing such plans. For deletion-only repair plans, we also presented a prototype implementation of the core components of the algorithm.

  • Based on the equivalent use of inconsistency-tolerant semantics for OBDA and logical instantiation of argumentation with existential rules, we highlighted some of the practical advantages that come from the interplay of the two techniques. More generally, we focussed on the generic problem of dealing with the uncertain knowledge (elicitation, representation and reasoning) involved at different levels of the food chain that model complex processes relying on numerous criteria, using various granularity of knowledge, most often inconsistent (due to the fact that complementary points of view can be expressed).

    Beside, regarding the various granularity of knowledge, inspired from a hierarchical graph-based definition, we introduced the possibility of representing hierarchical knowledge using existential rules.

  • Agent technology and notably argumentation can optimise food supply chain operation in presence of inconsistency by employing intelligent agent applications (as shown in supply chain management case) but also facilitate reasoning with incomplete, inconsistent and missing knowledge as shown in the results presented in the previous sections. We considered two main methods of handling inconsistency: repair-based techniques and argumentation techniques. We demonstrated how to benefit from structured argumentation frameworks in practice by means of their implementations. Such implementations provide reasoning capabilities under inconsistency-tolerant semantics by means of a workflow that will enable Datalog frameworks to handle inconsistencies in knowledge bases using existing structured argumentation implementations.

  • We provided a first implementation of the explanation based techniques using argumentation that can be used for inconsistent tolerant semantics. Such implementation served as a proof of concept of the usefulness of the interplay of the two techniques.

    Furthermore, we provided an existential rule benchmark inspired from a real practical setting in the DURDUR project.

    To refine this approach, we presented a generic framework of capturing reasoning errors by the interplay of strict logical rules and associative rules in knowledge bases (with the latter being elicited using a game with a purpose). We showed that such model can capture certain reasoning biases and could be eventually used as a predictive model for interacting with domain experts. We also showed empirically the difference of associations agronomy experts exhibit with respect to a random control population validated in the context of the DURDUR ANR project.

Decision Support in Agronomy

  • We addressed a crucial problem for decision-making tools that are using inconsistency-handling methods (either argumentation frameworks or inconsistency-tolerant semantics) and namely the existence of multiple extensions / repairs. We placed ourselves in an applicative scenario, in the Pack4Fresh project, that investigates the best packaging for strawberries. We showed that being given a set of preferences on the initial set of facts in the existential rule knowledge base we can output meaningful (i.e., agrifood chain expert validated) extensions / repairs that will assist the decision maker.

  • We proposed a novel approach for decision-making that allows not only to handle symbolic data but also handle numerical RDF datasets. To deal with the numerical data, a preprocessing step is applied to convert numerical data into symbolic data. Based on the obtained symbolic classes we discover keys that are valid in this preprocessed data. We tested this approach on a dataset that describes wines with the set of numerical values representing different chemical components that give the flavour of wines. In this application setting, the discovered keys can be used to discover flavour complementarity, unknown from the experts, that allow to distinguish various wine sorts amongst themselves. We then validated the keys obtained with domain experts and discussed their interest with respect to the statistical analysis.

  • We presented a decision support system (DSS) which permits to compare, in a multi-criteria approach, innovative biomass transformation processes for biorefinery. Considered criteria are process extraction rate and green indicators. This DSS implements a pipeline which permits to annotate in a RDF knowledge heterogeneous textual data sources using a OWL/SKOS termino-ontological resource, to assess data source reliabilty and to compute several green indicators taking into account data reliability.