WP3: Multi-Perspective Knowledge Representation

Task Lead: FBK, PI: Chiara Ghidini: Co-PI: Marco Roveri

We will research formalisms and algorithms to integrate the multiple perspectives of a real-world domain into holistic formal representations, and to discover, align, co-evolve, and exploit implicit and explicit knowledge. This will empower AI systems with the ability to build, make explicit, use, and share formal representations of the world. 

Effort 36p/m + cascading funds

Description of work 

This WP addresses new challenges in representation, discovery, alignment, and exploitation of multi-perspective knowledge.

Task 2.3.1 –  Multi-perspective Knowledge Representation and Discovery - Task Leader: Chiara Ghidini

Provide novel Knowledge Representation (KR) frameworks able to provide integrated representations of multiperspective domains  having a strong temporal flavor and a number of data-oriented structural constraints. Provide techniques and tools to extract and represent in the identified holistic formal framework multi-perspective knowledge contained in the data, with particular emphasis to events, (temporal) narratives, dynamic knowledge, and probabilities, so that the rules characterizing a specific portion of the word emerge. The new holistic formal framework will combine qualitative, quantitative, temporal and causal reasoning with data to capture the underlying (possibly hidden) relations. It will build on and extend existing symbolic frameworks such as DL, LTL/LTLf, ASP, (PO)MDP  to represent different aspects of knowledge in a formal manner, also keeping into account temporal, dynamic, uncertainty and probabilistic aspects, and will combine these formalisms with the theory of the data to deal with the data-oriented aspects (all essential to faithfully capture the inherent complexity of complex domains, their dynamics, and mutual interactions, as well as to incorporate flexibility, adaptiveness, openness, and uncertainty). The extraction techniques and tools will build on and extend existing techniques and tools with i) novel semantic data mapping and integration mechanisms; ii) integrated mechanisms to leverage ontological modeling, reasoning techniques such as ILP, ASP, SMT,  as well as classical Machine Learning techniques.

Task 2.3.2 –  Multi-perspective Alignment and Co-evolution Task Leader: Marco Roveri

Provide novel techniques and tools to address conformance checking between the formal models and their data instantiations, going beyond a simple “yes/no” answer, and actually exploit debugging information to compute and highlight deviations and alignments (e.g., ways to map the observed behaviors in the data to the “closest” behaviors prescribed by the model).  The novel contributions will concern i) the alignment techniques, with a special focus on conformance checking between data-aware process/temporal models and temporal event data; ii) algorithms and techniques able to identify and characterize/explain, in a semantically meaningful (why) manner, the reasons for (mis-)alignments in their different flavors (genuine nonconformance, deviations, exceptions, faulty data, and concept drifts, among others).  The developed techniques will provide i) the basis for investigating how to fix and compensate deviations in models and in data; ii) and the investigation of indications for an aligned co-evolution.  The approaches could leverage emerging qualitative reasoning modulo theory approaches that aim at integrating qualitative formalisms (e.g., DL/LTL/LTLf) with quantitative information (e.g. constraints on data), as well as probabilistic reasoning (e.g. probabilistic variants of the previously mentioned logics, or their integration with probabilistic systems such as (PO)MDPs).

Task 2.3.3 –  Exploiting  Multi-perspective Knowledge Task Leader: Chiara Ghidini

This task aims at developing novel techniques to build predictive and proactive explainable systems able i) to identify and exploit features/objectives that are important in a decision; and ii) to tell the human decision-makers why the algorithm considers those features as important. Such results will be, for instance, achieved by leveraging specific sources of knowledge, e.g., conceptual definitions, semantics, and data knowledge (from knowledge graphs to ontologies) while designing the predictive and proactive systems. Such techniques will answer the need to develop integrative approaches that increasingly leverage the power and flexibility of Machine Learning techniques, combined with the explicit (and often binding) rules governing the scenario, but from a different perspective, w.r.t integrative representational frameworks addressed in other WPs. To this extent, we may leverage the developed frameworks in T2.3.1 to integrate domain knowledge on one side and logic-based approaches (e.g., LTLf proofs of satisfiability or validity, possibly extended to deal with multiple perspectives at once) to extract and show explanations of the importance and rationale behind the chosen features, on the other side.