On combining probabilistic and semantic similarity-based methods toward off-domain reasoning for situational awareness

Hard and soft high-level fusion plays an important role in the situational awareness literature. To deal with complex real-world situations, it is highly desirable that such systems are able to effectively capture the rich semantics associated with soft information and world/domain knowledge, to efficiently reason with imperfect information, and to benefit and learn from any data that may be available, among others. In these respects, first-order probabilistic models, such as Markov Logic Networks, hold great promise and have received recent attention for high-level fusion. By combining the expressiveness of first-order logic and probabilistic graphical models, such models are able to facilitate representation, reasoning and learning with complex relational information and rich probabilistic structure within a unifying framework. However, first-order probabilistic models may face various challenges in dealing with real-world situational awareness, including scalability of reasoning, learning and knowledge base construction, and robustness in open worlds. In this paper, we motivate a new and pragmatic approach toward collectively addressing these concerns; that is, endowing high-level fusion systems a capability to perform off-domain reasoning, through the ability to reason about unknown/unmodelled concepts. In particular, we will discuss how such an approach could be achieved by means of combining probabilistic and semantic similarity-based methods. We will also explore the potential contribution of semantic similarity measures derived from both taxonomic knowledge (e.g., ontologies) and distributional semantic models (generated from text corpora) toward achieving this goal.