For multimedia interpretation, a semantically well-founded formalization is required. In accordance
with previous work, in CASAM a well-founded abduction-based approach is pursued. Extending
previous work, abduction is controlled by probabilistic knowledge, and it is done in terms of firstorder
logic.
This report describes the probabilistic abduction engine and the optimization techniques for
multimedia interpretation. It extends deliverable D3.2 by providing a probabilistic scoring function
for ranking interpretation alternatives. Parameters for the CASAM Abduction Engine (CAE)
introduced already in D3.2 are now appropriately formalized such that CAE is better integrated
into the probabilistic framework. In addition, this deliverable describes how media interpretation
services can be provided that work incrementally, i.e., are able to consume new analysis results,
or new input from a human annotator, and produce notifications for additional interpretation
results or, in some cases, revision descriptions for previous interpretations. Incremental processing
is nontrivial and is realized using an Abox dierence operator, which is used to interpretation
results obtained for extended inputs with one(s) previously obtained such that notifications about
additions and revisions can be computed.