Inference networks use a slightly different probabilistic model based on reasoning under uncertainty [30]. An inference network is an inverted tree structure comprised of multiple parents culminating to a single child. The tree contains four levels, with each level consisting of a different type of node:
A sample network is shown in
Figure 1: Sample inference network for query
``(Richard and Nixon) and not Watergate''
Figure 1. Here, the nodes labeled 
and
correspond to three of the n Document nodes. The nodes
labeled ``Richard,'' ``Nixon,'' and ``Watergate'' are
Representation nodes, the ``and'' and ``not'' nodes are the Query nodes,
and the ``information need'' node is the I node.
The horizontal line separates the Document and Representation nodes,
which are created at index time rather than at query time, from the
Query and I nodes, which are created dynamically for each query.
Probabilities are calculated starting from the Document nodes and propagated down to the I node. Each child node contains a link matrix [51] that details how to combine the probabilities of the child's parents into the probabilities for the child. Eventually, all probabilities are propagated down to the I node. The I node thus contains the probabilities of each parent node satisfying the information need. It is a simple matter then to sort the probabilities and display the most relevant ones, in accordance with Robertson's Probability Ranking Principle [35].