3.7.1: Inference Networks

An inference network consists of

To apply the model in IR Turtle and Croft (1990 [->], 1991 [->]) use a network consisting of two parts. The "document network" consists of three layers:

Edges are only leading from one layer to the next one. The document network is determined by the documents of the collection. The layers represent various levels of abstraction. All paths are starting at the document nodes and lead to the concept nodes.

The second part of the inference network is the query network. It is connected to the layer of concept representations and may consist of several layers detailing various levels of abstraction of a query. All paths in this part of the graph lead to a single node, that represents the importance of a document for the specified query. To obtain this value the node of the document is activated and this activation is propagated through the network until it reaches the last node of the query network. During this propagation several paths of "evidence" contribute to the final value.

The actual implementation in the IR system INQUERY is much simpler. It consists of only one layer representing terms, i. e. a vector space model with a sophisticated similarity measure implemented by an inverted list.


© 1998 / HTML-Version 17. 11. 1998: R. Ferber