FOA Home
The fact keywords can be associated with documents, that keywords can be
associated with one another, documents become associated by citations,
etc. lead early IR pioneers such as Doyle and Stiles to adopt simple
association as a unifying relation connecting many objects of
the FOA inquiry [REF317] [Doyle61] [Doyle62] [REF322] [Maron82] [Giuliano63] [Findler79] . Integrating information
across SEMANTIC NETWORKS of such associative relations has been
an important example of knowledge representation within AI since the
memory models of Collins and Qullian [REF170] . In its simplest form, a simple
quantity known as ACTIVITY is allowed to propagate through a
network like that shown in Figure (figure) from several
sources. SPREADING ACTIVATION SEARCH is the name for a broad
range techniques that find solutions (for example, a path from Start to
Goal in Figure (FOAref) ) by controlling the propagation of
activity through associative networks like this [REF321] [REF668] [REF667] .
The Adaptive Information
Retrieval (AIR) system was a prototype search engine built as part of my
dissertation at the University of Michigan in the mid-1980s [REF222] [REF622] . This research was part one of
several systems applying CONNECTIONIST (neural network) learning
methods to the IR search engine problem [REF568] [REF732] [REF773] [REF430] .
Figure (figure) shows
how many of the features discussed here can interact as part of a single
retrieval system. This figure comes from Dan Rose's SCALIR (Symbolic and
Connectionist Approach to Legal Information Retrieval) system, built to
investigate the use of both logical, ``symbolic'' modes of inference and
probabalitic, ``subsymbolic'' ones. This figure shows containment
relations between document elements, (like those shown in more detail in
(FOAref) ) topical connections between keywords, and
inter-document citations, all mixed and used as part of spreading
activation-based inference.
Top of Page
Spreading activation search
Subsections