Resonance as a neuronal mechanism is very cleverly linked to information retrieval (IR) in this interesting and well-written paper. The approach in the paper can be described as a “natural” interpretation of relevance in IR, through a theory (adaptive resonance theory (ART)) that models natural human cognitive processes.
In the approach described in the paper, relevant information, searched via query, is retrieved according to the level of resonance between the contents of a document and the query. The query is mapped to an internal state of the upper layer of a two-layered ART neural network, where the lower layer corresponds to the document under evaluation, namely, the input. Since the nodes of the two layers are connected through weighted directional links, when the incoming “stimulus” is similar to an internal state (that exists in the very values of the trained weights), strong signals back-propagate to the lower layer, this is repeated, and the system locks into a resonant state. If the document has no likeness to the query, the system does not lock into a resonant state because the back-propagated signals spread unevenly in the lower layer, blurring the picture even more.
How the query is initially mapped to the weights, and how these can be intelligently updated (re-trained) based on the successes and failures of the system according to user’s judgment of relevance, are also discussed in the paper. The paper presents some experimental data that indicates that the approach described can be effective in real-world IR projects.