A spreading activation network model for information retrieval
Scott E. Preece · 1981
This report addresses a new model of bibliographic data and a corresponding processing paradigm for information retrieval. The data model is based on a view of the database as a network of nodes (documents and attributes) and links (occurrence of attributes in documents). The processing model is based on the notion of node activation spreading from already active nodes to those associated with them. It is shown that the new model can be used at reasonable cost with existing databases and can improve retrieval performance over conventional search methods. The cost for storage should be no more than 30 percent higher than for a conventional inverted file. Update costs should be no more than 10 percent higher than inverted file update costs. Runtime processing costs should be comparable to Boolean searching on a per operation basis. The model uses only the data contained in existing databases, without costly calculation of similarity coefficients and allows easy incorporation of additional attribute types and processing rules. The Spreading Activation Model can simulate the performance of many retrieval enhancing approaches previously reported in the literature. In addition we report the results of a small validation experiment performed on a pilot implementation of the model in which a significant number of items were retrieved that would have been missed had only direct searching been used. Associative retrieval improved performance on the test sample from 50 percent recall and precision to 65 percent recall and precision.