Turtle and Croft advocate the use of inference networks in information retrieval systems. In their models, both document collections and queries are represented by networks; multiple representations can be handled. The authors compare their inference networks to probabilistic and Boolean models and show how networks can be used to simulate both of these models.

Experiments were conducted using two commonly used test collections: the CACM collection with 3204 documents and the CISI collection, published by the Compagnie Internationale de Services en Informatique (the International Information Services Company), with 1460 documents. The network model performed somewhat better than the probabilistic model and much better than the Boolean model. Combining results from different versions of the same queries gave improved performance. One important result is that the use of a nonzero default probability for term belief improves performance. Different results might have been obtained if different versions of the models had been implemented. Reasonable choices appear to have been made in all cases, however.

The paper is not easy to read and has few examples. It presents important results that are of interest to researchers in this area, however.