Adaptive Information Retrieval: Machine Learning in Associative Networks (Connectionist, Free-Text, Browsing, Feedback).

Abstract

One interesting issue in artificial intelligence (AI) currently is the relative merits of, and relationship between, the "symbolic" and "connectionist" approaches to intelligent systems building. The performance of more traditional symbolic systems has been striking, but getting these systems to learn truly new symbols has proven difficult. Recently, some researchers have begun to explore a distinctly different type of representation, similar in some respects to the nerve nets of several decades past. In these massively parallel, connectionist models, symbols arise implicitly, through the interactions of many simple and sub-symbolic elements. One of the advantages of using such simple elements as building blocks is that several learning algorithms work quite well. The range of application for connectionist models has remained limited, however, and it has been difficult to bridge the gap between this work and st and ard AI. The AIR system represents a connectionist approach to the problem of free-text information retrieval (IR). Not only is this an increasingly important type of data, but it provides an excellent demonstration of the advantages of connectionist mechanisms, particularly adaptive mechanisms. AIR's goal is to build an indexing structure that will retrieve documents that are likely to be found relevant. Over time, by using users' browsing patterns as an indication of approval, AIR comes to learn what the keywords (symbols) mean so as use them to retrieve appropriate documents. AIR thus attempts to bridge the gap between connectionist learning techniques and symbolic knowledge representations. The work described was done in two phases. The first phase concentrated on mapping the IR task into a connectionist network; it is shown that IR is very amenable to this representation. The second, more central phase of the research has shown that this network can also adapt. AIR translates the browsing behaviors of its users into a feedback signal used by a Hebbian-like local learning rule to change the weights on some links. Experience with a series of alternative learning rules are reported, and the results of experiments using human subjects to evaluate the results of AIR's learning are presented.