Automated evaluation of hypertext search strategies

This thesis investigates the potential of using search agents to analyze the performance of interactive information retrieval systems. An evaluation framework uses idealized and simulated individual differences in hypertext search strategies to simulate differences in human search performance and link selection behaviour.; Five distinct search strategies were identified and implemented in software. The software simulator allowed the search agents to interact with an information retrieval system, which employed dynamic hypertext as its interactive search mechanism.; Three studies investigated the performance and behaviour of the search agents. The first study compared three embedded hypertext link-ranking strategies (which used textual similarity assessment algorithms) with corresponding human link rankings. Although the three algorithms studied had low overall correlation with the human participants, the analysis revealed that the algorithms performed better (relative to human judgments) under specific conditions (e.g., short vs. long documents) and were more highly correlated with the participants under those conditions.; The second study compared performance when using different combinations of settings for the parameters that were relevant to each agent. Based on the results of this study, the best performing internal parameters for each agent were chosen and then used in a third study.; The third study compared the performance of the agents with two different experimental factors, one representing variations in query tail size (i.e., how much of the prior search history/link selections were reflected in the current automatically generated query), and in newness (i.e., the extent to which agents were permitted to return to previously viewed documents). Both query tail size and newness affected the performance of the agents (precision decreased when newness increased and precision increased when query tail increased).; This dissertation describes a framework for evaluating the effect of different search strategies and a search agent simulator based on the framework. The results of the simulation studies demonstrated a difference in terms of performance and link selection behaviour between the search agents. Effects of query difficulty and novelty on agent performance/behaviour were also identified. The simulator developed in this research should provide a useful platform for future studies of idealized search behaviour in interactive information retrieval.