Using the structure of overlap between search results to rank retrieval systems without relevance judgments

This paper addresses the problem of how to rank retrieval systems without the need for human relevance judgments, which are very resource intensive to obtain. Using TREC 3, 6, 7 and 8 data, it is shown how the overlap structure between the search results of multiple systems can be used to infer relative performance differences. In particular, the overlap structures for random groupings of five systems are computed, so that each system is selected an equal number of times. It is shown that the average percentage of a system’s documents that are only found by it and no other systems is strongly and negatively correlated with its retrieval performance effectiveness, such as its mean average precision or precision at 1000. The presented method uses the degree of consensus or agreement a retrieval system can generate to infer its quality. This paper also addresses the question of how many documents in a ranked list need to be examined to be able to rank the systems. It is shown that the overlap structure of the top 50 documents can be used to rank the systems, often producing the best results. The presented method significantly improves upon previous attempts to rank retrieval systems without the need for human relevance judgments. This “structure of overlap” method can be of value to communities that need to identify the best experts or rank them, but do not have the resources to evaluate the experts’ recommendations, since it does not require knowledge about the domain being searched or the information being requested.     

Assigning appropriate weights for the linear combination data fusion method in information retrieval

In data fusion, the linear combination method is a very flexible method since different weights can be assigned to different systems. However, it remains an open question which weighting schema should be used. In some previous investigations and experiments, a simple weighting schema was used: for a system, its weight is assigned as its average performance over a group of training queries. However, it is not clear if this weighting schema is good or not. In some other investigations, different numerical optimisation methods were used to search for appropriate weights for the component systems. One major problem with those numerical optimisation methods is their low efficiency. It might not be feasible to use them in some situations, for example in some dynamic environments, system weights need to be updated from time to time for reasonably good performance. In this paper, we investigate the weighting issue by extensive experiments. The key point is to try to find the relation between performances of component systems and their corresponding weights which can lead to good fusion performance. We demonstrate that a series of power functions of average performance, which can be implemented as efficiently as the simple weighting schema, is more effective than the simple weighting schema for the linear data fusion method. Some other features of the power function weighting schema and the linear combination method are also investigated. The observations obtained from this study can be used directly in fusion applications of component retrieval results. The observations are also very useful for optimisation methods to choose better starting points and therefore to obtain more effective weights more quickly.