Learning to select for information retrieval

The effective ranking of documents in search engines is based on various document features, such as the frequency of the query terms in each document, the length, or the authoritativeness of each document. In order to obtain a better retrieval performance, instead of using a single or a few features, there is a growing trend to create a ranking function by applying a learning to rank technique on a large set of features. Learning to rank techniques aim to generate an effective document ranking function by combining a large number of document features. Different ranking functions can be generated by using different learning to rank techniques or on different document feature sets. While the generated ranking function may be uniformly applied to all queries, several studies have shown that different ranking functions favour different queries, and that the retrieval performance can be significantly enhanced if an appropriate ranking function is selected for each individual query. This thesis proposes Learning to Select (LTS), a novel framework that selectively applies an appropriate ranking function on a per-query basis, regardless of the given query’s type and the number of candidate ranking functions. In the learning to select framework, the effectiveness of a ranking function for an unseen query is estimated from the available neighbouring training queries. The proposed framework employs a classification technique (e.g. k-nearest neighbour) to identify neighbouring training queries for an unseen query by using a query feature. In particular, a divergence measure (e.g. Jensen-Shannon), which determines the extent to which a document ranking function alters the scores of an initial ranking of documents for a given query, is proposed for use as a query feature. The ranking function which performs the best on the identified training query set is then chosen for the unseen query. The proposed framework is thoroughly evaluated on two different TREC retrieval tasks (namely, Web search and adhoc search tasks) and on two large standard LETOR feature sets, which contain as many as 64 document features, deriving conclusions concerning the key components of LTS, namely the query feature and the identification of neighbouring queries components. Two different types of experiments are conducted. The first one is to select an appropriate ranking function from a number of candidate ranking functions. The second one is to select multiple appropriate document features from a number of candidate document features, for building a ranking function. Experimental results show that our proposed LTS framework is effective in both selecting an appropriate ranking function and selecting multiple appropriate document features, on a per-query basis. In addition, the retrieval performance is further enhanced when increasing the number of candidates, suggesting the robustness of the learning to select framework. This thesis also demonstrates how the LTS framework can be deployed to other search applications. These applications include the selective integration of a query independent feature into a document weighting scheme (e.g. BM25), the selective estimation of the relative importance of different query aspects in a search diversification task (the goal of the task is to retrieve a ranked list of documents that provides a maximum coverage for a given query, while avoiding excessive redundancy), and the selective application of an appropriate resource for expanding and enriching a given query for document search within an enterprise. The effectiveness of the LTS framework is observed across these search applications, and on different collections, including a large scale Web collection that contains over 50 million