Language models and smoothing methods for information retrieval

Designing an effective retrieval model that can rank documents accurately for a given query has been a central problem in information retrieval for several decades. An optimal retrieval model that is both effective and efficient and that can learn from feedback information over time is needed. Language models are new generation of retrieval models and have been applied since the last ten years to solve many different information retrieval problems. Compared with the traditional models such as the vector space model, they can be more easily adapted to model non traditional and complex retrieval problems and empirically they tend to achieve comparable or better performance than the traditional models. Developing new language models is currently an active research area in information retrieval. In the first stage of this thesis we present a new language model based on an odds formula, which explicitly incorporates document length as a parameter. To address the problem of data sparsity where there is rarely enough data to accurately estimate the parameters of a language model, smoothing gives a way to combine less specific, more accurate information with more specific, but noisier data. We introduce a new smoothing method called exponential smoothing, which can be combined with most language models. We present experimental results for various language models and smoothing methods on a collection with large document length variation, and show that our new methods compare favourably with the best approaches known so far. We discuss the collection effect on the retrieval function, where we investigate the performance of well known models and compare the results conducted using two variant collections. In the second stage we extend the current model from flat text retrieval to XML retrieval since there is a need for content-oriented XML retrieval systems that can efficiently and effectively store, search and retrieve information from XML document collections. Compared to traditional information retrieval, where whole documents are usually indexed and retrieved as single complete units, information retrieval from XML documents creates additional retrieval challenges. By exploiting the logical document structure, XML allows for more focussed retrieval that identifies elements rather than documents as answers to user queries. Finally we show how smoothing plays a role very similar to that of the idf function: beside the obvious role of smoothing, it also improves the accuracy of the estimated language model. The within document frequency and the collection frequency of a term actually influence the probability of relevance, which led us to a new class of smoothing function based on numeric prediction, which we call empirical smoothing. Its retrieval quality outperforms that of other smoothing methods.