Models of document indexing and document retrieval have been extensively studied. The integration of these two classes of models has been the goal of several researchers but it is a very difficult problem. We argue that much of the reason for this is the lack of an adequate indexing model. This suggests that perhaps a better indexing model would help solve the problem. However, we feel that making unwarranted parametric assumptions will not lead to better retrieval performance. Furthermore, making prior assumptions about the similarity of documents is not warranted either. Instead, we propose an approach to retrieval based on probabilistic language modeling. We estimate models for each document individually. Our approach to modeling is non-parametric and integrates document indexing and document retrieval into a single model. One advantage of our approach is that collection statistics which are used heuristically in many other retrieval models are an integral part of our model. We have...

this document length correction factor is #global": it is added at the end, after the weights for the individual terms have been summed, and is independentofwhich terms match.

Automatic information retrieval systems have to deal with documents of varying lengths in a text collection. Document length normalization is used to fairly retrieve documents of all lengths. In this study, we observe that a normalization scheme that retrieves documents of all lengths with similar chances as their likelihood of relevance will outperform another scheme which retrieves documents with chances very different from their likelihood of relevance. We show that the retrieval probabilities for a particular normalization method deviate systematically from the relevance probabilities across different collections. We present pivoted normalization, a technique that can be used to modify any normalization function thereby reducing the gap between the relevance and the retrieval probabilities. Training pivoted normalization on one collection, we can successfully use it on other (new) text collections, yielding a robust, collection independent normalization technique. We use the idea o...

The naive Bayes classifier, currently experiencing a renaissance in machine learning, has long been a core technique in information retrieval. We review some of the variations of naive Bayes models used for text retrieval and classification, focusing on the distributional assump- tions made about word occurrences in documents.

We explore the relationship between time and relevance using TREC ad-hoc queries. A type of query is identified that favors very recent documents. We propose a time-based language model approach to retrieval for these queries. We show how time can be incorporated into both query-likelihood models and relevance models. We carried out experiments to compare time-based language models to heuristic techniques for incorporating document recency in the ranking. Our results show that time-based models perform as well as or better than the best of the heuristic techniques.

The paper combines a comprehensive account of the probabilistic model of retrieval with new systematic experiments on TREC Programme material. It presents the model from its foundations through its logical development to cover more aspects of retrieval data and a wider range of system functions. Each step in the argument is matched by comparative retrieval tests, to provide a single coherent account of a major line of research. The experiments demonstrate, for a large test collection, that the probabilistic model is effective and robust, and that it responds appropriately, with major improvements in performance, to key features of retrieval situations.

A multi-database model of distributed information retrieval is presented, in which people are assumed to have access to many searchable text databases. In such an environment, full-text information retrieval consists of discovering database contents, ranking databases by their expected ability to satisfy the query, searching a small number of databases, and merging results returned by different databases. This paper presents algorithms for each task. It also discusses how to reorganize conventional test collections into multi-database testbeds, and evaluation methodologies for multi-database experiments. A broad and diverse group of experimental results is presented to demonstrate that the algorithms are effective, efficient, robust, and scalable. 1. INTRODUCTION Wide area networks, particularly the Internet, have transformed how people interact with information. Much of the routine information access by the general public is now based on full-text information retrieval, as opposed t...

An important class of searches on the world-wide-web has the goal to find an entry page (homepage) of an organisation. Entry page search is quite different from Ad Hoc search. Indeed a plain Ad Hoc system performs disappointingly. We explored three non-content features of web pages: page length, number of incoming links and URL form. Especially the URL form proved to be a good predictor. Using URL form priors we found over 70% of all entry pages at rank 1, and up to 89% in the top 10. Non-content features can easily be embedded in a language model framework as a prior probability.

This paper describes a simple way of adapting the BM25 ranking formula to deal with structured documents. In the past it has been common to compute scores for the individual fields (e.g. title and body) independently and then combine these scores (typically linearly) to arrive at a final score for the document. We highlight how this approach can lead to poor performance by breaking the carefully constructed non-linear saturation of term frequency in the BM25 function. We propose a much more intuitive alternative which weights term frequencies before the nonlinear term frequency saturation function is applied. In this scheme, a structured document with a title weight of two is mapped to an unstructured document with the title content repeated twice. This more verbose unstructured document is then ranked in the usual way. We demonstrate the advantages of this method with experiments on Reuters Vol1 and the TREC dotGov collection.

This paper describes the o cial runs of the Twenty-One group for TREC-7. The Twenty-One group participated in the ad-hoc and the cross-language track and made the following accomplishments: We developed a new weighting algorithm, which outperforms the popular Cornell version of BM25 on the ad-hoc collection. For the CLIR task we developed a fuzzy matching algorithm to recover from missing translations and spelling variants of proper names. Also for CLIR we investigated translation strategies that make extensive use of information from our dictionaries by identifying preferred translations, main translations and synonym translations, by de ning weights of possible translations and by experimenting with probabilistic boolean matching strategies.