Using an open-source, Java toolkit of name-matching methods, we experimentally compare string distance metrics on the task of matching entity names. We investigate a number of different metrics proposed by different communities, including edit-distance metrics, fast heuristic string comparators, token-based distance metrics, and hybrid methods. Overall, the best-performing method is a hybrid scheme combining a TFIDF weighting scheme, which is widely used in information retrieval, with the Jaro-Winkler string-distance scheme, which was developed in the probabilistic record linkage community.

The optimal settings of retrieval parameters often depend on both the document collection and the query, and are usually found through empirical tuning. In this paper, we propose a family of two-stage language models for information retrieval that explicitly captures the different influences of the query and document collection on the optimal settings of retrieval parameters. As a special case, we present a two-stage smoothing method that allows us to estimate the smoothing parameters completely automatically. In the first stage, the document language model is smoothed using a Dirichlet prior with the collection language model as the reference model. In the second stage, the smoothed document language model is further interpolated with a query background language model. We propose a leave-one-out method for estimating the Dirichlet parameter of the first stage, and the use of document mixture models for estimating the interpolation parameter of the second stage. Evaluation on five different databases and four types of queries indicates that the twostage smoothing method with the proposed parameter estimation methods consistently gives retrieval performance that is close to— or better than—the best results achieved using a single smoothing method and exhaustive parameter search on the test data.

The language modeling approach to retrieval has been shown to perform well empirically. One advantage of this new approach is its statistical foundations. However, feedback, as one important component in a retrieval system, has only been dealt with heuristically in this new retrieval approach: the original query is usually literally expanded by adding ditional terms to it. Such expansion-based feedback creates an inconsistent interpretation of the original and the expanded query. In this paper, we present a more principled approach to feedback in the language modeling approach. Specifically, we treat feedback as updating the query language model based on the extra evidence carried by the feedback documents. Such a model-based feedback strategy easily fits into an extension of the language modeling approach. We propose and evaluate two different approaches to updating a query language model based on feedback documents, one based on a generarive probabilistic model of feedback documents and one based on minimization of the KL-divergence over feedback documents. Experiment resuits show that both approaches are effective and outperform the Rocchio feedback approach.

We present a non-traditional retrieval problem we call subtopic retrieval. The subtopic retrieval problem is concerned with finding documents that cover many different subtopics of a query topic. This means that the utility of a document in a ranking is dependent on other documents in the ranking, violating the assumption of independent relevance which is assumed in most traditional retrieval methods. Subtopic retrieval poses challenges for evaluating performance, as well as for developing effective algorithms. We propose a framework for evaluating subtopic retrieval which generalizes the traditional precision and recall metrics by accounting for intrinsic topic difficulty as well as redundancy in documents. We propose and systematically evaluate several methods for performing subtopic retrieval using statistical language models and a maximal marginal relevance (MMR) ranking strategy. A mixture model combined with query likelihood relevance ranking is shown to modestly outperform a baseline relevance ranking on a data set used in the TREC interactive track.

This paper addresses the problem of extending an adaptive information filtering system to make decisions about the novelty and redundancy of relevant documents. It argues that relevance and redundance should each be modelled explicitly and separately. A set of five redundancy measures are proposed and evaluated in experiments with and without redundancy thresholds. The experimental results demonstrate that the cosine similarity metric and a redundancy measure based on a mixture of language models are both effective for identifying redundant documents.

This paper is concerned with learning to rank for information retrieval (IR). Ranking is the central problem for information retrieval, and employing machine learning techniques to learn the ranking function is viewed as a promising approach to IR. Unfortunately, there was no benchmark dataset that could be used in comparison of existing learning algorithms and in evaluation of newly proposed algorithms, which stood in the way of the related research. To deal with the problem, we have constructed a benchmark dataset referred to as LETOR and distributed it to the research communities. Specifically we have derived the LETOR data from the existing data sets widely used in IR, namely, OHSUMED and TREC data. The two collections contain queries, the contents of the retrieved documents, and human judgments on the relevance of the documents with respect to the queries. We have extracted features from the datasets, including both conventional features, such as term frequency, inverse document frequency, BM25, and language models for IR, and features proposed recently at SIGIR, such as HostRank, feature propagation, and topical PageRank. We have then packaged LETOR with the extracted features, queries, and relevance judgments. We have also provided the results of several state-ofthe-arts learning to rank algorithms on the data. This paper describes in details about LETOR.

This paper investigates the pre-conditions for successful combination of document representations formed from structural markup for the task of known-item search. As this task is very similar to work in meta-search and data fusion, we adapt several hypotheses from those research areas and investigate them in this context. To investigate these hypotheses, we present a mixturebased language model and also examine many of the current metasearch algorithms. We find that compatible output from systems is important for successful combination of document representations. We also demonstrate that combining low performing document representations can improve performance, but not consistently. We find that the techniques best suited for this task are robust to the inclusion of poorly performing document representations. We also explore the role of variance of results across systems and its impact on the performance of fusion, with the surprising result that the correct documents have higher variance across document representations than highly ranking incorrect documents.

Discriminative models have been preferred over generative models in many machine learning problems in the recent past owing to some of their attractive theoretical properties. In this paper, we explore the applicability of discriminative classifiers for IR. We have compared the performance of two popular discriminative models, namely the maximum entropy model and support vector machines with that of language modeling, the state-of-the-art generative model for IR. Our experiments on ad-hoc retrieval indicate that although maximum entropy is significantly worse than language models, support vector machines are on par with language models. We argue that the main reason to prefer SVMs over language models is their ability to learn arbitrary features automatically as demonstrated by our experiments on the home-page finding task of TREC-10.

Inspired by the PageRank and HITS (hubs and authorities) algorithms for Web search, we propose a structural re-ranking approach to ad hoc information retrieval: we reorder the documents in an initially retrieved set by exploiting asymmetric relationships between them. Specifically, we consider generation links, which indicate that the language model induced from one document assigns high probability to the text of another; in doing so, we take care to prevent bias against long documents. We study a number of re-ranking criteria based on measures of centrality in the graphs formed by generation links, and show that integrating centrality into standard language-model-based retrieval is quite effective at improving precision at top ranks.

A family of kernels for statistical learning is introduced that exploits the geometric structure of statistical models. The kernels are based on the heat equation on the Riemannian manifold defined by the Fisher information metric associated with a statistical family, and generalize the Gaussian kernel of Euclidean space. As an important special case, kernels based on the geometry of multinomial families are derived, leading to kernel-based learning algorithms that apply naturally to discrete data. Bounds on covering numbers and Rademacher averages for the kernels are proved using bounds on the eigenvalues of the Laplacian on Riemannian manifolds. Experimental results are presented for document classification, for which the use of multinomial geometry is natural and well motivated, and improvements are obtained over the standard use of Gaussian or linear kernels, which have been the standard for text classification.