Collaborative filtering is concerned with making recommendations about items to users. Most formulations of the problem are specifically designed for predicting user ratings, assuming past data of explicit user ratings is available. However, in practice we may only have implicit evidence of user preference; and furthermore, a better view of the task is of generating a top-N list of items that the user is most likely to like. In this regard, we argue that collaborative filtering can be directly cast as a relevance ranking problem. We begin with the classic Probability Ranking Principle of information retrieval, proposing a probabilistic item ranking framework. In the framework, we derive two different ranking models, showing that despite their common origin, different factorizations reflect two distinctive ways to approach item ranking. For the model estimations, we limit our discussions to implicit user preference data, and adopt an approximation method introduced in the classic text retrieval model (i.e. the Okapi BM25 formula) to effectively decouple frequency counts and presence/absence counts in the preference data. Furthermore, we extend the basic formula by proposing the Bayesian inference to estimate the probability of relevance (and non-relevance), which largely alleviates the data sparsity problem. Apart from a theoretical contribution, our experiments on real data sets demonstrate that the proposed methods perform significantly better than other strong baselines.

Over the last three decades, research in Information Retrieval (IR) shows performance improvement when many sources of evidence are combined to produce a ranking of documents. Most current approaches assess document relevance by computing a single score which aggregates values of some attributes or criteria. They use analytic aggregation operators which either lead to a loss of valuable information, e.g., the min or lexicographic operators, or allow very bad scores on some criteria to be compensated with good ones, e.g., the weighted sum operator. Moreover, all these approaches do not handle imprecision of criterion scores. In this paper, we propose a multiple criteria framework using a new aggregation mechanism based on decision rules identifying positive and negative reasons for judging whether a document should get a better ranking than another. The resulting procedure also handles imprecision in criteria design. Experimental results are reported showing that the suggested method performs better than standard aggregation operators.

Smoothing of document language models is critical in language modeling approaches to information retrieval. In this paper, we present a novel way of smoothing document language models based on propagating term counts probabilistically in a graph of documents. A key difference between our approach and previous approaches is that our smoothing algorithm can iteratively propagate counts and achieve smoothing with remotely related documents. Evaluation results on several TREC data sets show that the proposed method significantly outperforms the simple collection-based smoothing method. Compared with those other smoothing methods that also exploit local corpus structures, our method is especially effective in improving precision in top-ranked documents through "filling in" missing query terms in relevant documents, which is attractive since most users only pay attention to the top-ranked documents in search engine applications.

This paper reports on the underlying IR problems encountered when indexing and searching with the Bulgarian language. For this language we propose a general light stemmer and demonstrate that it can be quite effective, producing significantly better MAP (around + 34%) than an approach not applying stemming. We implement the GL2 model derived from the Divergence from Randomness paradigm and find its retrieval effectiveness better than other probabilistic, vector-space and language models. The resulting MAP is found to be about 50% better than the classical tf idf approach. Moreover, increasing the query size enhances the MAP by around 10% (from T to TD). In order to compare the retrieval effectiveness of our suggested stopword list and the light stemmer developed for the Bulgarian language, we conduct a set of experiments on another stopword list and also a more complex and aggressive stemmer. Results tend to indicate that there is no statistically significant difference between these variants and our suggested approach. This paper evaluates other indexing strategies such as 4-gram indexing and indexing based on the automatic decompounding of compound words. Finally, we analyze certain queries to discover why we obtained poor results, when indexing Bulgarian documents using the suggested word-based approach.

The effectiveness of a video retrieval system largely depends on the choice of underlying text and image retrieval components. The unique properties of video collections (e.g., multiple sources, noisy features and temporal relations) suggest we examine the performance of these retrieval methods in such a multimodal environment, and identify the relative importance of the underlying retrieval components. In this paper, we review a variety of text/image retrieval approaches as well as their individual components in the context of broadcast news video. Numerous components of text/image retrieval have been discussed in detail, including retrieval models, text sources, temporal expansion methods, query expansion methods, image features, and similarity measures. For each component, we conduct a series of retrieval experiments on TRECVID video collections to identify their advantages and disadvantages. To provide a more complete coverage of video retrieval, we briefly discuss an emerging approach called concept-based video retrieval, and review strategies for combining multiple retrieval outputs.

Although considerable research has been conducted in the field of hierarchical text categorization, little has been done on automatically collecting labeled corpus for building hierarchical taxonom...

In retrieving medical free text, users are often interested in answers pertinent to certain scenarios that correspond to common tasks performed in medical practice, e.g., treatment or diagnosis of a disease. A major challenge in handling such queries is that scenario terms in the query (e.g., treatment) are often too general to match specialized terms in relevant documents (e.g., chemotherapy). In this paper, we propose a knowledge-based query expansion method that exploits the UMLS knowledge source to append the original query with additional terms that are specifically relevant to the query's scenario(s). We compared the proposed method with traditional statistical expansion that expands terms which are statistically correlated but not necessarily scenario specific. Our study on two standard testbeds shows that the knowledge-based method, by providing scenario-specific expansion, yields notable improvements over the statistical method in terms of average precision-recall. On the OHSUMED testbed, for example, the improvement is more than 5% averaging over all scenario-specific queries studied and about 10% for queries that mention certain scenarios, such as treatment of a disease and differential diagnosis of a symptom/disease.

The paper presents several techniques for selecting noun phrases for interactive query expansion following pseudo-relevance feedback and a new phrase-based document ranking method. A combined syntactico-statistical method was used for the selection of phrases for query expansion. Several statistical measures of phrase selection were evaluated. Experiments were also conducted studying the effectiveness of noun phrases in document ranking. One of the major problems in phrase-based document retrieval is weighting of overlapping and non-contiguous word sequences in documents. The paper presents a new method of phrase weighting, which addressed this problem, and its evaluation on the TREC dataset.

Sentence level novelty detection aims at spotting sentences with novel information from an ordered sentence list. In the task, sentences appearing later in the list with no new meanings are eliminated. For the task of novelty detection, the contributions of this paper are three-fold. First, conceptually, this paper reveals the computational nature of the task currently overlooked by the Novelty community--Novelty as a combination of partial overlap (PO) and complete overlap (CO) relations between sentences. We define partial overlap between two sentences as a sharing of common facts, while complete overlap is when one sentence covers all of the meanings of the other sentence. Second, technically, a novel approach, the selected pool method is provided which follows naturally from the PO-CO computational structure. We provide formal error analysis for selected pool and methods based on this PO-CO framework. We address the question how accurate must the PO judgments be to outperform the baseline pool method. Third, experimentally, results were presented for all the three novelty datasets currently available. Results show that the selected pool is significantly better or no worse than the current methods, an indication that the term overlap criterion for the PO judgments could be adequately accurate.

The complexity and diversity of government regulations make understanding and retrieval of regulations a non-trivial task. One of the issues is the existence of multiple sources of regulations and interpretive guides with differences in format, terminology and context. This paper describes a comparative analysis scheme developed to help retrieval of related provisions from different regulatory documents. Specifically, the goal is to identify the most strongly related provisions between regulations. The relatedness analysis makes use of not only traditional term match but also a combination of feature matches, and not only content comparison but also structural analysis. Regulations are first compared based on conceptual information as well as domain knowledge through feature matching. Regulations also possess specific organizational structures, such as a tree hierarchy of provisions and heavy referencing between provisions. These structures represent useful information in locating related provisions, and are therefore exploited in the comparison of regulations for completeness. System performance is evaluated by comparing a similarity ranking produced by users with the machine-predicted ranking. Ranking produced by the relatedness analysis system shows a reduction in error compared to that of Latent Semantic Indexing. Various pairs of regulations are compared and the results are analyzed along with observations based on different feature usages. An example of an e-rulemaking scenario is shown to demonstrate capabilities and limitations of the prototype relatedness analysis system.