Search

The impressive rise of user-generated content on the web in the hands of sites like Twitter imposes new challenges to search systems. The concept of real-time search emerges, increasing the role that efficient indexing and retrieval algorithms play in this scenario. Thousands of new updates need to be processed in the very moment they are generated and users expect content to be “searchable” within seconds. This lead to the develop of efficient data structures and algorithms that may face this challenge efficiently. In this work, we introduce the concept of index entry invalidator, a strategy responsible for keeping track of the evolution of the underlying vocabulary and selectively invalidate and evict those inverted index entries that do not considerably degrade retrieval effectiveness. Consequently, the index becomes smaller and may increase overall efficiency. We introduce and evaluate two approaches based on Time-to-Live and Sliding Windows criteria. We also study the dynamics of the vocabulary using a real dataset while the evaluation is carry out using a search engine specifically designed for real-time indexing and search., Facultad de Informática

International audience; Social media provides an environment of information exchange. They principally rely on their users to create content, to annotate others’ content and to make on-line relationships. The user activities reflect his opinions, interests, etc. in this environment. We focus on analysing this social environment to detect user interests which are the key elements for improving adaptation. This choice is motivated by the lack of information in the user profile and the inefficiency of the information issued from methods that analyse the classic user behaviour (e.g. navigation, time spent on web page, etc.). So, having to cope with an incomplete user profile, the user social network can be an important data source to detect user interests. The originality of our approach is based on the proposal of a new technique of interests' detection by analysing the accuracy of the tagging behaviour of a user in order to figure out the tags which really reflect the content of the resources. So, these tags are somehow comprehensible and can avoid tags “ambiguity” usually associated to these social annotations. The approach combines the tag, user and resource in a way that guarantees a relevant interests detection. The proposed approach has been tested and evaluated in the Delicious social database. For the evaluation, we compare the result issued from our approach using the tagging behaviour of the neighbours (the egocentric network and the communities) with the information yet known for the user (his profile). A comparative evaluation with the classical tag-based method of interests detection shows that the proposed approach is better.

We consider the recoverability of traditional R-tree index structures under concurrent updating transactions, an important issue that is neglected or treated inadequately in many proposals of R-tree concurrency control. We present two solutions to ARIES-based recovery of transactions on R-trees. These assume a standard fine-grained single-version update model with physiological write-ahead logging and steal-and-no-force buffering where records with uncommitted updates by a transaction may migrate from their original page to another page due to structure modifications caused by other transactions. Both solutions guarantee that an R-tree will remain in a consistent and balanced state in the presence of any number of concurrent forward-rolling and (totally or partially) backward-rolling multiaction transactions and in the event of process failures and system crashes. One solution maintains the R-tree in a strictly consistent state in which the bounding rectangles of pages are as tight as possible, while in the other solution this requirement is relaxed. In both solutions only a small constant number of simultaneous exclusive latches (write latches) are needed, and in the solution that only maintains relaxed consistency also the number of simultaneous nonexclusive latches is similarly limited. In both solutions, deletions are handled uniformly with insertions, and a logarithmic insertion-path length is maintained under all circumstances.

We present a query formulation language (called MashQL) in order to easily query and fuse structured data on the web. The main novelty of MashQL is that it allows people with limited IT skills to explore and query one (or multiple) data sources without prior knowledge about the schema, structure, vocabulary, or any technical details of these sources. More importantly, to be robust and cover most cases in practice, we do not assume that a data source should havean offline or inlineschema. This poses several language-design and performance complexities that we fundamentally tackle. To illustrate the query formulation power of MashQL, and without loss of generality, we chose the Data web scenario. We also chose querying RDF, as it is the most primitive data model hence, MashQL can be similarly used for querying relational databases and XML. We present two implementations of MashQL, an online mashup editor, and a Firefox add on. The former illustrates how MashQL can be used to query and mash up the Data web as simple as filtering and piping web feeds and the Firefox add on illustrates using the browser as a web composer rather than only a navigator. To end, we evaluate MashQL on querying two data sets, DBLP and DBPedia, and show that our indexing techniques allow instant user interaction. © 1989-2012 IEEE. 24 5 783 798 Cited By :15

Retrieving text from early printed books is particularly difficult because in these documents, the words are very close one to the other and, similarly to medieval manuscripts, there is a large use of ligatures and abbreviations. To address these problems, we propose a word indexing and retrieval technique that does not require word segmentation and is tolerant to errors in character segmentation. Two main principles characterize the approach. First, characters are identified in the pages and clustered with self-organizing map (SOM). During the retrieval, the similarity of characters is estimated considering the proximity of cluster centroids in the SOM space, rather than directly comparing the character images. Second, query words are matched with the indexed sequence of characters by means of a dynamic time warping (DTW)-based approach. The proposed technique integrates the SOM similarity and the information about the width of characters in the string matching process. The best path in the DTW array is identified considering the widths of matching words with respect to the query so as to deal with broken or touching symbols. The proposed method is tested on four copies of the Gutenberg Bibles.

Nowadays, most enterprises still use relational databases to store their data. However, in a modern and dynamic environment the relationships connecting stored data are gaining more importance than the data itself. In this situation, graph databases have proven to be quite useful. Regardless of which database management system is being used to manage the company data, the issue of database query performance must be considered due to the increasing amount of data. Indexes are data structures, which can be used to improve database query performance. Due to their significant impact on the database performance, indexes have been developed and implemented both in relational and graph databases. In this paper, an overview of various indexing methods in the previously mentioned database categories has been made in order to detect their similarities and differences.

Introduction: According to mortality rates in Iran, cardiovascular diseases, neoplasms, perinatal mortality, and respiratory tract diseases were top rate mortality in 2003(1382). To reduce mortality rate, Iranian medical community need to know more about recent therapeutic regimens. Two main medical databases are Pubmed and Embase. Researching Pubmed and Embase indexing methods and comparing MeSH with Emtree help users to do more successful search in these databases. Consequently, designers of national medical information database in Iran may construct a model for updating indexing methods and thesaurus. This study aimed at comparing indexing methods in Pubmed and Embase. Methods: This was an applied descriptive - analytical research. Research population was all of the descriptors in MeSH and Emtree and indexed articles from Pubmed and Embase about four selected fields. In the last 3 months in 2006, all of the descriptors of selected fields were extracted through a structured search strategy. Then needed data was extracted from 6381 descriptors and 3358 articles. For collecting data we used a checklist and a questionnaire. Nine factors (including phrased descriptors versus single word descriptors, number of words in phrased descriptor, descriptor on adjectives and substantives format versus prepositions and conjunctives format, transforming versus non-transforming descriptors, using different quotation sign in descriptor structure, using abbreviations and commencer as descriptor, using definitions in descriptors, descriptors with explanations, and providing comments) selected from standard and analyzed in thesauri. Data were analyzed by SPSS using t-test and z test. Results: Emtree in six factors, and MeSH in four factors are more similar to standards. Pubmed articles are indexed with average number of 21-30 indexing terms. Embase uses average number 31-40 indexing terms for each article. Conclusions: Emtree structure is more suitable for modeling. Embase indexing method is assignmentive and derivative indexing and does it specific and more exhaustive. Pubmed indexing method is derivative and exhaustive indexing.

The increasing availability of large motion databases, in addition to advancements in motion synthesis, has made motion indexing and classification essential for better motion composition. However, in order to achieve good connectivity in motion graphs, it is important to understand human behaviour human movement though is complex and difficult to completely describe. In this paper, we investigate the similarities between various emotional states with regards to the arousal and valence of the Russell's circumplex model. We use a variety of features that encode, in addition to the raw geometry, stylistic characteristics of motion based on Laban Movement Analysis (LMA). Motion capture data from acted dance performances were used for training and classification purposes. The experimental results show that the proposed features can partially extract the LMA components, providing a representative space for indexing and classification of dance movements with regards to the emotion. This work contributes to the understanding of human behaviour and actions, providing insights on how people express emotional states using their body, while the proposed features can be used as complement to the standard motion similarity, synthesis and classification methods. © 2015 The Eurographics Association and John Wiley & Sons Ltd. 34 6 262 276 Cited By :4

Special methods are required to interpret sparse diffraction patterns collected from peptide crystals at X-ray free-electron lasers. Bragg spots can be indexed from composite-image powder rings, with crystal orientations then deduced from a very limited number of spot positions., Still diffraction patterns from peptide nanocrystals with small unit cells are challenging to index using conventional methods owing to the limited number of spots and the lack of crystal orientation information for individual images. New indexing algorithms have been developed as part of the Computational Crystallography Toolbox (cctbx) to overcome these challenges. Accurate unit-cell information derived from an aggregate data set from thousands of diffraction patterns can be used to determine a crystal orientation matrix for individual images with as few as five reflections. These algorithms are potentially applicable not only to amyloid peptides but also to any set of diffraction patterns with sparse properties, such as low-resolution virus structures or high-throughput screening of still images captured by raster-scanning at synchrotron sources. As a proof of concept for this technique, successful integration of X-ray free-electron laser (XFEL) data to 2.5 Å resolution for the amyloid segment GNNQQNY from the Sup35 yeast prion is presented.

Valid and reliable information on the use and effects of chemicals is a key factor in the industry and not least within many regulatory agencies. Identification data from lists of substances sometimes leads to incomplete bibliographic analysis in the major chemical databases. The present study takes as its starting point environmentally important chemicals and the retrieval of selectively chosen substances in the four databases: SciFinder, Web of Science (WoS), Scopus and Google Scholar. The way chemical data are stored in the databases plays a major role in the recovery process but differences in coverage, sometimes major, are still found. No single database records all publications about a substance. Inspection of individual titles is necessary when performing a complete count of references. Special care is taken in order to make data from the different databases comparable using the same journals and time periods (2000-2009). A number of nomenclature as well as problems related to the chemical structure and function, often inherent in quantitative or qualitative bibliographic studies of chemicals, are discussed. The practical implications for registration of chemicals in different databases are demonstrated.

Image collections today are increasingly larger in size, and they continue to grow constantly. Without the help of image search systems these abundant visual records collected in many different fields and domains may remain unused and inaccessible. Many available image databases often contain complementary modalities, such as attached text resources, which can be used to build an index for querying with keywords. However, sometimes users do not have or do not know the right words to express what they need, and, in addition, keywords do not express all the visual variations that an image may contain. Using example images as queries can be viewed as an alternative in different scenarios such as searching images using a mobile phone with a coupled camera, or supporting medical diagnosis by searching a large medical image collection. Still, matching only visual features between the query and image databases may lead to undesirable results from the user's perspective. These conditions make the process of finding relevant images for a specific information need very challenging, time consuming or even frustrating. Instead of considering only a single data modality to build image search indexes, the simultaneous use of both, visual and text data modalities, has been suggested. Non-visual information modalities may provide complementary information to enrich the image representation. The goal of this research work is to study the relationships between visual contents and text terms to build useful indexes for image search. A family of algorithms based on matrix factorization are proposed for extracting the multimodal aspects from an image collection. Using this knowledge about how visual features and text terms correlate, a search index is constructed, which can be searched using keywords, example images or combinations of both. Systematic experiments were conducted on different data sets to evaluate the proposed indexing algorithms. The experimental results showed that multimodal indexing is an effective strategy for designing image search systems. Las colecciones de imágenes hoy en día son muy grandes y crecen constantemente. Sin la ayuda de sistemas para la búsqueda de imágenes esos abundantes registros visuales que han sido recolectados en diferentes areas del conocimiento pueden permanecer aislados sin uso. Muchas bases de datos de imágenes contienen modalidades de datos complementarias, como los recursos textuales que pueden ser utilizados para crear índices de búsqueda. Sin embargo, algunas veces los usuarios no tienen o no saben qué palabras utilizar para encontrar lo que necesitan, y adicionalmente, las palabras clave no expresan todas las variaciones visuales que una imagen puede tener. Utilizar imágenes de ejemplo para expresar la consulta puede ser visto como una alternativa, por ejemplo buscar imágenes con teléfonos móviles, o dar soporte al diagnóstico médico con las imágenes de los pacientes. Aún así, emparejar correctamente las características visuales de la consulta y las imágenes en la base de datos puede llevar a resultados semánticamente incorrectos. Estas condiciones hacen que el proceso de buscar imágenes relevantes para una necesidad de información particular sea una tarea difícil, que consume mucho tiempo o que incluso puede ser frustrante. En lugar de considerar solo una modalidad de datos para construir índices de búsqueda para imágenes, el uso simultáneo de las modalidades visual y textual ha sido sugerido. Las modalidades no visuales pueden proporcionar información complementaria para enriquecer la representación de las imágenes. El objetivo de este trabajo de investigación es estudiar las relaciones entre los contenidos visuales y los términos textuales, para construir índices de búsqueda útiles. Este trabajo propone una familia de algoritmos basados en factorización de matrices para extraer los aspectos multimodales de una colección de imágenes. Utilizando este conocimiento acerca de cómo las características visuales se correlacionan con los términos textuales, se construye un índice que puede ser consultado con palabras clave, imágenes de ejemplo o por combinaciones de estas dos. Se realizaron experimentos sistemáticos en diferentes conjuntos de datos para evaluar los algoritmos de indexamiento propuestos. Los resultados muestran que el indexamiento multimodal es una estrategia efectiva para diseñar sistemas de búsqueda de imágenes. Doctorado

El creciente volumen de datos que se usan en análisis de negocios y minería de datos ha ocasionado que se formularan y usaran tecnologías diferentes a las basadas exclusivamente en bases de datos relcionales. El uso de índices invertidos generalizados constituye un elemento fundamental de muchos productos que se ofrecen y desarrollan en la actualidad. El paquete que se describe proporciona precisamente una variación de tales índices. Sirve para indizar un conjunto cualquiera de elementos, tales como registros, pero no se circunscribe a hacerlo para una sola tabla, como es el caso en una base de datos. Como una versión anterior del paquete no proporcionaba la eficiencia suficiente, especialmente en cuanto al número de operaciones de entrada-salida, pero también en cuanto al volumen de datos a almacenar en disco, se crearon e implementaron nuevos modos de gestionar las listas invertidas. En especial, en lugar de usar listas parciales y árboles B, se incluyeron nuevos modos de armar y almacenar los arreglos, y en especial, utilizar bitmaps – cadenas de bits – para las listas invertidas. Un ejemplo del uso del paquete por una aplicación se incluyó para aclarar los conceptos. Se describen los algoritmos para efectuar operaciones entre subconjuntos de números de registros resultantes de consultas formuladas con anterioridad. _______________ A SYSTEM TO MANAGE KEYWORDS BY CONTEXT FOR A DATA COLLECTION. ABSTRACT: The increase in size of data collections used for business analytics and data mining has caused technologies which are not based exclusively on relational databases to be developed and used. The use of generalized inverted indices is one of the main elements of many products offered. The software product described here offers a variant of such indexes. It serves to index any set of elements such as records, but is not confined to a single table, as is the case in a database. Since a previous version of the same product was not efficient enough, especially concerning input-output operations and the volume of data to be stored, new ways to store and manage inverted lists were created and implemented. Particularly, instead of using partial lists and B trees, new methods including the use of bitmpas were introduced to build and store the arrays corresponding to the inverted lists. An example of an application that uses the software product was included to shed lights on the different concepts. Some of the algorithms used to perform the operations between subsets of the record numbers resulting from queries are described.

Die Dissertation „Ebenen des Themas. Zur Interaktion von Thema, Text und Wissen“ beschäftigt sich mit einem Kernkonzept der Textlinguistik, dem Thema. Dieses Konzept wird einerseits in einer Fülle von linguistischen Aufsätzen – allerdings nur einer einzigen Monographie – zumindest am Rande behandelt, andererseits besteht bislang weder ein Konsens über die Definition des Themas noch über seine Analyse. Die vorliegende Arbeit setzt an diesem Punkt an und entwickelt unter Einbeziehung der bestehenden sehr heterogenen Satz- und Textthema-Konzepte ein integratives Thema-Modell, das den Ausgangspunkt für ein Modell zur Operationalisierung der Analyse des Themas im Text bildet. Dabei wird aufgezeigt, welche Analyseschritte mittels welcher computerlinguistischen Verfahren automatisiert werden können, welche Ressourcen für eine „Verrechnung“ der Analysen der einzelnen Ebenen nötig und bereits verfügbar sind und wie die Architektur eines auf den theoretischen Überlegungen der vorliegenden Arbeit basierenden Softwaresystems zur maschinellen Themenidentifikation aussehen könnte. Die Entwicklung des integrativen Thema-Konzepts basiert auf einer Neubewertung bisheriger Forschungsansätze. Dabei wird gezeigt, dass die meisten der linguistischen Thema-Konzepte keineswegs konkurrierende und einander ausschließende Thema-Auffassungen darstellen, sondern dass sie lediglich verschiedene Perspektiven auf das Thema widerspiegeln, d.h. als einander ergänzende Auffassungen interpretiert werden können. Auf Basis eines Multi-Ebenen-Modells des Textes, das in den größeren Kontext von Kommunikation und Kognition eingebettet wird, wird dargestellt, wie die verschiedenen Ebenen des Themas und die verschiedenen Wissensbestände im Prozess der Kommunikation (Textproduktion und -rezeption) interagieren. Die Diskussion der linguistischen Thema-Forschung wird ergänzt durch die der computerlinguistischen Forschungsansätze zur maschinellen Themenidentifikation. Dabei wird gezeigt, dass diese nahezu immer ein Theoriedefizit im Hinblick auf die ihnen zugrunde liegende Thema-Konzeption aufweisen. Dieses Defizit ist mit dafür verantwortlich, dass zwischen den meisten Ansätzen zur Themenidentifikation und zur Keyword- bzw. Schlüsselwort-Identifikation de facto kein großer Unterschied besteht. An die theoretische Klärung und Modellentwicklung schließt sich die Entwicklung einer Architektur und die Beschreibung und Evaluation der Komponenten eines Software-Systems zur maschinellen Themenidentifikation an. In diesem Zusammenhang werden einerseits die Wissensbestände, die für die Analyse des Themas nötig sind, erörtert und auf existierende oder in der Entwicklung befindliche computerlinguistische Ressourcen bezogen. Andererseits wird dargestellt, in welcher Weise die verschiedenen Beschreibungsebenen des Textes selbst (Informationsstruktur, lexikalische, anaphorische, relationale, visuelle, logische sowie generische Texttypstruktur) Aufschluss über das Thema geben, d.h. das Thema reflektieren oder durch indexikalische Zeichen auf das Thema verweisen. Für jede Beschreibungsebene wird schließlich beschrieben und anhand eines Beispieltextes illustriert, welche bereits implementierten computerlinguistischen Verfahren zur maschinellen Analyse des Themas auf der jeweiligen Ebene eingesetzt werden können. Damit wurden mit der vorliegenden Arbeit zwei Desiderate erfüllt: Zum einen wurde ein Thema-Konzept entwickelt, das die bisherigen, sehr vielfältigen linguistischen Thema-Konzepte in ein komplexes Erklärungsmodell zu integrieren erlaubt. Zum anderen wurde ein Analysemodell für das Thema präsentiert, das eine Operationalisierung der Themenanalyse ermöglicht und als Grundlage für computerlinguistische Arbeiten zur maschinellen Themenidentifikation geeignet ist und dabei eine eindeutige Differenzierung der Themenidentifikation von der Keyword-Identifikation ermöglicht.

International audience; As a text, each job advertisement expresses rich information about the occupation at hand, such as competence needs (i.e. required degrees, field knowledge, task expertise or technical skills). To facilitate the access to this information, the SIRE project conducted a corpus based study of how to articulate HR expert ontologies with modern semi-supervised information extraction techniques. An adaptive semantic labeling framework is developed through a parallel work on retrieval rules and on latent semantic lexicons of terms and jargon phrases. In its operational stage, our prototype will collect online job ads and index their content into detailed RDF triples compatible with applications ranging from enhanced job search to automated labor-market analysis.

We describe one tool for Table of Content (ToC) identification and recognition from PDF books. This task is part of ongoing research on the development of tools for the semi-automatic conversion of PDF documents in the Epub format that can be read on several E-book devices. Among various sub-tasks, the ToC extraction and recognition is particularly useful for an easy navigation of book contents.The proposed tool first identifies the ToC pages. The bounding boxes of ToC titles in the book body are subsequently found in order to add suitable links in the Epub ToC. The proposed approach is tolerant to discrepancies between the ToC text and the corresponding titles. We evaluated the tool on several open access books edited by University Presses that are partner of the OAPEN EcontentPlus project

MashQL, a novel query formulation language for querying and mashing up structured data on the Web, doesn't require users to know the queried data's structure or the data itself to adhere to a schema. In this article, the authors address the fact that being a language, not merely an interface (and simultaneously schema-free), MashQL faces a particular set of challenges. In particular, the authors propose and evaluate a novel technique for optimizing queries over large data sets to allow instant user interaction. © 2006 IEEE. 14 3 58 67 Cited By :9

Network querying is a growing domain with vast applications ranging from screening compounds against a database of known molecules to matching sub-networks across species. Graph indexing is a powerful method for searching a large database of graphs. Most graph indexing methods to date tackle the exact matching (isomorphism) problem, limiting their applicability to specific instances in which such matches exist. Here we provide a novel graph indexing method to cope with the more general, inexact matching problem. Our method, SIGMA, builds on approximating a variant of the set-cover problem that concerns overlapping multi-sets. We extensively test our method and compare it to a baseline method and to the state-of-the-art Grafil. We show that SIGMA outperforms both, providing higher pruning power in all the tested scenarios.

National audience; Automatic construction of ontologies from text is generally based on retrieving text content. For a much more rich ontology we extend these approaches by taking into account the document structure and some external resources (like thesaurus of indexing terms of near domain). In this paper we describe how these external resources are at first analyzed and then exploited. This method has been applied on a geographical domain and the benefit has been evaluated.

Three post-processing methods are described that can be used to enhance the performance of concept classifiers in news video.Two of them, consider the semantic relationships between concepts. The first method, Ancestor Boosting, aims to improve classifier performance for the case where one concept class is the semantic ancestor (superset) of another concept class, which is a child (subset) of the first. The second method, Sibling Confusion Removal, considers the case where concepts are semantically mutually exclusive in order to lessen the frequency of false positives amongst sibling classifiers. A final method, Chi-square Boosting, improves classifier performance by considering the influence of frequently co-occurring concepts. Publication of master thesis ispartof: pages:81-92 ispartof: Proceedings of the 8th Dutch-Belgian information retrieval workshop (DIR 2008) pages:81-92 ispartof: The 8th Dutch-Belgian information retrieval workshop (DIR 2008) location:Maastricht date:14 Apr - 15 Apr 2008 status: published

The metric spaces model formalizes the similarity search concept in nontraditional databases. The goal is to build an index designed to save distance computations when answering similarity queries later. A large class of algorithms to build the index are based on partitioning the space in zones as compact as possible. Each zone stores a representative point, called center, and a few extra data that allow to discard the entire zone at query time without measuring the actual distance between the elements of the zone and the query object. The way in which the centers are selected affects the performance of the algorithm. In this paper, we introduce two new center selection techniques for compact partition based indexes. These techniques were evaluated using the Geometric Near-neighbor Access Tree (GNAT). We experimentally showed that they achieve good performance., Facultad de Informática

Information retrieval focuses on content-based searching in text documents. For this purpose, first text content must be represented, by using a representation language (like thesauri or classification schemes) or by performing free-text search. The latter approach uses either string-based or computer-linguistic methods (stemming, dictionary lookup, syntax analysis). For retrieval, weighting and ranking methods give better results than Boolean retrieval, and some of them also allow for relevance feedback. Retrieval of XML documents requires new methods for support weighting and ranking, specificity-oriented search, data types with vague predicates and vague structural conditions

We here expand on a previous paper concerning the role of logic in information retrieval (IR) modelling. In that paper, among other things, we had pointed out how different ways of understanding the contribution of logic to IR have sprung from the (always unstated) adherence to either the total or the partial knowledge assumption. Here we make our analysis more precise by relating this dichotomy to the notion of vividness, as used in knowledge representation, and to another dichotomy which has had a profound influence in database theory, namely the distinction between the proof-theoretic and the model-theoretic views of a database, spelled out by Reiter in his "logical econstruction of database theory". We show that precisely the same distinction can be applied to logical models of IR developed so far. The strengths and weaknesses of the adoption of either approach in logical models of IR are discussed.

The categorisation of documents into subject-specific categories is a useful enhancement for large document collections addressed by information retrieval systems, as a user can first browse a category tree in search of the category that best matches her interests, and then issue a query for more specific documents "from within the category". This approach combines two modalities in information seeking that are most popular in Web-based search engines, i.e. category-based site browsing (as exemplified by e.g. YAHOO) and keyword-based document querying (as exemplified by e.g. ALTAVISTA). Appropriate query expansion tools need to be provided, though, in order to allow the user to incrementally refine her query through further retrieval passes, thus allowing the system to produce a series of subsequent document rankings that hopefully converge to the user's expected ranking. In this work we propose that automatically generated, category-specific "associative" thesauri be used for such purpose. We discuss a method for their generation, and discuss how the thesaurus specific to a given category may usefully be endowed with "gateways" to the thesauri specific to its parent and children categories.