As advances in information and networking technologies, huge amount of information typically in the form of text documents are available online. To facilitate efficient and effective access to documents relevant to users¡¦ information needs, information retrieval systems have been imposed a more significant role than ever. One challenging issue in information retrieval is word mismatch that refers to the phenomenon that concepts may be described by different words in user queries and/or documents. The word mismatch problem, if not appropriately addressed, would degrade retrieval effectiveness critically of an information retrieval system. In this thesis, we develop a cluster-based query expansion technique to solve the word mismatch problem. Using the traditional query expansion techniques (i.e., global analysis and local feedback) as performance benchmarks, the empirical results suggest that when a user query only consists of one query term, the global analysis technique is more effective. However, if a user query consists of two or more query terms, the cluster-based query expansion technique can provide a more accurate query result, especially within the first few top-ranked documents retrieved.

Over the last decades, there have been remarkable shifts in the area of Information Retrieval (IR) as huge amount of information is increasingly accumulated on the Web. The gigantic information explosion increases the need for discovering new tools that retrieve meaningful knowledge from various complex information sources. Thus, techniques primarily used to search and extract important information from numerous database sources have been a key challenge in current IR systems. Topic modeling is one of the most recent techniquesthat discover hidden thematic structures from large data collections without human supervision. Several topic models have been proposed in various fields of study and have been utilized extensively for many applications. Latent Dirichlet Allocation (LDA) is the most well-known topic model that generates topics from large corpus of resources, such as text, images, and audio.It has been widely used in many areas in information retrieval and data mining, providing efficient way of identifying latent topics among document collections. However, LDA has a drawback that topic cohesion within a concept is attenuated when estimating infrequently occurring words. Moreover, LDAseems not to consider the meaning of words, but rather to infer hidden topics based on a statisticalapproach. However, LDA can cause either reduction in the quality of topic words or increase in loose relations between topics. In order to solve the previous problems, we propose a domain specific topic model that combines domain concepts with LDA. Two domain specific algorithms are suggested for solving the difficulties associated with LDA. The main strength of our proposed model comes from the fact that it narrows semantic concepts from broad domain knowledge to a specific one which solves the unknown domain problem. Our proposed model is extensively tested on various applications, query expansion, classification, and summarization, to demonstrate the effectiveness of the model. Experimental results show that the proposed model significantly increasesthe performance of applications.

This Master’s thesis is a comparative study of information retrieval results between three distinct search strategies in simulated automatic query expansion in a bibliographic database. Our purpose is to investigate which of the search strategies score the most effective precision and to what extent the same relevant documents are retrieved (overlapped). A thesaurus attached to the database is used to select appropriate descriptors for the baseline query formulations which subsequently are expanded with hierarchical relations. The search strategies are s1: A baseline query with two or three descriptors, s2: The baseline descriptors combined with at least one Narrower Term, s3: The baseline descriptors combined with Narrower Term and at least one Broader Term. A Document Cutoff Value of 15 is used and only the 15 highest ranked documents are judged by relevancy. The measurements used are precision for effectiveness and Jaccard’s index for overlap. In terms of precision, results reveal that s1 scores the highest value (average 84,8 %) with s2 and s3 in decreasing order (average 81,94 % and 61,41 % respectively). The overlap varies greatly depending on topic and the average is between s1 and s2 78,81 %, between s2 and s3 58,48 % and between s3 and s1 40,41 %. In short, average precision decreases as well as average overlap. The use of thesaurus in the applied strategy of automatic query expansion is not recommended in this specific database, if the aim is to increase precision. However, in single searches with the structure like s1 the thesaurus can be of assistance in the selection of specific search terms. / Uppsatsnivå: D

In query expansion (QE) terms are added to an initial query in order to improve retrieval effectiveness. In this thesis we use QE in the sense that a reformulation of the query is done by deleting the terms in the initial query and instead replacing them with terms from the documents retrieved in the initial run. The aim of this thesis is to, in a experimental full text invironment, study and compare the retrieval result of two different query expansion strategies in relation to each other. The following questions are addressed by the study: How do the two strategies perform in relation to each other regarding recall?What may be causing the result?Are the two strategies retrieving the same relevant documents?Two strategies are designed to simulate a searcher using automatic query expansion (AQE) either with or without relevance feedback. Strategy I is simulating AQE without relevance feedback by taking the top five documents that are retrieved in the initial run and then extracting the top ten most frequently occurring terms in these to create a new query. Correspondingly the Strategy II, is simulating AQE with relevance feedback by taking the top five relevant documents and extracting the top ten terms in these to create a new query. It is concluded that both of the strategies’ retrieval performance was improved for most of the topics. In average Strategy II did achieve 54.63 percent recall compared to Strategy I which did achieve 45.59 percent recall. The two strategies did retrieve different relevant documents for majority of the topics. Hence, it would be reasonable to base a system on both of them. / Uppsatsnivå: D

A busca de informações em bases de dados de instituições que possuem grande volume de dados necessita cada vez mais de processos mais eficientes para realização dessa tarefa. Problemas de grafia, idioma, sinonímia, abreviação de termos e a falta de padronização dos termos, tanto nos argumentos de busca, quanto na indexação dos documentos, interferem diretamente nos resultados. Diante disso, este estudo teve como objetivo avaliar o impacto da adição de vocabulários estruturados da área de Ciências da Saúde no Currículo Lattes, na recuperação de perfis similares de pesquisadores das áreas de Ciências Biológicas e Ciências da Saúde, utilizando técnicas de mineração de dados, expansão de consultas, modelos vetoriais de consultas e utilização de algoritmo de trigramas. Foram realizados cruzamentos de informações entre as palavras-chaves de artigos publicados registrados no Currículo Lattes e as informações contidas no Medical Subject Headings (MeSH) e nos Descritores em Ciências da Saúde (DeCS), bem como comparações entre os resultados das consultas, utilizando as palavras-chaves originais e adicionando-lhes os termos resultantes do processo de expansão de consultas. Os resultados mostram que a metodologia adotada neste estudo pode incrementar qualitativamente o universo de perfis recuperados, podendo dessa forma contribuir para a melhoria dos Sistemas de Informações do Conselho Nacional de Desenvolvimento Científico e Tecnológico - CNPq. / Information retrieval in large databases need increasingly more efficient ways for accomplishing this task. There are many problems, like spelling, language, synonym, acronyms, lack of standardization of terms, both in the search arguments, as in the indexing of documents. They directly interfere in the results. Thus, this study aimed to evaluate the impact of the addition of structured vocabularies of Health Sciences area in Lattes Database, in the recovery of similar profiles of researchers that work in Biological Sciences and Health Sciences, using Query Expansion, Data Mining procedures, Vector Models and Trigram Phrase Matching algorithm. Crosschecking keywords of articles registered in Lattes Database and Medical Subject Headings (MeSH) and Health Sciences Descriptors (DeCS) terms, as well as comparisons between the results of queries using the original keywords and adding them to query expansion terms. The results show that the methodology used in this study can qualitatively increase the set of recovered profiles, contributing to the improvement of CNPq Information Systems.

As advances of information technology, engineering consulting firms have gradually digitalized their documents and line-drawing images. Such digital libraries greatly facilitate document retrievals. However, engineers still face a challenging issue: searches and retrievals of line-drawing images in a digital library. With a small number of line-drawing images in a digital library, engineers can browse thumbnails for locating relevant images. As the number of line-drawing images increases, the manual browsing process is time-consuming and frustrated. In response to the need and importance of supporting efficient and effective retrieval of line-drawing images, this thesis aims to develop a line-drawing image retrieval system. Typically, a line-drawing image within an engineering document is associated with surrounding text for description or illustration purpose. Such surrounding text provides important information for automatically indexing the line-drawing image. With extracted indexes (or keywords), retrieval of line-drawing images can be accomplished using a traditional information retrieval technique. Specifically, in this study, we propose a line-drawing image retrieval system based on surrounding text. We develop four models for defining surrounding text boundaries for line-drawing images. Furthermore, two information retrieval techniques (one with and one without query expansion) are implemented and evaluated. According to our empirical evaluations, the surrounding text boundary model with image caption together with three sentences (preceding, image anchoring, and successive sentences) would result in the best retrieval effectiveness, as measured by recall and precision rates.

Στα πλαίσια της παρούσας διπλωματικής εργασίας, ασχολούμαστε με την αυτόματη κατηγοριοποίηση των αποτελεσμάτων των αναζητήσεων στον Παγκόσμιο Ιστό. Αρχικά, ορίζουμε τα χαρακτηριστικά των σελίδων που είναι κατάλληλα για κατηγοριοποίηση με βάση την πρόθεση του χρήστη. Έπειτα, με χρήση μεθόδων μείωσης της διαστατικότητας επιλέγουμε τα πιο αντιπροσωπευτικά από τα χαρακτηριστικά αυτά και αξιολογούμε την απόδοση διάφορων αλγορίθμων κατηγοριοποίησης. Ακολούθως, επιλέγουμε τον αλγόριθμο κατηγοριοποίησης που βασίζεται στα επιλεγμένα χαρακτηριστικά και επιτυγχάνει την καλύτερη απόδοση. Εφαρμόζοντας τον αλγόριθμο, κατηγοριοποιούμε τα αποτελέσματα των αναζητήσεων στον Παγκόσμιο Ιστό. Τέλος, προτείνουμε μια μέθοδο εξαγωγής όρων από τα κατηγοριοποιημένα αποτελέσματα και επαναδιατύπωσης του ερωτήματος με βάση τον σκοπό αναζήτησης του χρήστη. / In this thesis, we tackle the problem of automatic classification of search results in Web environment. First, we define web pages features that are convenient for classification based on the user’s intent. Next, we use dimensionality reduction techniques to choose the most representative features and we evaluate different classification algorithms. We choose the most efficient classification algorithm based on chosen features and by using it, we classify the results retrieved from web searches. In the end, we propose a method to extract terms from the classified results and to reformulate the query based on user intent.

Uma das principais tarefas de Recuperação de Informações é encontrar documentos que sejam relevantes a uma consulta. Esta tarefa é difícil porque, em muitos casos os termos de busca escolhidos pelo usuário são diferentes dos termos utilizados pelos autores dos documentos. Ao longo dos anos, várias abordagens foram propostas para lidar com este problema. Uma das técnicas mais utilizadas, com o objetivo de expandir o número de documentos relevantes recuperados é a Expansão de Consultas, que consiste em expandir a consulta com a adição de termos relacionados. Este trabalho propõe um método que utiliza o modelo de Cell Assemblies para a expansão da consulta. Cell Assemblies são grupos de neurônios conectados, com padrões de disparo, que permitem que a atividade persista mesmo após a remoção dos estímulos externos. A modificação das sinapses entre os neurônios é feita através de regras de aprendizagem Hebbiana. Neste trabalho, o modelo Cell Assemblies foi adaptado a fim de aprender os relacionamentos entre os termos de uma coleção de documentos. Esses relacionamentos são utilizados para expandir a consulta original com termos relacionados. A avaliação experimental sobre uma coleção de testes padrão em Recuperação de Informações mostrou que algumas consultas melhoraram significativamente seus resultados com a técnica proposta. / One of the main tasks in Information Retrieval is to match a user query to the documents that are relevant for it. This matching is challenging because in many cases the keywords the user chooses will be different from the words the authors of the relevant documents have used. Throughout the years, many approaches have been proposed to deal with this problem. One of the most popular consists in expanding the query with related terms with the goal of retrieving more relevant documents. In this work, we propose a new method in which a Cell Assembly model is applied for query expansion. Cell Assemblies are reverberating circuits of neurons that can persist long beyond the initial stimulus has ceased. They learn through Hebbian Learning rules and have been used to simulate the formation and the usage of human concepts. We adapted the Cell Assembly model to learn relationships between the terms in a document collection. These relationships are then used to augment the original queries. Our experiments use standard Information Retrieval test collections and show that some queries significantly improved their results with the proposed technique.

Submitted by Jorge Janaite Neto (janaite@gmail.com) on 2018-06-24T23:56:37Z No. of bitstreams: 1 janaite_neto_j_me_mar.pdf: 1649007 bytes, checksum: 66467a076d4f716197896c6dc3c5ee2b (MD5) / Approved for entry into archive by Satie Tagara (satie@marilia.unesp.br) on 2018-06-25T13:46:39Z (GMT) No. of bitstreams: 1 janaiteneto_j_me_mar.pdf: 1649007 bytes, checksum: 66467a076d4f716197896c6dc3c5ee2b (MD5) / Made available in DSpace on 2018-06-25T13:46:39Z (GMT). No. of bitstreams: 1 janaiteneto_j_me_mar.pdf: 1649007 bytes, checksum: 66467a076d4f716197896c6dc3c5ee2b (MD5) Previous issue date: 2018-05-30 / Não recebi financiamento / A recuperação de informação ocorre por meio da comparação entre as representações dos documentos de um acervo e a representação da necessidade de informação do usuário. Um documento é recuperado quando sua representação coincidir total ou parcialmente com a representação da necessidade de informação do usuário. O processo de recuperação de informação pode ser visto como um problema linguístico no qual o conteúdo informacional dos documentos e a necessidade de informação do usuário são representados por um conjunto de termos. A eficiência do processo de recuperação de informação depende da qualidade das representações dos documentos e dos termos empregados pelo usuário para representar sua necessidade de informação. Quanto mais compatíveis forem essas representações maior será a eficiência do processo de recuperação. A partir de uma pesquisa exploratória e descritiva fundamentada em bibliografia específica, este trabalho propõe a utilização de ontologias computacionais em sistemas de recuperação de informação baseados no Modelo Espaço Vetorial. As ontologias são empregadas como estrutura terminológica externa utilizadas tanto na expansão dos termos de indexação quanto na expansão dos termos que compõe a expressão de busca. A expansão dos termos de indexação é feita logo após a extração dos termos mais representativos do documento em análise durante o processo de indexação, consistindo na adição de novos termos conceitualmente relacionados a fim de enriquecer a representação do documento. A expansão da consulta é obtida a partir da adição de novos termos relacionados aos já existentes na expressão de busca com o objetivo de melhor contextualizá-los. Nesta proposta utiliza-se apenas a estrutura terminológica e hierárquica oferecida por uma ontologia computacional OWL, sem considerar os demais tipos de relações possíveis nem as restrições lógicas que podem ser descritas, podendo esses recursos serem utilizados em trabalhos futuros na tentativa de melhorar ainda mais a eficiência do processo de recuperação. A proposta apresentada neste estudo pode ser implementada e futuramente tornar-se um sistema de recuperação de informação totalmente operacional. / The information retrieval occurs by means of match between the representations of documents from a collection and the representation of user information’s needs. A document is retrieved when its representation matches totally or partially to the user information’s needs. The process of information retrieval can be seen as a linguistic issue in which the document information content and the user information need are represented by a set of terms. Its efficiency depends on the quality of the representations of the documents and the terms used to represent the user’s information need. The more compatible these representations were, the more efficient the retrieval process. Based on an exploratory and descriptive research substantiated in a specific bibliography, this paper offers to use computational ontologies in information retrieval systems based on the Vector Space Model. The ontologies are applied as external terminological structures used in the indexing terms expansion as well as in the expansion of the terms which compound the query expression. The indexing terms expansion is made as soon as the extraction of the more representative terms of the document in analysis during the indexing process, consisting on the adding of new conceptually related terms in order to improve the document representation. Query expansion is obtained from adding new related terms to the existent ones in the query expression to better contextualize them. In this propose, only the terminological and hierarchical structure offered by an OWL computational ontology was used, regardless other possible relations and logical restrictions that could be descripted, saving these resources to be used in further works in an attempt to improve the retrieval process efficiency. The shown proposition can be implemented and become a fully operational information retrieval system.

Uma das principais tarefas de Recuperação de Informações é encontrar documentos que sejam relevantes a uma consulta. Esta tarefa é difícil porque, em muitos casos os termos de busca escolhidos pelo usuário são diferentes dos termos utilizados pelos autores dos documentos. Ao longo dos anos, várias abordagens foram propostas para lidar com este problema. Uma das técnicas mais utilizadas, com o objetivo de expandir o número de documentos relevantes recuperados é a Expansão de Consultas, que consiste em expandir a consulta com a adição de termos relacionados. Este trabalho propõe um método que utiliza o modelo de Cell Assemblies para a expansão da consulta. Cell Assemblies são grupos de neurônios conectados, com padrões de disparo, que permitem que a atividade persista mesmo após a remoção dos estímulos externos. A modificação das sinapses entre os neurônios é feita através de regras de aprendizagem Hebbiana. Neste trabalho, o modelo Cell Assemblies foi adaptado a fim de aprender os relacionamentos entre os termos de uma coleção de documentos. Esses relacionamentos são utilizados para expandir a consulta original com termos relacionados. A avaliação experimental sobre uma coleção de testes padrão em Recuperação de Informações mostrou que algumas consultas melhoraram significativamente seus resultados com a técnica proposta. / One of the main tasks in Information Retrieval is to match a user query to the documents that are relevant for it. This matching is challenging because in many cases the keywords the user chooses will be different from the words the authors of the relevant documents have used. Throughout the years, many approaches have been proposed to deal with this problem. One of the most popular consists in expanding the query with related terms with the goal of retrieving more relevant documents. In this work, we propose a new method in which a Cell Assembly model is applied for query expansion. Cell Assemblies are reverberating circuits of neurons that can persist long beyond the initial stimulus has ceased. They learn through Hebbian Learning rules and have been used to simulate the formation and the usage of human concepts. We adapted the Cell Assembly model to learn relationships between the terms in a document collection. These relationships are then used to augment the original queries. Our experiments use standard Information Retrieval test collections and show that some queries significantly improved their results with the proposed technique.