Knowledge Management Techniques to Augment Textual CBR (PhD Thesis)

The past decade had witnessed an unprecedented growth in the amount of available digital content, and its volume is expected to continue to grow the next few years. Unstructured text data generated from web and enterprise sources form a large fraction of such content. Many of these contain large volumes of reusable data such as solutions to frequently occurring problems, and general know-how that may be reused in appropriate contexts. In this work, we address issues around leveraging unstructured text data from sources as diverse as the web and the enterprise within the Case-based Reasoning framework. Case-based Reasoning (CBR) provides a framework and methodology for systematic reuse of historical knowledge that is available in the form of problemsolution
pairs, in solving new problems. Here, we consider possibilities of enhancing Textual CBR systems under three main themes: procurement, maintenance and retrieval. We adapt and build upon the stateof-the-art techniques from data mining and natural language processing in addressing various challenges therein. Under procurement, we investigate the problem of extracting cases (i.e., problem-solution pairs) from data sources such as incident/experience
reports. We develop case-base maintenance methods specifically tuned to text targeted towards retaining solutions such that the utility of the filtered case base in solving new problems is maximized. Further, we address the problem of query suggestions for textual case-bases and show that exploiting the problem-solution partition can enhance retrieval effectiveness by prioritizing more useful query suggestions. Additionally, we illustrate interpretable clustering as a tool to drill-down to domain specific text collections (since CBR systems are usually very domain specific) and develop techniques for improved similarity assessment in social media sources such as microblogs. Through extensive empirical evaluations, we illustrate the improvements that we are able to
achieve over the state-of-the-art methods for the respective tasks.

Medical information extraction in European Portuguese

The electronic storage of medical patient data is becoming a daily experience in most of the practices and hospitals worldwide. However, much of the data available is in free-form text, a convenient way of expressing concepts and events, but especially challenging if one wants to perform automatic searches, summarization or statistical analysis. Information Extraction can relieve some of these problems by offering a semantically informed interpretation and abstraction of the texts. MedInX, the Medical Information eXtraction system presented in this document, is the first information extraction system developed to process textual clinical discharge records written in Portuguese. The main goal of the system is to improve access to the information locked up in unstructured text, and, consequently, the efficiency of the health care process, by allowing faster and reliable access to quality information on health, for both patient and health professionals. MedInX components are based on Natural Language Processing principles, and provide several mechanisms to read, process and utilize external resources, such as terminologies and ontologies, in the process of automatic mapping of free text reports onto a structured representation. However, the flexible and scalable architecture of the system, also allowed its application to the task of Named Entity Recognition on a shared evaluation contest focused on Portuguese general domain free-form texts. The evaluation of the system on a set of authentic hospital discharge letters indicates that the system performs with 95% F-measure, on the task of entity recognition, and 95% precision on the task of relation extraction. Example applications, demonstrating the use of MedInX capabilities in real applications in the hospital setting, are also presented in this document. These applications were designed to answer common clinical problems related with the automatic coding of diagnoses and other health-related conditions described in the documents, according to the international classification systems ICD-9-CM and ICF. The automatic review of the content and completeness of the documents is an example of another developed application, denominated MedInX Clinical Audit system.

A corpus-based translation study on english-persian verb phrase ellipsis

Dissertação de mest., Natural Language Processing & Human Language Technology, Faculdade de Ciências Humanas e Sociais, Univ. do Algarve, 2011

Equações de 1.º grau : estratégias e erros na resolução e simplificação de equações de 1.º grau

Relatório da prática de ensino supervisionada, Mestrado em Ensino da Matemática, Universidade de Lisboa, 2011

Representações em situações problemáticas que envolvem inequações do 1.º grau a uma incógnita : um estudo com alunos do 9.º ano de escolaridade

Relatório da Prática de Ensino Supervisionada, Ensino da Matemática, Universidade de Lisboa, 2013

A resolução de problemas envolvendo figuras geométricas : um estudo no 7º ano

Relatório da Prática de Ensino Supervisionada, Mestrado em Ensino de Matemática, Universidade de Lisboa, 2014

Open-domain web-based multiple document : question answering for list questions with support for temporal restrictors

Tese de doutoramento, Informática (Ciências da Computação), Universidade de Lisboa, Faculdade de Ciências, 2015

A aprendizagem do conceito de limite no 11º ano de escolaridade

Relatório da Prática de Ensino Supervisionada, Mestrado em Ensino da Matemática, Universidade de Lisboa, 2015

Entangled semantics

In the context of monolingual and bilingual retrieval, Simple Knowledge Organisation System (SKOS) datasets can play a dual role as knowledge bases for semantic annotations and as language-independent resources for translation. With no existing track of formal evaluations of these aspects for datasets in SKOS format, we describe a case study on the usage of the Thesaurus for the Social Sciences in SKOS format for a retrieval setup based on the CLEF 2004-2006 Domain-Specific Track topics, documents and relevance assessments. Results showed a mixed picture with significant system-level improvements in terms of mean average precision in the bilingual runs. Our experiments set a new and improved baseline for using SKOS-based datasets with the GIRT collection and are an example of component-based evaluation.

F-LQE: a Fuzzy Link Quality Estimator for wireless sensor networks

Radio Link Quality Estimation (LQE) is a fundamental building block for Wireless Sensor Networks, namely for a reliable deployment, resource management and routing. Existing LQEs (e.g. PRR, ETX, Fourbit, and LQI ) are based on a single link property, thus leading to inaccurate estimation. In this paper, we propose F-LQE, that estimates link quality on the basis of four link quality properties: packet delivery, asymmetry, stability, and channel quality. Each of these properties is defined in linguistic terms, the natural language of Fuzzy Logic. The overall quality of the link is specified as a fuzzy rule whose evaluation returns the membership of the link in the fuzzy subset of good links. Values of the membership function are smoothed using EWMA filter to improve stability. An extensive experimental analysis shows that F-LQE outperforms existing estimators.

Segmentation and Clustering of Textual Sequences: a Typological Approach

The long term goal of this research is to develop a program able to produce an automatic segmentation and categorization of textual sequences into discourse types. In this preliminary contribution, we present the construction of an algorithm which takes a segmented text as input and attempts to produce a categorization of sequences, such as narrative, argumentative, descriptive and so on. Also, this work aims at investigating a possible convergence between the typological approach developed in particular in the field of text and discourse analysis in French by Adam (2008) and Bronckart (1997) and unsupervised statistical learning.

Generating Relation Algebras for Qualitative Spatial Reasoning

Basic relationships between certain regions of space are formulated in natural language in everyday situations. For example, a customer specifies the outline of his future home to the architect by indicating which rooms should be close to each other. Qualitative spatial reasoning as an area of artificial intelligence tries to develop a theory of space based on similar notions. In formal ontology and in ontological computer science, mereotopology is a first-order theory, embodying mereological and topological concepts, of the relations among wholes, parts, parts of parts, and the boundaries between parts. We shall introduce abstract relation algebras and present their structural properties as well as their connection to algebras of binary relations. This will be followed by details of the expressiveness of algebras of relations for region based models. Mereotopology has been the main basis for most region based theories of space. Since its earliest inception many theories have been proposed for mereotopology in artificial intelligence among which Region Connection Calculus is most prominent. The expressiveness of the region connection calculus in relational logic is far greater than its original eight base relations might suggest. In the thesis we formulate ways to automatically generate representable relation algebras using spatial data based on region connection calculus. The generation of new algebras is a two pronged approach involving splitting of existing relations to form new algebras and refinement of such newly generated algebras. We present an implementation of a system for automating aforementioned steps and provide an effective and convenient interface to define new spatial relations and generate representable relational algebras.

Region Connection Calculus: Composition Tables and Constraint Satisfaction Problems

Qualitative spatial reasoning (QSR) is an important field of AI that deals with qualitative aspects of spatial entities. Regions and their relationships are described in qualitative terms instead of numerical values. This approach models human based reasoning about such entities closer than other approaches. Any relationships between regions that we encounter in our daily life situations are normally formulated in natural language. For example, one can outline one's room plan to an expert by indicating which rooms should be connected to each other. Mereotopology as an area of QSR combines mereology, topology and algebraic methods. As mereotopology plays an important role in region based theories of space, our focus is on one of the most widely referenced formalisms for QSR, the region connection calculus (RCC). RCC is a first order theory based on a primitive connectedness relation, which is a binary symmetric relation satisfying some additional properties. By using this relation we can define a set of basic binary relations which have the property of being jointly exhaustive and pairwise disjoint (JEPD), which means that between any two spatial entities exactly one of the basic relations hold. Basic reasoning can now be done by using the composition operation on relations whose results are stored in a composition table. Relation algebras (RAs) have become a main entity for spatial reasoning in the area of QSR. These algebras are based on equational reasoning which can be used to derive further relations between regions in a certain situation. Any of those algebras describe the relation between regions up to a certain degree of detail. In this thesis we will use the method of splitting atoms in a RA in order to reproduce known algebras such as RCC15 and RCC25 systematically and to generate new algebras, and hence a more detailed description of regions, beyond RCC25.

Apprentissage de représentations sur-complètes par entraînement d’auto-encodeurs

Les avancés dans le domaine de l’intelligence artificielle, permettent à des systèmes informatiques de résoudre des tâches de plus en plus complexes liées par exemple à la vision, à la compréhension de signaux sonores ou au traitement de la langue. Parmi les modèles existants, on retrouve les Réseaux de Neurones Artificiels (RNA), dont la popularité a fait un grand bond en avant avec la découverte de Hinton et al. [22], soit l’utilisation de Machines de Boltzmann Restreintes (RBM) pour un pré-entraînement non-supervisé couche après couche, facilitant grandement l’entraînement supervisé du réseau à plusieurs couches cachées (DBN), entraînement qui s’avérait jusqu’alors très difficile à réussir. Depuis cette découverte, des chercheurs ont étudié l’efficacité de nouvelles stratégies de pré-entraînement, telles que l’empilement d’auto-encodeurs traditionnels(SAE) [5, 38], et l’empilement d’auto-encodeur débruiteur (SDAE) [44]. C’est dans ce contexte qu’a débuté la présente étude. Après un bref passage en revue des notions de base du domaine de l’apprentissage machine et des méthodes de pré-entraînement employées jusqu’à présent avec les modules RBM, AE et DAE, nous avons approfondi notre compréhension du pré-entraînement de type SDAE, exploré ses différentes propriétés et étudié des variantes de SDAE comme stratégie d’initialisation d’architecture profonde. Nous avons ainsi pu, entre autres choses, mettre en lumière l’influence du niveau de bruit, du nombre de couches et du nombre d’unités cachées sur l’erreur de généralisation du SDAE. Nous avons constaté une amélioration de la performance sur la tâche supervisée avec l’utilisation des bruits poivre et sel (PS) et gaussien (GS), bruits s’avérant mieux justifiés que celui utilisé jusqu’à présent, soit le masque à zéro (MN). De plus, nous avons démontré que la performance profitait d’une emphase imposée sur la reconstruction des données corrompues durant l’entraînement des différents DAE. Nos travaux ont aussi permis de révéler que le DAE était en mesure d’apprendre, sur des images naturelles, des filtres semblables à ceux retrouvés dans les cellules V1 du cortex visuel, soit des filtres détecteurs de bordures. Nous aurons par ailleurs pu montrer que les représentations apprises du SDAE, composées des caractéristiques ainsi extraites, s’avéraient fort utiles à l’apprentissage d’une machine à vecteurs de support (SVM) linéaire ou à noyau gaussien, améliorant grandement sa performance de généralisation. Aussi, nous aurons observé que similairement au DBN, et contrairement au SAE, le SDAE possédait une bonne capacité en tant que modèle générateur. Nous avons également ouvert la porte à de nouvelles stratégies de pré-entraînement et découvert le potentiel de l’une d’entre elles, soit l’empilement d’auto-encodeurs rebruiteurs (SRAE).

Désambiguisation de sens par modèles de contextes et son application à la Recherche d’Information

Il est connu que les problèmes d'ambiguïté de la langue ont un effet néfaste sur les résultats des systèmes de Recherche d'Information (RI). Toutefois, les efforts de recherche visant à intégrer des techniques de Désambiguisation de Sens (DS) à la RI n'ont pas porté fruit. La plupart des études sur le sujet obtiennent effectivement des résultats négatifs ou peu convaincants. De plus, des investigations basées sur l'ajout d'ambiguïté artificielle concluent qu'il faudrait une très haute précision de désambiguation pour arriver à un effet positif. Ce mémoire vise à développer de nouvelles approches plus performantes et efficaces, se concentrant sur l'utilisation de statistiques de cooccurrence afin de construire des modèles de contexte. Ces modèles pourront ensuite servir à effectuer une discrimination de sens entre une requête et les documents d'une collection. Dans ce mémoire à deux parties, nous ferons tout d'abord une investigation de la force de la relation entre un mot et les mots présents dans son contexte, proposant une méthode d'apprentissage du poids d'un mot de contexte en fonction de sa distance du mot modélisé dans le document. Cette méthode repose sur l'idée que des modèles de contextes faits à partir d'échantillons aléatoires de mots en contexte devraient être similaires. Des expériences en anglais et en japonais montrent que la force de relation en fonction de la distance suit généralement une loi de puissance négative. Les poids résultant des expériences sont ensuite utilisés dans la construction de systèmes de DS Bayes Naïfs. Des évaluations de ces systèmes sur les données de l'atelier Semeval en anglais pour la tâche Semeval-2007 English Lexical Sample, puis en japonais pour la tâche Semeval-2010 Japanese WSD, montrent que les systèmes ont des résultats comparables à l'état de l'art, bien qu'ils soient bien plus légers, et ne dépendent pas d'outils ou de ressources linguistiques. La deuxième partie de ce mémoire vise à adapter les méthodes développées à des applications de Recherche d'Information. Ces applications ont la difficulté additionnelle de ne pas pouvoir dépendre de données créées manuellement. Nous proposons donc des modèles de contextes à variables latentes basés sur l'Allocation Dirichlet Latente (LDA). Ceux-ci seront combinés à la méthodes de vraisemblance de requête par modèles de langue. En évaluant le système résultant sur trois collections de la conférence TREC (Text REtrieval Conference), nous observons une amélioration proportionnelle moyenne de 12% du MAP et 23% du GMAP. Les gains se font surtout sur les requêtes difficiles, augmentant la stabilité des résultats. Ces expériences seraient la première application positive de techniques de DS sur des tâches de RI standard.