Search

Esiste un numero significativo di documenti, report e pagine Web – un'analisi riporta 233 milioni di tabelle relazionali nel repository Common Crawl contenente un totale 2,85 miliardi di documenti – che fanno uso di tabelle per fornire informazioni che non possono essere facilmente elaborate dagli umani o capite dai computer. Per risolvere questo problema proponiamo un nuovo approccio che permetterà ai computer di interpretare la semantica di una tabella, e fornirà agli umani una rappresentazione più accessibile dei dati contenuti in essa. Per conseguire questo obiettivo, il problema principale è stato suddiviso in tre sotto-problemi: (i) la definizione di un metodo per fornire un'interpretazione semantica dei dati di una tabella; (ii) la definizione di un modello descrittivo che permetta ai computer di capire e condividere dati di una tabella; e (iii) la definizione di processi, tecniche e algoritmi per generare rappresentazioni dei dati in linguaggio naturale. Per quanto riguarda il sotto-problema (i), la rappresentazione semantica dei dati è stata ottenuta attraverso l'applicazione di tecniche di interpretazione di tabelle (table interpretation), che aiuta gli utenti ad identificare, in una maniera semi-automatica, il significato dei dati di una tabella e le relazioni tra di essi. Queste tecniche considerano in input una tabella e un Knowledge Graph, e restituiscono una rappresentazione RDF – un set di tuple – del contenuto della tabella, facendo riferimento ai concetti e alle proprietà del KG. Questa dissertazione presenta un nuovo approccio che, a partire dai lavori presenti in letteratura, ha portato allo sviluppo di un nuovo strumento, chiamato MantisTable, che effettua automaticamente un'interpretazione semantica completa della tabella. Gli esperimenti condotti hanno mostrato buoni risultati, rispetto alle tecniche e ai tool simili. Il sotto-problema (ii) è stato affrontato con la definizione di nuovi modi di rappresentazione dei dati: è stato definito un nuovo tipo di descrizione che combina la specifica OpenAPI con il linguaggio JSON-LD. I risultati delle tecniche di interpretazione semantica delle tabelle vengono così sfruttati per migliorare un formato già popolare, permettendo il recupero e il processamento dei dati tabellari. Il sotto-problema (iii) è stato affrontato definendo una tecnica di generazione del linguaggio naturale che utilizza una rete neurale per trasformare dati RDF, ottenuti dall'interpretazione delle tabelle, in frasi. Grazie a queste frasi, è possibile creare una rappresentazione testuale del contenuto delle tabelle. Questa è poi estendibile con informazioni aggiuntive provenienti da fonti che possono essere selezionate automaticamente utilizzando l'annotazione semantica. A significant number of documents, reports and Web pages –an analysis reports 233M relational tables within the Common Crawl repository of 1.81 billion documents– makes use of tables to convey information that cannot be easily processed by humans, and understood by computers. To address this issue, we propose a new approach that allows computers to interpret the semantics of a table, and provides humans with a more accessible representation of the data contained in a table. To achieve the objective, the general problem has been broken down into three sub-problems: (i) define a method to provide a semantic interpretation of table data; (ii) define a descriptive model that allows computers to understand and share table data; and (iii) define processes, techniques and algorithms to generate natural language representation of the table data. Regarding sub-problem (i), the semantic representation of a data has been obtained through the application of table interpretation techniques, which supports users to identify in a semi-automatic way the meaning of the data in the table and the relationships between them. Such techniques take a table and a Knowledge Graph (KG) as input, and deliver as output an RDF representation –a set of tuples –. The output contains the input table annotated with the KG concepts and properties. This thesis presents a new approach, rooted in the existing literature, to laid the foundations for the development of a new tool -called MantisTable- which automatically performs a complete semantic interpretation of a table. The conducted experiments have shown good results compared to similar techniques. Sub-problem (ii) has been tackled by defining new ways of representing data. A new kind of description has been defined that combines the OpenAPI specification with the JSON-LD. The results of semantic table interpretation techniques are exploited to enhance a popular description format and allow automatic retrieval and processing of table data. Sub-problem (iii) has been addressed by defining a natural language generation technique that uses a neural network to translate RDF data obtained from table interpretation into sentences. Thanks to these sentences, it is possible to create a textual representation of the content of the table, and possibly extend it with additional information from data sources that can be selected automatically using semantic annotations.

Lo scopo di questa tesi è di studiare due principali categorie di algoritmi per la detection di oggetti e il loro uso in particolari applicazioni. La prima categoria esaminata riguarda approcci basati su Keypoint. Diversi esperimenti comparativi vengono eseguiti all'interno della pipeline standard del modello di test MPEG CDVS e viene proposta una pipeline estesa che fa uso di informazione colore. La seconda categoria di object detectors oggetto di indagine si basa su Reti neurali convoluzionali. In particolare, vengono affrontate due applicazioni di reti neurali convoluzionali per il riconoscimento di oggetti. Il primo riguarda il riconoscimento di loghi commerciali. Due pipeline di classificazione sono progettate e testate su un set di immagini raccolte da Flickr. La prima architettura utilizza una rete pre-addestrata come feature extractor e raggiunge risultati comparabili a quelli di algoritmi basati Keypoint. La seconda architettura si avvale di una rete neurale che supera le performances di metodi stato dell'arte basati su Keypoint. L'altra applicazione esaminata è la categorizzazione di dipinti che consiste nell'associare l'autore, nell'assegnare un dipinto alla scuola o al movimento artistico a cui appartiene, e classificare il genere del dipinto, ad es. paesaggio, ritratto, illustrazione ecc. Per affrontare questo problema, viene proposta una struttura di rete neurale multibranch e multitask che beneficia dell'uso congiunto di approcci basati su keypoint e di features neurali. In entrambe le applicazioni viene anche esaminato l'uso di tecniche di data augmentation per ampliare il training set. In particolare per i dipinti, un algoritmo di trasferimento di stile pittorico basato su reti neurali viene sfruttato per generare quadri sintetici da utilizzare in fase di training. The aim of this thesis is to study two main categories of algorithms for object detection and their use in particular applications. The first category that is investigated concerns Keypoint-based approaches. Several comparative experiments are performed within the standard testing pipeline of the MPEG CDVS Test Model and an extended pipeline which make use of color information is proposed. The second category of object detectors that is investigated is based on Convolutional Neural Networks. Two applications of Convolutional Neural Networks for object recognition are in particular addressed. The first concerns logo recognition. Two classification pipelines are designed and tested on a real-world dataset of images collected from Flickr. The first architecture makes use of a pre-trained network as feature extractor and it achieves comparable results keypoint based approaches. The second architecture makes use of a tiny end-to-end trained Neural Network that outperformed state-of-the-art keypoint based methods. The other application addressed is Painting Categorization. It consists in associating the author, assigning a painting to the school or art movement it belongs to, and categorizing the genre of the painting, e.g. landscape, portrait, illustration etc. To tackle this problem, a novel multibranch and multitask Neural Network structure is proposed which benefit from joint use of keypoint-based approaches and neural features. In both applications the use of data augmentation techniques to enlarge the training set is also investigated. In particular for paintings, a neural style transfer algorithm is exploited for generating synthetic paintings to be used in training.

Nonostante l'elevato numero di dati pubblicati come LD, il loro utilizzo non ha ancora mostrato il loro potenziale per l’assenza di comprensione dei metadati. I consumatori di dati hanno bisogno di ottenere informazioni dai dataset in modo veloce e concentrato per poter decidere se sono utili per il loro problema oppure no. Le tecniche di profilazione dei dati offrono una soluzione efficace a questo problema in quanto sono utilizzati per generare metadati e statistiche che descrivono il contenuto dei dataset. Questa tesi presenta una ricerca, che affronta i problemi legati alla profilazione Linked Data. Nonostante il termine profilazione dei dati è usato in modo generico per diverse informazioni che descrivono i dataset, in questa tesi noi andiamo a ricoprire tre aspetti della profilazione; topic-based, schema-based e linkage-based. Il profilo proposto in questa tesi è fondamentale per il processo decisionale ed è la base dei requisiti che portano verso la comprensione dei dataset. In questa tesi presentiamo un approccio per classificare automaticamente insiemi di dati in una delle categorie utilizzate nel mondo dei LD. Inoltre, indaghiamo il problema della profilazione multi-topic. Per la profilazione schema-based proponiamo un approccio riassuntivo schema-based, che fornisce una panoramica sui rapporti nei dati. I nostri riassunti sono concisi e chiari sufficientemente per riassumere l'intero dataset. Inoltre, essi rivelano problemi di qualità e possono aiutare gli utenti nei compiti di formulazione dei query. Molti dataset nel LD cloud contengono informazioni simili per la stessa entità. Al fine di sfruttare appieno il suo potenziale LD bisogna far vedere questa informazione in modo esplicito. Profiling Linkage fornisce informazioni sul numero di entità equivalenti tra i dataset e rivela possibili errori.Le tecniche di profiling sviluppate durante questo lavoro sono automatiche e possono essere applicate a differenti insiemi di dati indipendentemente dal dominio. Recently, the increasing diffusion of Linked Data (LD) as a standard way to publish and structure data on the Web has received a growing attention from researchers and data publishers. LD adoption is reflected in different domains such as government, media, life science, etc., building a powerful Web available to anyone. Despite the high number of datasets published as LD, their usage is still not exploited as they lack comprehensive metadata. Data consumers need to obtain information about datasets content in a fast and summarized form to decide if they are useful for their use case at hand or not. Data profiling techniques offer an efficient solution to this problem as they are used to generate metadata and statistics that describe the content of the dataset. Existing profiling techniques do no cover a wide range of use cases. Many challenges due to the heterogeneity nature of Linked Data are still to overcome. This thesis presents the doctoral research which tackles the problems related to Profiling Linked Data. Even though the term of data profiling is the umbrella term for diverse descriptive information that describes a dataset, in this thesis we cover three aspects of profiling; topic-based, schema-based and linkage-based. The profile provided in this thesis is fundamental for the decision-making process and is the basic requirement towards the dataset understanding. In this thesis we present an approach to automatically classify datasets in one of the topical categories used in the LD cloud. Moreover, we investigate the problem of multi-topic profiling. For the schema-based profiling we propose a schema-based summarization approach, that provides an overview about the relations in the data. Our summaries are concise and informative enough to summarize the whole dataset. Moreover, they reveal quality issues and can help users in the query formulation tasks. Many datasets in the LD cloud contain similar information for the same entity. In order to fully exploit its potential LD should made this information explicit. Linkage profiling provides information about the number of equivalent entities between datasets and reveal possible errors. The techniques of profiling developed during this work are automatic and can be applied to different datasets independently of the domain.

Over the last few years, there has been an increasing di↵usion of Linked Data as a standard way to publish interlinked structured data on the Web, which allows users, and public and private organizations to fully exploit a large amount of data from several domains that were not available in the past. Although gathering and publishing such massive amount of structured data is certainly a step in the right direction, quality still poses a significant obstacle to the uptake of data consumption applications at large-scale. A crucial aspect of quality regards the dynamic nature of Linked Data where information can change rapidly and fail to reflect changes in the real world, thus becoming out-date. Quality is characterised by di↵erent dimensions that capture several aspects of quality such as accuracy, currency, consistency or completeness. In particular, the aspects of Linked Data dynamicity are captured by Time-Related Quality Dimen- sions such as data currency. The assessment of Time-Related Quality Dimensions, which is the task of measuring the quality, is based on temporal information whose collection poses several challenges regarding their availability, representation and diversity in Linked Data. The assessment of Time-Related Quality Dimensions supports data consumers in their decisions whether information are valid or not. The main goal of this thesis is to develop techniques for assessing Time-Related Quality Dimensions in Linked Data, which must overcome several challenges posed by Linked Data such as third-party applications, variety of data, high volume of data or velocity of data. The major contributions of this thesis can be summarized as follows: it presents a general settings of definitions for quality dimensions and measures adopted in Linked Data; it provides a large-scale analysis of approaches for representing temporal information in Linked Data; it provides a sharable and interoperable conceptual model which integrates vocabularies used to represent temporal information required for the assessment of Time-Related Quality Di- mensions; it proposes two domain-independent techniques to assess data currency that work with incomplete or inaccurate temporal information and finally it pro- vides an approach that enrich information with time intervals representing their temporal validity.

Genetics and Molecular Biology are keys for the understanding the mechanisms of many of the human diseases that have strong harmful effects. The empirical mission of Genetics is to translate these mechanisms into Clinical benefits, thus bridging in-silico findings to patient bed side: approaching this goal means achieving what is commonly referred as clinical genomics or personalized medicine. In this process, technologies are assuming an increasing role. With the introduction of new experimental platforms (microarrays, sequencing, etc), today's analyses are much more detailed and can cover a wide spectrum of applications, from gene expression to Copy Number Variants detection. The advantages of technological improvements are usually followed by data management drawbacks due to the explosion of data throughput that reflects on a real need for new systems of data rationalization and management, data access, query and extraction. Our genetic laboratories partners encountered all those issues: what they need is a tool that allows data-integration and supports biological data analysis exploiting computational infrastructures on distributed environment. From such needs, we defined two main goals: (1) Computer Science goal: to design and implement a framework that integrates and manages data and genetic analyses; (2) Genetics and Molecular Biology goal (application domains): to solve biological problems through the framework and develop new methods. Given these requirements and related specifications, we designed an extensible framework based on three inter-connected layers: (1) Experimental data layer, that provides data integration of data from high-throughput platforms (also called horizontal data integration); (2) Knowledge data layer, that provides data integration of knowledge data (also called vertical integration); (3) Computational layer, that provides access to distributed environments for data analysis, in our cases GRID and Cluster technologies. Above the three design blocks, single biological problems can be supported and custom user interfaces are implemented. From our partner laboratories, two main relevant biological problems have been addressed: (1) Linkage Analysis: given a large pedigree in which subjects were genotyped with chips of 1 million of SNPs, the linkage analysis problem presented real computational limits. We designed a heuristic method to overcome computational restrictions and implemented it within our framework, exploiting GRID and Cluster environments. Using our approach, we obtained genetic results, successfully validated by end-users. We also tested performances of the system, reporting compared results. (2) SNP selection and ranking: given the problem of ranking SNPs based on a-priori information, we developed a novel method for biological data mining on genes' annotations. The method has been implemented as a web tool, SNP Ranker, that is under deep validation by our partners laboratories. The framework here designed and implemented demonstrated that this approach is consistent and can have potential impacts on the scientific community.