Search

The move towards preserving judgement disagreements in NLP requires the identification of adequate evaluation metrics. We identify a set of key properties that such metrics should have, and assess the extent to which natural candidates for soft evaluation such as Cross Entropy satisfy such properties. We employ a theoretical framework, supported by a visual approach, by practical examples, and by the analysis of a real case scenario. Our results indicate that Cross Entropy can result in fairly paradoxical results in some cases, whereas other measures Manhattan distance and Euclidean distance exhibit a more intuitive behavior, at least for the case of binary classification.

La logica Order-Sorted Feature (OSF) è un linguaggio di rappresentazione della conoscenza nato dal lavoro di ricerca di Hassan Aït-Kaci su un sistema logico per modellare le nozioni di sussunzione e unificazione nei formalismi basati sull'ereditarietà. La logica OSF è stata applicata alla linguistica computazionale e implementata in linguaggi di constraint logic programming e ragionatori automatici. Il linguaggio della logica OSF si basa su funzioni (attributi, o features), e su tipi (concetti, o sorts) ordinati in una relazione di sussunzione. In questa logica il ragionamento è fondato sull'unificazione di strutture chiamate termini OSF, un processo che mira a combinare i vincoli espressi da due termini OSF in un unico termine. Un vantaggio della logica OSF è che il suo algoritmo di unificazione tiene conto dell'ordinamento di sussunzione tra i tipi, che può consentire a una singola unificazione di sostituire diversi passi di inferenza, portando a calcoli più efficienti. Questa tesi si occupa dello sviluppo teorico del ragionamento approssimato con la logica OSF. Il primo contributo della tesi è la definizione della logica OSF fuzzy, una generalizzazione fuzzy della semantica della logica OSF in cui i tipi denotano insiemi fuzzy, consentendo di rappresentare concetti vaghi. Inoltre, i tipi della logica OSF fuzzy sono ordinati in una relazione di sussunzione fuzzy, fornendo maggiore flessibilità nella modellazione. La relazione di sussunzione fuzzy viene dotata di una semantica che generalizza la definizione di inclusione degli insiemi fuzzy di Zadeh. In questo lavoro indaghiamo se diverse proprietà semantiche e computazionali della logica OSF siano preservate nel contesto fuzzy. Dimostriamo, per esempio, che i termini OSF sono ordinati in un reticolo di sussunzione fuzzy che estende l'ordinamento fuzzy tra i tipi, e dimostriamo che l'unificazione di due termini OSF produce il loro estremo inferiore. Definiamo anche procedure per calcolare il grado di sus, Order-Sorted Feature (OSF) logic is a Knowledge Representation and Reasoning (KRR) language originating in Hassan Aït-Kaci's work on designing a calculus of partially ordered type structures. The language was developed to model the notions of subsumption and unification in inheritance-based KRR formalisms, and it has been applied in computational linguistics and implemented in constraint logic programming languages and automated reasoners. The language of OSF logic is based on function-denoting feature symbols and on set-denoting sort symbols ordered in a subsumption (is-a) lattice. Reasoning with OSF logic relies on the unification of set-denoting structures called OSF terms, a process that aims to combine the constraints expressed by two OSF terms into a single term. An advantage of OSF logic is that its unification algorithm takes into account the subsumption ordering between sorts, which may enable a single unification step to replace several inference steps, leading to more efficient computations. This thesis deals with the theoretical development of approximate reasoning within the framework of OSF logic. The first key contribution of the thesis is the definition of fuzzy OSF logic, a fuzzy generalization of the semantics of OSF logic where sorts denote fuzzy sets rather than crisp sets, allowing to represent vague concepts. Moreover, the sorts of fuzzy OSF logic are ordered in a fuzzy subsumption relation (formally a fuzzy lattice) rather than a crisp one, which provides more modeling flexibility by allowing to represent graded subsumption relations. The fuzzy sort subsumption relation is given a special semantics which generalizes Zadeh's definition of inclusion of fuzzy sets. We investigate whether several semantic and computational properties of crisp OSF logic are preserved in the fuzzy setting. For instance, we show that OSF terms are ordered in a fuzzy subsumption relation which extends the fuzzy ordering between sorts, and we prove that the unificatio

The large availability of processable textual resources for Classical Latin has made it possible to study Latin literature through methods and tools that support distant reading. This paper describes a number of experiments carried out to test the possibility of investigating the thematic distribution of the Classical Latin corpus Opera Latina by means of topic modeling. For this purpose, we train, optimize and compare two neural models, Product-of-Experts LDA (ProdLDA) and Embedded Topic Model (ETM), opportunely revised to deal with the textual data from a Classical Latin corpus, to evaluate which one performs better both on the basis of topic diversity and topic coherence metrics, and from a human judgment point of view. Our results show that the topics extracted by neural models are coherent and interpretable and that they are significant from the perspective of a Latin scholar. The source code of the proposed model is available at https://github.com/MIND-Lab/LatinProdLDA.

La conoscenza testuale è uno dei pilastri principali della nostra società. Infatti, la conoscenza umana è spesso trasmessa attraverso le parole. Dall'invenzione della scrittura, gli uomini hanno raccontato e descritto la loro esistenza con parole su pezzi di carta. Questa quantità di conoscenza si accumula fino a ciò che l'intera civiltà ha raccolto in più di 5000 anni. Gli storici e gli scienziati sociali e politici cercano modi per capire meglio questa vasta quantità di conoscenza collettiva che non può essere esplorata manualmente. A questo scopo, i ricercatori di machine learning, statistica e linguistica computazionale hanno sviluppato i topic model, una suite di algoritmi che mirano ad annotare grandi archivi di documenti con informazioni tematiche. La popolarità di questi modelli è dovuta al fatto che sono non supervisionati e che sono interpretabili. I topic model analizzano e riassumono i temi principali, anche detti topic, di grandi collezioni di documenti, presentando le informazioni in una forma compatta e comprensibile. La maggior parte dei topic model si concentrano solo sulle parole codificate nei documenti. Tuttavia, informazioni aggiuntive possono essere introdotte nei topic model per migliorare le loro prestazioni. Infatti, in molti casi del mondo reale, raramente abbiamo solo i semplici testi da analizzare. Invece, abbiamo informazioni aggiuntive o metadati relativi ai documenti, ad esempio, l'autore del documento, la data, collegamenti ipertestuali ad altri documenti, un insieme di hashtag, menzioni o etichette. Possiamo usare queste informazioni precedenti per aiutare un topic model a scoprire topic di qualità. Per esempio, sapere che un documento cita un altro documento aumenta la nostra fiducia che i documenti parlano degli stessi argomenti.Inoltre, i topic model spesso ignorano l'ordine delle parole e le informazioni contestuali, rendendo difficile dedurre gli topic coerenti e significativi. Un altro problema nel campo è legato a, Textual knowledge is one of the main pillars of our society. Indeed, human knowledge is often passed along using words. Since the invention of writing, humans have narrated and described their existence with words over pieces of papers. This amount of knowledge builds up to what the entire civilization has collected over more than 5'000 years. Historians and social and political scientists look for ways to understand better this vast amount of collective knowledge that cannot be manually explored. To this end, researchers from machine learning, statistics and computational linguistic have developed topic models, a suite of algorithms that aim to annotate large archives of documents with thematic information. The popularity of these models is due to the fact that they are unsupervised and that they are interpretable. Topic models analyze and summarize the main themes, or topics, of large collections of documents, presenting the information in a compact and understandable form. Most topic models focus only on the words encoded in the documents. However, additional information can be introduced into topic models to improve their performance. In fact, in many real-world cases, we seldom have only the mere texts to analyze. Instead, we have additional information or metadata related to the documents, e.g., the document's author, the date, hyperlinks to other documents, a set of hashtags, mentions or labels. We can use this prior information to help a topic model discover better topics. For example, knowing that a document cites another document increases our confidence that the documents talk about the same topics. Also, topic models often ignore word order and contextual information, making it difficult to infer high-quality topics. Another problem in the field is related to the hyperparameters used to train the topics models. The hyperparameters are often fixed in experimental settings and few researchers have tried to study their impact on the results. In this thesis

Proceedings of the Eighth Italian Conference on Computational Linguistics (CLiC-it 2021). Milan, Italy, January 26-28, 2022

Preface of the volume

In this paper we present a set of annotated data and the results of a number of unsupervised experiments for the analysis of sentiment in Latin poetry. More specifically, we describe a small gold standard made of eight poems by Horace, in which each sentence is labeled manually for the sentiment using a four-value classification (positive, negative, neutral and mixed). Then, we report on how this gold standard has been used to evaluate two automatic approaches for sentiment classification: one is lexicon-based and the other adopts a zero-shot transfer approach.

The Liber Abbaci (13th century) is a milestone in the history of mathematics and accounting. Due to the late stage of Latin, its features and its very specialized content, it also represents a unique resource for scholars working on Latin corpora. In this paper we present the annotation and linking work carried out in the frame of the project Fibonacci 1202-2021. A gold-standard lemmatization and part-ofspeech tagging allow us to elaborate some first observations on the linguistic and historical features of the text, and to link the text to the Lila Knowledge Base, that has as its goal to make distributed linguistic resources for Latin interoperable by following the principles of the Linked Data paradigm. Starting from this specific case, we discuss the importance of annotating and linking scientific and technical texts, in order to (a) compare and search them together with other (non-technical) Latin texts (b) train, apply and evaluate NLP resources on a non-standard variety of Latin. The paper also describes the fruitful interaction and coordination between NLP experts and traditional Latin scholars on a project requiring a large range of expertise.

This paper describes the steps taken to include data from the Lewis & Short bilingual Latin-English dictionary into the Knowledge Base of linguistic resources for Latin LiLa. First, data were extracted from the original XML and matched with entries in LiLa, overcoming ambiguities and structural inconsistencies in the source. Subsequently, senses were modelled using the Ontolex Lemon Lexicographic module (lexicog), so that they could be included in the LiLa Knowledge Base and thus made interoperable with the (meta)data of the linguistic resources for Latin therein interlinked.

Relational topic models (RTM) have been widely used to discover hidden topics in a collection of networked documents. In this paper, we introduce the class of Constrained Relational Topic Models (CRTM), a semi-supervised extension of RTM that, apart from modeling the structure of the document network, explicitly models some available domain knowledge. We propose two instances of CRTM that incorporate prior knowledge in the form of document constraints. The models smooth the probability distribution of topics such that two constrained documents can either share the same topics or denote distinct themes. Experimental results on benchmark relational datasets show significant performances of CRTM on a semi-supervised document classification task.

In this paper we deal with complex attributed graphs which can exhibit rich connectivity patterns and whose nodes are often associated with attributes, such as text or images. In order to analyze these graphs, the primary challenge is to find an effective way to represent them by preserving both structural properties and node attribute information. To create low-dimensional and meaningful embedded representations of these complex graphs, we propose a fully unsupervised model based on Deep Learning architectures, called Constrained Attributed Graph Embedding model (CAGE). The main contribution of the proposed model is the definition of a novel two-phase optimization problem that explicitly models node attributes to obtain a higher representation expressiveness while preserving the local and the global structural properties of the graph. We validated our approach on two different benchmark datasets for node classification. Experimental results demonstrate that this novel representation provides significant improvements compared to state of the art approaches, also showing higher robustness with respect to the size of the training data.

The aim of Sentiment Analysis is to define automatic tools able to extract subjective information from texts in natural language, such as opinions and sentiments, in order to create structured and actionable knowledge to be used by either a decision support system or a decision maker. Sentiment analysis has gained even more value with the advent and growth of social networking. Sentiment Analysis in Social Networks begins with an overview of the latest research trends in the field. It then discusses the sociological and psychological processes underling social network interactions. The book explores both semantic and machine learning models and methods that address context-dependent and dynamic text in online social networks, showing how social network streams pose numerous challenges due to their large-scale, short, noisy, context- dependent and dynamic nature. Further, this volume: Takes an interdisciplinary approach from a number of computing domains, including natural language processing, machine learning, big data, and statistical methodologiesProvides insights into opinion spamming, reasoning, and social network analysisShows how to apply sentiment analysis tools for a particular application and domain, and how to get the best results for understanding the consequencesServes as a one-stop reference for the state-of-the-art in social media analyticsTakes an interdisciplinary approach from a number of computing domains, including natural language processing, big data, and statistical methodologiesProvides insights into opinion spamming, reasoning, and social network miningShows how to apply opinion mining tools for a particular application and domain, and how to get the best results for understanding the consequencesServes as a one-stop reference for the state-of-the-art in social media analytics

La mole di dati generata dagli utenti sul Web è esponenzialmente cresciuta negli ultimi dieci anni, creando nuove e rilevanti opportunità per ogni tipo di dominio applicativo. Per risolvere i problemi derivanti dall’eccessiva quantità di dati, la ricerca nell’ambito dell’elaborazione del linguaggio naturale si è mossa verso lo sviluppo di modelli computazionali capaci di capirlo ed interpretarlo senza (o quasi) alcun intervento umano. Recentemente, questo campo di studi è stato testimone di un incremento sia in termini di efficienza computazionale che di risultati, per merito dell’avvento di una nuova linea di ricerca nell’apprendimento automatico chiamata Deep Learning. Questa tesi si focalizza in modo particolare su una specifica classe di modelli di Deep Learning atta ad apprendere rappresentazioni di alto livello, e conseguentemente più significative, dei dati di input in ambiente non supervisionato. Nelle tecniche di Deep Learning, queste rappresentazioni sono ottenute tramite multiple trasformazioni non lineari di complessità e astrazione crescente a partire dai dati di input. Questa fase, in cui vengono elaborate le sopracitate rappresentazioni, è un processo cruciale per l’elaborazione del linguaggio naturale in quanto include la procedura di trasformazione da simboli discreti (es. lettere) a una rappresentazione vettoriale che può essere facilmente trattata da un elaboratore. Inoltre, questa rappresentazione deve anche essere in grado di codificare la sintattica e la semantica espressa nel linguaggio utilizzato nei dati. La prima direzione di ricerca di questa tesi mira ad evidenziare come i modelli di elaborazione del linguaggio naturale possano essere potenziati dalle rappresentazioni ottenute con metodi non supervisionati di Deep Learning al fine di conferire un senso agli ingenti contenuti generati dagli utenti. Nello specifico, questa tesi si focalizza su diversi ambiti che sono considerati cruciali per capire di cosa il testo tratti (Named En, The huge amount of textual user-generated content on the Web has incredibly grown in the last decade, creating new relevant opportunities for different real-world applications and domains. To overcome the difficulties of dealing with this large volume of unstructured data, the research field of Natural Language Processing has provided efficient solutions developing computational models able to understand and interpret human natural language without any (or almost any) human intervention. This field has gained in further computational efficiency and performance from the advent of the recent machine learning research lines concerned with Deep Learning. In particular, this thesis focuses on a specific class of Deep Learning models devoted to learning high-level and meaningful representations of input data in unsupervised settings, by computing multiple non-linear transformations of increasing complexity and abstraction. Indeed, learning expressive representations from the data is a crucial step in Natural Language Processing, because it involves the transformation from discrete symbols (e.g. characters) to a machine-readable representation as real-valued vectors, which should encode semantic and syntactic meanings of the language units. The first research direction of this thesis is aimed at giving evidence that enhancing Natural Language Processing models with representations obtained by unsupervised Deep Learning models can significantly improve the computational abilities of making sense of large volume of user-generated text. In particular, this thesis addresses tasks that were considered crucial for understanding what the text is talking about, by extracting and disambiguating the named entities (Named Entity Recognition and Linking), and which opinion the user is expressing, dealing also with irony (Sentiment Analysis and Irony Detection). For each task, this thesis proposes a novel Natural Language Processing model enhanced by the data representation obtained by

This article presents a comparison of different Word Sense Induction (wsi) clustering algorithms on two novel pseudoword data sets of semantic-similarity and co-occurrence-based word graphs, with a special focus on the detection of homonymic polysemy. We follow the original definition of a pseudoword as the combination of two monosemous terms and their contexts to simulate a polysemous word. The evaluation is performed comparing the algorithm’s output on a pseudoword’s ego word graph (i.e., a graph that represents the pseudoword’s context in the corpus) with the known subdivision given by the components corresponding to the monosemous source words forming the pseudoword. The main contribution of this article is to present a self-sufficient pseudoword-based evaluation framework for wsi graph-based clustering algorithms, thereby defining a new evaluation measure (top2) and a secondary clustering process (hyperclustering). To our knowledge, we are the first to conduct and discuss a large-scale systematic pseudoword evaluation targeting the induction of coarse-grained homonymous word senses across a large number of graph clustering algorithms.

This article presents a comparison of different Word Sense Induction (wsi) clustering algorithms on two novel pseudoword data sets of semantic-similarity and co-occurrence-based word graphs, with a special focus on the detection of homonymic polysemy. We follow the original definition of a pseudoword as the combination of two monosemous terms and their contexts to simulate a polysemous word. The evaluation is performed comparing the algorithm’s output on a pseudoword’s ego word graph (i.e., a graph that represents the pseudoword’s context in the corpus) with the known subdivision given by the components corresponding to the monosemous source words forming the pseudoword. The main contribution of this article is to present a self-sufficient pseudoword-based evaluation framework for wsi graph-based clustering algorithms, thereby defining a new evaluation measure (top2) and a secondary clustering process (hyperclustering). To our knowledge, we are the first to conduct and discuss a large-scale systematic pseudoword evaluation targeting the induction of coarse-grained homonymous word senses across a large number of graph clustering algorithms.

Knowledge Graphs (KG) represent a large amount of Semantic Associations (SAs), i.e., chains of relations that may reveal interesting and unknown connections between different types of entities. Applications for the contextual exploration of KGs help users explore information extracted from a KG, including SAs, while they are reading an input text. Because of the large number of SAs that can be extracted from a text, a first challenge in these applications is to effectively determine which SAs are most interesting to the users, defining a suitable ranking function over SAs. However, since different users may have different interests, an additional challenge is to personalize this ranking function to match individual users’ preferences. In this paper we introduce a novel active learning to rank model to let a user rate small samples of SAs, which are used to iteratively learn a personalized ranking function. Experiments conducted with two data sets show that the approach is able to improve the quality of the ranking function with a limited number of user interactions.

Information Extraction (IE) is a process focused on automatic extraction of structured information from unstructured text sources. One open research field of IE relates to Named Entity Recognition (NER), aimed at identifying and associating atomic elements in a given text to a predefined category such as names of persons, organizations, locations and so on. This problem can be formalized as the assignment of a finite sequence of semantic labels to a set of interdependent variables associated with text fragments, and can modelled through a stochastic process involving both hidden variables (semantic labels) and observed variables (textual cues). In this work we investigate one of the most promising model for NER based on Conditional Random Fields (CRFs). CRFs are enhanced in a two stages approach to include in the decision process logic rules that can be either extracted from data or defined by domain experts. The problem is defined as a Resource Constrained Maximum Path Problem (RCMPP) associating a resource with each logic rule. Proper resource Extension Functions (REFs) and upper bound on the resource consumptions are defined in order to model the logic rules as knapsack-like constraints. A well-tailored dynamic programming procedure is defined to address the RCMPP

Sentiment Analysis is a broad task that involves the analysis of various aspect of the natural language text. However, most of the approaches in the state of the art usually investigate independently each aspect, i.e. Subjectivity Classification, Sentiment Polarity Classification, Emotion Recognition, Irony Detection. In this paper we present a Multi-View Sentiment Corpus (MVSC), which comprises 3000 English microblog posts related the movie domain. Three independent annotators manually labelled MVSC, following a broad annotation schema about different aspects that can be grasped from natural language text coming from social networks. The contribution is therefore a corpus that comprises five different views for each message, i.e. subjective/objective, sentiment polarity, implicit/explicit, irony, emotion. In order to allow a more detailed investigation on the human labelling behaviour, we provide the annotations of each human annotator involved.

La presente tesi si occupa dell’induzione dei significati delle parole (ISP o word sense induction - WSI), una branca dell’elaborazione dei linguaggi naturali il cui scopo è individuare ed elencare automaticamente e in modo non supervisionato i possibili significati o sensi che un termine può assumere relativamente ai differenti contesti in cui esso viene impiegato, senza ricorrere a risorse esterne quali dizionari od ontologie. Fra i vari approcci alla ISP esistenti, ci siamo concentrati in particolare su quelli che modellano il contesto di una parola tramite un grafo, su cui viene fatto operare un algoritmo per partizionarlo: la partizione che ne risulta ha come interpretazione l’implicita descrizione dei possibili sensi di quella parola. Le nozioni fondamentali dell’ISP, alcuni concetti base e approcci alla ISP scelti fra la letteratura del campo vengono presentati e analizzati nella prima parte del lavoro. Nella seconda parte introduciamo il nostro triplice contributo. Per prima cosa definiamo ed esploriamo la distanza di Jaccard pesata (assieme a una sua versione non pesata), cioè una distanza sui nodi di un grafo non orientato con pesi positivi, che usiamo per ottenere relazioni di secondo ordine dalle relazioni di primo ordine eventualmente modellate dal grafo (p.e. cooccorrenze). Inoltre, definiamo la nozione correlata di “arco passerella”, un arco “separatore” il cui peso è superiore alla media dei pesi sugli archi uscenti da una delle sue due estremità, e definiamo anche una nuova interpretazione sintetica delle curvatura di un grafo, vista come la differenza delle distanze di Jaccard pesata e non pesata sugli archi. La distanza di Jaccard da noi definita è alla base del nostro secondo contributo: tre nuovi algoritmi di clustering su grafo espressamente creati per l’ISP, rispettivamente l’algoritmo di clustering per passerelle, un algoritmo di clustering aggregativo e uno basato sulla curvatura. Il terzo contributo di questa tesi è un nuovo sistema, This dissertation is about Word Sense Induction (WSI), a branch of Natural Language Processing concerned with the automated, unsupervised detection and listing of the possible senses that a word can assume relative to the different contexts in which it appears. To this end, no external resources like dictionaries or ontologies are used. Among the many existing approaches to WSI, we focus specifically on modelling the context of a word through a graph and on running a clustering algorithm on it: the resulting clusters are interpreted as implicitly describing the possible senses of the word. Fundamental notions of WSI, basic concepts and some WSI approaches selected from literature are presented and examined in the first part of this work. In the second part, we introduce our threefold contribution. Firstly, we define and explore a weighted (together with an unweighted) Jaccard distance, i.e. a distance on the nodes of a positively weighted undirected graph which we use to obtain second-order relations from the first-order ones modelled by the graph (e.g. co-occurrences). Moreover, we define the related notion of gangplank edge, a separator edge with weight greater than the mean weights of the edges incident to either of its two ends, and finally a new synthetic interpretation of the curvature on a graph, seen as the difference between weighted and unweighted Jaccard distances between node couples. Our Jaccard distance is at the basis of the second contribution: three novel graph-based clustering algorithms expressly created for the task of WSI, respectively the gangplank clustering algorithm, an aggregative clustering algorithm and a curvature-based clustering algorithm. The third contribution is a novel evaluation framework for graph-based clustering algorithms for WSI, consisting of two word graph data sets (one for co-occurrences and one for semantic similarities) and a new ad hoc evaluation measure built around pseudowords. A pseudoword is the artificial conflati

L'estrazione di informazioni strutturate a partire dal “web non strutturato”, ha suscitato un notevole interesse da parte delle comunità scientifiche che si occupano di elaborazione del linguaggio naturale e di sistemi basati sulla conoscenza per sviluppare a pieno la visione del “web semantico”. Nell'era moderna, l'uso pervasivo e diffuso delle reti sociali ha portato alla produzione di un flusso continuo di informazioni su piattaforme quali Twitter o Facebook, definite anche piattaforme di microblogging. Tali sorgenti informative, accessibili in tempo reale, producono informazioni caratterizzate dalla presenza costante di rumore e ambiguità linguistiche che rendono particolarmente difficoltoso il compito di estrarre informazioni strutturate. Tale estrazione è tuttavia cruciale per poter arricchire grandi basi di conoscenza, oggi usate in molte applicazioni industriali e di ricerca, con informazioni nuove e rilevanti. Ne risulta che nell'ultimo decennio sono aumentati significativamente gli sforzi della ricerca nel campo dell’elaborazione del linguaggio naturale per l'estrazione di informazioni da piattaforme di microblogging, con particolare attenzione nei confronti dell’estrazione e identificazione di entità nominali (anche Named Entity Extraction and Linking o NEEL). Oggigiorno esistono numerosi sistemi di NEEL, di cui la maggior parte però creati a scopo commerciale. La calibrazione dei componenti di un sistema di NEEL, cioè dei componenti per la rilevazione, la disambiguazione e l'identificazione di entità nominali, nel caso di piattaforme di microblogging come Twitter e Facebook è difficile in particolare a causa delle tipologie di testo considerato. Mancano approcci di ricerca sistematici volti a guidare l'utilizzo e il miglioramento di tali componenti, per la realizzazione di sistemi più robusti, in grado di meglio adattarsi all’emergere di nuove informazioni, e nuovi interessi (ad esempio, a estrarre tipi di entità nuovi rispetto a quelli conside, Natural Language Processing and Knowledge Base Experts are actively involved in extracting structured information from the Unstructured Web in order to realize the Semantic Web Vision. Diverse forms of unstructured information is easily available today to research scientists from social media platforms such as Twitter and Facebook in real time. %Knowledge discovery from the diverse forms of unstructured information available today from social media platforms such as Twitter and Facebook in real time, thus, plays a key role for this goal. The comprehensive and widespread use of such platforms in the modern age has led to a continuous stream of evolving information along with a constant presence of noise, and ambiguity which makes the task of extracting structured information difficult. An essential step is therefore identification of relevant information from the point of view of knowledge base enrichment. As a result, research efforts towards Information Extraction and Natural Language Processing Frameworks have increased significantly over the past decade, Named Entity Extraction and Linking (NEEL) Frameworks being one of the very prevalent ones. Numerous NEEL frameworks exist today, however, mostly for commercial purposes. The orchestration of components of a NEEL framework, i.e., named entity recognition component, named entity disambiguation and named entity linking component, for microblogging platforms such as Twitter and Facebook is difficult in particular due to the type of text under consideration. As a result, there is little research in the use and improvement of such components towards a more robust framework that can be adapted to emerging information in real time. This thesis discusses the challenges faced by conventional NEEL frameworks when faced with textual formats such as tweets and investigates several approaches to improve the performance of the components and of the NEEL framework as a whole. A key hypothesis is that the performance of such a f

In the recent years, the amount of user generated contents shared on the Web has significantly increased, especially in social media environment, e.g. Twitter, Facebook, Google+. This large quantity of data has generated the need of reactive and sophisticated systems for capturing and understanding the underlying information enclosed in them. In this paper we present TWINE, a real-time system for the big data analysis and exploration of information extracted from Twitter streams. The proposed system based on a Named Entity Recognition and Linking pipeline and a multi-dimensional spatial geo-localization is managed by a scalable and flexible architecture for an interactive visualization of micropost streams insights. The demo is available at http://twinemind. cloudapp.net/streaming1,2.

Numerous state-of-the-art Named Entity Recognition (NER) systems use different classification schemas/ontologies. Comparisons and integration among NER systems, thus, becomes complex. In this paper, we propose a transfer-learning approach where we use supervised learning methods to automatically learn mappings between ontologies of NER systems, where an input probability distribution over a set of entity types defined in a source ontology is mapped to a target distribution over the entity types defined for a target ontology. Experiments conducted with benchmark data show valuable re-classification performance of entity mentions, suggesting our approach as a promising one for domain adaptation of NER systems.

Real world applications of machine learning in natural language processing can span many different domains and usually require a huge effort for the annotation of domain specific training data. For this reason, domain adaptation techniques have gained a lot of attention in the last years. In order to derive an effective domain adaptation, a good feature representation across domains is crucial as well as the generalisation ability of the predictive model. In this paper we address the problem of domain adaptation for sentiment classification by combining deep learning, for acquiring a cross-domain high-level feature representation, and ensemble methods, for reducing the cross-domain generalization error. The proposed adaptation framework has been evaluated on a benchmark dataset composed of reviews of four different Amazon category of products, significantly outperforming the state of the art methods.