Search

This article presents an operational system for the automatic production of high-resolution (HR) large-scale land cover (LC) maps in a fast, efficient, and unsupervised manner. This is based on a scalable and parallelizable tile-based approach, which does not require the collection of new training data. The method leverages the complementary information provided by the existing LC maps and recent acquisitions of HR Earth observation (EO) images to identify map units that have the highest probability of being correctly associated with their labels, and exploit the obtained “weak” training set to produce an updated HR LC map by classifying the recently acquired EO data. Both steps, performed at tile level, can be implemented on a high-performance computing (HPC) environment, which simultaneously process all required tiles (independently of each other) for the entire study area. The method was tested considering the publicly available 2018 Corine LC map having a minimum mapping unit of 25 ha and the Sentinel-2 images to generate an HR LC map of Italy. The obtained map has a spatial resolution of 10 m and presents the nine major LC types (i.e., “artificial land,” “bareland,” “grassland,” “cropland,” “broadleaves,” “conifers,” “snow,” “water,” and “shrubland”). Validation was performed using the 2018 European Land Use and Coverage Area Frame Survey database made up of in situ data. The overall accuracy achieved for the Northern, Southern, and Central part of Italy and the Italian Islands is 91.29%, 91.63%, 92.21%, and 91.06%, respectively.

Η έκταση, η πιθανότητα και η πολυπλοκότητα της βιομηχανίας υγειονομικής περίθαλψης έχει προσελκύσει εκτεταμένη απάτη που έχει σημαντικό αντίκτυπο στην οικονομία. Οι δόλιες δραστηριότητες όχι μόνο συμβάλλουν στην αύξηση του προβλήματος των δαπανών της υγειονομικής περίθαλψης, αλλά επηρεάζουν επίσης την υγεία των ασθενών. Η πρόκληση στα τρέχοντα συστήματα εντοπισμού απάτης έγκειται κυρίως στην κατανόηση της επιβάρυνσης της χρηματικής απώλειας και των ασυνήθιστων συμπεριφορών. Παρά την εφαρμογή διαφόρων τεχνολογιών και στρατηγικών για την καταπολέμηση της απάτης, όπως προγραμματισμένοι, στοχευμένοι και τυχαίοι έλεγχοι, καταγγελίες και βιομετρικά συστήματα, η απάτη στην ασφάλιση συνεχίζει να αποτελεί μεγάλο πρόβλημα για τους περισσότερους παρόχους ασφάλισης υγείας. Σκοπός της παρούσας διπλωματικής εργασίας είναι να ευαισθητοποιήσει σχετικά με αυτόν τον σημαντικό τομέα το ευρύτερο κοινό αλλά και να καταδείξει την αναγκαιότητα για τους σχετικούς δημόσιους οργανισμούς και τις εταιρείες του χώρου της υγείας να επενδύσουν ένα σημαντικό μέρος του κεφαλαίου τους αλλά και του χρόνου τους στη σοφή αξιοποίηση των δεδομένων, τα οποία διατίθενται πλέον στον κόσμο, στην στατιστική μηχανική μάθηση, καθώς και σε τεχνικές προβλεπτικής αναλυτικής με σκοπό την μείωση της απάτης στην ασφάλιση υγείας. Αρχικά, δίνονται οι σχετικοί ορισμοί για την απάτη στην ασφάλισης υγείας καθώς και μία επισκόπηση του προβλήματος και των τρόπων και μορφών με τις οποίες αυτή εμφανίζεται. Στη συνέχεια, υπογραμμίζεται με διάφορα παραδείγματα απάτης στον τομέα της υγειονομικής περίθαλψης πόσο σημαντικό είναι το πρόβλημα. Επιπλέον, παρατίθενται και άλλοι τομείς εντοπισμού απάτης, οι οποίοι υποστηρίζονται με διάφορες μελέτες περιπτώσεων. Στη συνέχεια γίνεται μία διεξοδική αναζήτηση στη βιβλιογραφία για τις εφαρμογές της ανάλυσης δεδομένων και της στατιστικής μηχανικής μάθησης που έχουν χρησιμοποιηθεί μαζί με μελέτες περιπτώσεων. Έπειτα, αναφέρονται οι συνήθεις μέθοδοι μηχανικής μάθησης στον εντοπισμό της απάτης. Τέλος, επιλεγμένες μεθοδολογίες εφαρμόστηκαν σε πραγματικά δεδομένα προκειμένου να παρουσιαστεί μια ολοκληρωμένη μελέτη περίπτωσης με στόχο τη δημιουργία ενός μοντέλου που μπορεί να εντοπίσει με χαμηλότερο κόστος και δαπάνες δόλιους παρόχους στην ασφάλιση υγείας., The extent, probability, and complexity of the healthcare industry have attracted widespread fraud that has a significant impact on the economy; fraudulent activities not only contribute to the problem of increasing health care costs but also affect patient health; the challenge in current fraud detection systems lies mainly in understanding the burden of financial loss and unusual areas of behavior. Despite the implementation of various anti-fraud technologies and strategies, such as planned, targeted and random checks, complaints, and biometric systems, insurance fraud is still a major problem for most health insurance providers. The purpose of this thesis is to raise awareness of this important area among the general public and to demonstrate the need for relevant public bodies and health companies to invest a significant part of their capital and their time in the wise exploitation of the data now available in the world, statistical machine learning, as well as predictive analysis techniques to reduce health insurance fraud. Initially, the relevant definitions for health insurance fraud are given, and it is provided an overview of the problem and the ways and forms in which health insurance fraud occurs. Next, it is highlighted how important the problem is with various examples of healthcare fraud worldwide, and a presentation of punishments for health insurance fraud. In addition, other fraud detection areas are listed, supported by case studies. Then there will be a thorough search in the literature for the applications of data analytics and statistical machine learning that have been used along with case studies. At a later stage, the methods of machine learning in fraud detection that are commonly used are mentioned. Finally, selected methodologies were applied to real-world data in order to present a complete case study that aims to build a model that can detect fraudulent providers with lower costs and expenditures.

International audience; For a solar adaptive optics system equipped with a Shack-Hartmann wavefront sensor, the local wavefront slopes are measured from the displacements of the images produced by the sub-pupils of the sensor with respect to a given reference image. Measuring these displacements is challenging because of the very low contrast (at most a few percent) of the structures at the Sun surface. Additional difficulties arise from the fact that these structures evolve in time and are slightly distorted in each sub-image. In this contribution, we describe a novel approach to process the images of a solar wavefront sensor which jointly estimates the wavefront slopes, their noise covariance matrix, and the reference image. Spatio-temporal constraints are imposed on the reference image to regularize the problem and stabilize the global tip-tilt. Automatically tuned correction factors are introduced to account for the scintillation and the local distortions. Our method yields a sufficient statistic which enables an optimal wavefront reconstruction. We propose an alternating strategy to quickly solve the joint estimation problem. Special attention has been paid to make the numerical algorithm usable in real-time. Our method is implemented in the adaptive optics system of the THEMIS solar telescope equipped with a 10 × 10 Shack-Hartmann wavefront sensor delivering 400 × 400 pixel images at 1kHz. On a single CPU core, the Julia version of our algorithm provides the measurements with 90μs of latency after the image acquisition and takes an additional 200µs to update the reference image.

Anomaly Detection (AD) is to identify samples that differ from training observations in some way. Those samples that do not follow the distribution of normal data are called outliers or anomalies. In this thesis, we examined two different challenges related to deep learning-based anomaly detection methods. The first challenge is the generalizability to outliers. A wide range of unsupervised anomaly detection methods use deep autoencoders as a foundation. However, a notable limitation of deep autoencoders is that they generalize to outliers and reconstruct them with low error. In order to overcome this issue, we propose an adversarial framework consisting of two competing components, an adversarial distorter, and an autoencoder. During training, the adversarial distorter produces perturbations that are applied to the encoder’s latent space to maximize the reconstruction error. The autoencoder attempts to neutralize the effects of these perturbations to minimize the reconstruction error. Another challenge is the high computational cost, complexity, and unstable training procedures of deep anomaly detection methods. Despite being successful at anomaly detection, deep neural networks are difficult to deploy in real-world applications because of this challenge. We overcome this problem by using a simple learning procedure that trains a lightweight convolutional neural network. We propose to solve anomaly detection as a supervised regression problem. We label normal and anomalous data using two separable distributions of continuous values. As a way to compensate for the lack of anomalous samples during training, we use straightforward image augmentation techniques to create a distinct set of anomalous samples. An augmented set has a distribution that is similar to normal data, but deviates slightly from it, while real anomalies should have a further distribution. Consequently, training a regressor on normal and these augmented samples will result in more distinct distributions of labels for normal and real anomalous data points. In several image and video anomaly detection benchmarks, our methods outperform cutting-edge approaches.

In this paper, we propose a novel unsupervised approach for sequence matching by explicitly accounting for the locality properties in the sequences. In contrast to conventional approaches that rely on frame-to-frame matching, we conduct matching using sequencelet or seqlet , a sub-sequence wherein the frames share strong similarities and are thus grouped together. The optimal seqlets and matching between them are learned jointly, without any supervision from users. The learned seqlets preserve the locality information at the scale of interest and resolve the ambiguities during matching, which are omitted by frame-based matching methods. We show that our proposed approach outperforms the state-of-the-art ones on datasets of different domains including human actions, facial expressions, speech, and character strokes.

High annotation costs are a substantial bottleneck in applying deep learning architectures to clinically relevant use cases, substantiating the need for algorithms to learn from unlabeled data. In this work, we propose employing self-supervised methods. To that end, we trained with three self-supervised algorithms on a large corpus of unlabeled dental images, which contained 38K bitewing radiographs (BWRs). We then applied the learned neural network representations on tooth-level dental caries classification, for which we utilized labels extracted from electronic health records (EHRs). Finally, a holdout test-set was established, which consisted of 343 BWRs and was annotated by three dental professionals and approved by a senior dentist. This test-set was used to evaluate the fine-tuned caries classification models. Our experimental results demonstrate the obtained gains by pretraining models using self-supervised algorithms. These include improved caries classification performance (6 p.p. increase in sensitivity) and, most importantly, improved label-efficiency. In other words, the resulting models can be fine-tuned using few labels (annotations). Our results show that using as few as 18 annotations can produce ≥45% sensitivity, which is comparable to human-level diagnostic performance. This study shows that self-supervision can provide gains in medical image analysis, particularly when obtaining labels is costly and expensive.

The hippocampus and amygdala are functionally coupled brain regions that play a crucial role in processes involving memory and learning. Because inter-areal communication has been reported both during specific sleep stages and in awake, behaving animals, these brain regions can serve as an archetype to establish that measuring functional interactions is important for comprehending neural systems. To this end, we analyze here a public data set of local field potentials (LFPs) recorded in rats simultaneously from the hippocampus and amygdala during different behaviors. Employing a specific, time-lagged embedding technique, named topological causality, we infer directed interactions between the LFP band powers of the two regions across six frequency bands in a time-resolved manner. The combined power and interaction signals are processed with our own unsupervised tools developed originally for the analysis of molecular dynamics simulations to effectively visualize and identify putative, internal states that are visited by the animals repeatedly. Our proposed methodology minimizes impositions onto the data, such as isolating specific epochs, or averaging across externally annotated behavioral stages, and succeeds in separating internal states by external labels such as sleep or stimulus events. We show that this works better for 2 of the 3 rats we analyzed, and highlight the need to acknowledge individuality in analyses of this type. Importantly, we demonstrate that the quantification of functional interactions is a significant factor in discriminating these external labels, and we suggest our methodology as a general tool for large, multi-site recordings. Significance statement We develop an analysis pipeline for neuroscience data sets. We test it on a published example of multi-electrode recordings of rats in a range of behaviors: running on a track, sleeping, collecting rewards, etc. We adopt nonlinear analysis techniques that are able to quantify directed interactions between different signals, here oscillations of two brain regions in different frequency bands. Using the entire recordings and, importantly, distinguishing each animal, we provide a high-resolution overview of the functional interplay of the two regions. Putative neural states that the animals can be in are derived from a time-aware clustering of the large data sets. When discriminating experimental annotations like run speed, we provide evidence that our methodology outperforms common clustering techniques.

Ponencia presentada en la 2021 International Conference on Indoor Positioning and Indoor Navigation (IPIN), 29 Nov.-2 Dec. 2021, Lloret de Mar (Spain) This paper presents a set of unsupervised methods based on Bluetooth Low Energy (BLE) technology to obtain the position and transitions that users perform between the zones where they usually stay longer in their own homes. In particular, two different methods are studied to detect the user's position, the first based on proximity and the second based on fingerprinting techniques with self-training. The proposed methodology allows a very easy-to-use deployment by the user and with a very low economic cost. A set of experiments have been carried out to assess the performance of the proposed techniques presented in this paper, where a real user has captured data for several days at home. The results obtained show that the fingerprinting-based method is the most effective to accurately detect the user's position and the transitions performed between the different areas of interest at home.

Supervised classification algorithms require a sufficiently large set of representative training samples to generate accurate land-cover maps. Collecting reference data is difficult, expensive, and unfeasible at the large scale. To solve this problem, this article introduces a novel approach that aims to extract reliable labeled data from existing thematic products. Although these products represent a potentially useful information source, their use is not straightforward. They are not completely reliable since they may present classification errors. They are typically aggregated at polygon level, where polygons do not necessarily correspond to homogeneous areas. Finally, usually, there is a semantic gap between map legends and remote sensing (RS) data. In this context, we propose an approach that aims to: 1) perform a domain understanding to detect the discrepancies between the thematic map domain and the RS data domain; 2) use RS data contemporary to the map to decompose the thematic product from the semantic and spatial viewpoints; and 3) extract a database of informative and reliable training samples. The database of weak labeled units is used for training an ensemble of classifiers on recent data whose results are then combined in a majority voting rule. Two sets of experimental results obtained on MS images by extracting training samples from a crop type map and the 2018 Corine Land Cover (CLC) map, respectively, confirm the effectiveness of the proposed approach.

In this paper we present the problem of a noisy lexical taxonomy and suggest two tasks as potential remedies. The first task is to identify and eliminate incorrect hypernymy links, and the second is to repopulate the taxonomy with new relations. The first task consists of revising the entire taxonomy and returning a Boolean for each assertion of hypernymy between two nouns (e.g. brie is a kind of cheese). The second task consists of recursively producing a chain of hypernyms for a given noun, until the most general node in the taxonomy is reached (e.g. brie → cheese → food → etc.). In order to achieve these goals, we implemented a hybrid hypernym-detection algorithm that incorporates various intuitions, such as syntagmatic, paradigmatic and morphological association measures as well as lexical patterns. We evaluate these algorithms individually and collectively and report findings in Spanish, English and French.

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) Hyperspectral imagery have hundreds of bands and greater subtle differences discrimination capacity in compariosion with multispectral imagery, which benefits precision applications. However, inherent high spectral resolution and high band correlation in these imagery suggests curse of dimensionality occurrence possibility in pattern recognition processes. So, study of the effects of dimensionality reduction is relevant for this kind of image. Additionaly, is relevant to compare the behavior between linear and non linear reduction methods. In this scenario, the purpose of the present work is to analyze how unsupervised feature extraction and its different approaches affect an unsupervised learning task in hyperspectral imagery. In order to conduct such analysis, the algorithms Principal Component Analysis, Isometric Feature Mapping and Locally Linear Embedding were executed in a set of seven images. Clusterings by K-Means and Expectation Maximization algorithms were built under each execution. Performances were measured by Rand, Jaccard, Kappa, Entropy and Purity indexes and compared by Friedman and Nemenyi statistical tests. Hypothesis tests results have shown that, for 70% of the images, feature extraction deployment raised clustering performance significantly and, in 60% of those cases, nonlinear extraction yielded better results than linear Imagens hiperespectrais possuem centenas de bandas e maior capacidade de discriminação de diferenças sutis em comparação a imagens multiespectrais, o que beneficia aplicações de precisão. Entretanto, a alta resolução espectral e alta correlação de bandas inerentes dessas imagens, sugerem a possibilidade de ocorrência da maldição da dimensionalidade em processos de reconhecimento de padrões. Dessa forma, o estudo dos efeitos de métodos de redução de dimensionalidade é relevante para esse tipo de imagem. Adicionalmente, é relevante a comparação de comportamento de métodos de redução linear e não-linear. Nesse cenário, o objetivo do presente trabalho é analisar como a extração não-supervisionada de características e suas diferentes abordagens afetam uma tarefa de aprendizado não-supervisionado em imagens hiperespectrais. Para conduzir tal análise, os algoritmos Principal Component Analysis, Isometric Feature Mapping e Locally Linear Embedding foram executados em um conjunto de sete imagens. Sob cada execução, foram realizados agrupamentos pelos algoritmos K-Means e Expectation Maximization. Os desempenhos foram mensurados pelas medidas Rand, Jaccard, Kappa, Entropy, Purity e comparados pelos testes estatísticos de Friedman e Nemenyi. Os resultados dos testes de hipótese mostraram que, para 70% das imagens, a aplicação da extração de características aumentou significativamente o desempenho da tarefa de agrupamento e, em 60% desses casos, a extração não-linear gerou melhores resultados em comparação à linear. CAPES: Código de Financiamento 001 CAPES: 88882.426564/2019-01

Unsupervised methods, such as principal component analysis, have gained popularity and wide-spread acceptance in the chemometrics and applied statistics communities. Unsupervised random forest is an additional method capable of discovering underlying patterns in the data. However, the number of applications of unsupervised random forest in chemometrics has been limited. One possible cause for this is the belief that random forest can only be used in a supervised analysis setting. This tutorial introduces the basic concepts of unsupervised random forest and illustrates several applications in chemometrics through worked examples.

This paper presents an unsupervised approach that extracts reliable labeled units from outdated maps to update them using time series (TS) of recent multispectral (MS) images. The method assumes that: 1) the source of the map is unknown and may be different from remote sensing data; 2) no ground truth is available; 3) the map is provided at polygon level, where the polygon label represents the dominant class; and 4) the map legend can be converted into a set of classes discriminable with the TS of images (i.e., no land-use classes that require manual analysis are considered). First, the outdated map is adapted to the spatial and spectral properties of the MS images. Then, the method identifies reliable labeled units in an unsupervised way by a two-step procedure: 1) a clustering analysis performed at polygon level to detect samples correctly associated to their labels and 2) a consistency analysis to discard polygons far from the distribution of the related land-cover class (i.e., having high probability of being mislabeled). Finally, the map is updated by classifying the recent TS of MS image with an ensemble of classifiers trained using only the reference data derived from the map. The experimental results obtained updating the 2012 Corine Land Cover (CLC) and the GlobLand30 in Trentino Alto Adige (Italy) achieved 93.2% and 93.3% overall accuracy (OA) on the validation data set. The method increased the OA up to 18% and 11.5% with respect to the reference methods on the 2012 CLC and the GlobLand30, respectively.

Word Sense Disambiguation (WSD) is the task of identifying the correct sense of a word in a specific context when the word has multiple meaning. WSD is very important as an intermediate step in many Natural Language Processing (NLP) tasks, especially in Information Extraction(IE), Machine Translation(MT) and Question/Answering Systems. Word sense ambiguity arises when a particular word has more than one possible sense. The peculiarity of any language is that it includes a lot of ambiguous words. Since the sense of a word depends on its context of use, disambiguation process requires the understanding of word knowledge. Automatic WSD systems are available for structured languages like English, Chinese, etc. But Indian languages are morphologically rich and thus the processing task is very complex. The aim of this work is to develop a WSD system for Malayalam, a language spoken in India, predominantly used in the state of Kerala. The proposed system uses a corpus which is collected from various Malayalam web documents. For each possible sense of the ambiguous word, a relatively small set of training examples (seed sets) are identified which represents the sense. Collocations and most co-occurring words are considered as training examples. Seed set expansion module extends the seed set by adding most similar words to the seed set elements. These extended sets act as sense clusters. The most similar sense cluster to the input text context is considered as the sense of the target word.

Unsupervised methods, such as principal component analysis, have gained popularity and wide-spread acceptance in the chemometrics and applied statistics communities. Unsupervised random forest is an additional method capable of discovering underlying patterns in the data. However, the number of applications of unsupervised random forest in chemometrics has been limited. One possible cause for this is the belief that random forest can only be used in a supervised analysis setting. This tutorial introduces the basic concepts of unsupervised random forest and illustrates several applications in chemometrics through worked examples.

Doutoramento em Informática Genomic sequences are large codi ed messages describing most of the structure of all known living organisms. Since the presentation of the rst genomic sequence, a huge amount of genomics data have been generated, with diversi ed characteristics, rendering the data deluge phenomenon a serious problem in most genomics centers. As such, most of the data are discarded (when possible), while other are compressed using general purpose algorithms, often attaining modest data reduction results. Several speci c algorithms have been proposed for the compression of genomic data, but unfortunately only a few of them have been made available as usable and reliable compression tools. From those, most have been developed to some speci c purpose. In this thesis, we propose a compressor for genomic sequences of multiple natures, able to function in a reference or reference-free mode. Besides, it is very exible and can cope with diverse hardware speci cations. It uses a mixture of nite-context models (FCMs) and eXtended FCMs. The results show improvements over state-of-the-art compressors. Since the compressor can be seen as a unsupervised alignment-free method to estimate algorithmic complexity of genomic sequences, it is the ideal candidate to perform analysis of and between sequences. Accordingly, we de ne a way to approximate directly the Normalized Information Distance, aiming to identify evolutionary similarities in intra- and inter-species. Moreover, we introduce a new concept, the Normalized Relative Compression, that is able to quantify and infer new characteristics of the data, previously undetected by other methods. We also investigate local measures, being able to locate speci c events, using complexity pro les. Furthermore, we present and explore a method based on complexity pro les to detect and visualize genomic rearrangements between sequences, identifying several insights of the genomic evolution of humans. Finally, we introduce the concept of relative uniqueness and apply it to the Ebolavirus, identifying three regions that appear in all the virus sequences outbreak but nowhere in the human genome. In fact, we show that these sequences are su cient to classify di erent sub-species. Also, we identify regions in human chromosomes that are absent from close primates DNA, specifying novel traits in human uniqueness. As sequências genómicas podem ser vistas como grandes mensagens codificadas, descrevendo a maior parte da estrutura de todos os organismos vivos. Desde a apresentação da primeira sequência, um enorme número de dados genómicos tem sido gerado, com diversas características, originando um sério problema de excesso de dados nos principais centros de genómica. Por esta razão, a maioria dos dados é descartada (quando possível), enquanto outros são comprimidos usando algoritmos genéricos, quase sempre obtendo resultados de compressão modestos. Têm também sido propostos alguns algoritmos de compressão para sequências genómicas, mas infelizmente apenas alguns estão disponíveis como ferramentas eficientes e prontas para utilização. Destes, a maioria tem sido utilizada para propósitos específicos. Nesta tese, propomos um compressor para sequências genómicas de natureza múltipla, capaz de funcionar em modo referencial ou sem referência. Além disso, é bastante flexível e pode lidar com diversas especificações de hardware. O compressor usa uma mistura de modelos de contexto-finito (FCMs) e FCMs estendidos. Os resultados mostram melhorias relativamente a compressores estado-dearte. Uma vez que o compressor pode ser visto como um método não supervisionado, que não utiliza alinhamentos para estimar a complexidade algortímica das sequências genómicas, ele é o candidato ideal para realizar análise de e entre sequências. Em conformidade, definimos uma maneira de aproximar directamente a distância de informação normalizada (NID), visando a identificação evolucionária de similaridades em intra e interespécies. Além disso, introduzimos um novo conceito, a compressão relativa normalizada (NRC), que é capaz de quantificar e inferir novas características nos dados, anteriormente indetectados por outros métodos. Investigamos também medidas locais, localizando eventos específicos, usando perfis de complexidade. Propomos e exploramos um novo método baseado em perfis de complexidade para detectar e visualizar rearranjos genómicos entre sequências, identificando algumas características da evolução genómica humana. Por último, introduzimos um novo conceito de singularidade relativa e aplicamo-lo ao Ebolavirus, identificando três regiões presentes em todas as sequências do surto viral, mas ausentes do genoma humano. De facto, mostramos que as três sequências são suficientes para classificar diferentes sub-espécies. Também identificamos regiões nos cromossomas humanos que estão ausentes do ADN de primatas próximos, especificando novas características da singularidade humana.

Doutoramento em Informática Submitted by Patrícia Correia (patriciacorreia@ua.pt) on 2016-11-15T11:51:34Z No. of bitstreams: 1 Compressão e análise de dados genómicos.pdf: 28471056 bytes, checksum: 1f3688406df1249d02a1024e178fbd1a (MD5) Made available in DSpace on 2016-11-15T11:51:34Z (GMT). No. of bitstreams: 1 Compressão e análise de dados genómicos.pdf: 28471056 bytes, checksum: 1f3688406df1249d02a1024e178fbd1a (MD5) Previous issue date: 2015

In this paper we are going to detail an unsupervised, graph-based approach for word sense discrimination on tweets. We deal with this problem by constructing a word graph of co-occurrences. By defining a distance on this graph, we obtain a word metric space, on which we can apply an aggregative algorithm for word clustering. As a result, we will get word clusters representing contexts that discriminate the possible senses of a term. We present some experimental results both on a data set consisting of tweets we collected and on the data set of task 14 at SemEval-2010.

In this paper we present an unsupervised, graph-based approach for Word Sense Discrimination. Given a set of text sentences, a word co-occurrence graph is derived and a distance based on Jaccard index is defined on it; subsequently, the new distance is used to cluster the neighbour nodes of ambiguous terms using the concept of “gangplanks” as edges that separate denser regions (“islands”) in the graph. The proposed approach has been evaluated on a real data set, showing promising performance in Word Sense Discrimination. L’obiettivo di questo articolo è descrivere un approccio di clustering non supervisionato e basato su grafi per individuare e discriminare i differenti sensi che un termine può assumere all’interno di un testo. Partendo da un grafo di cooccorrenze, vi definiamo una distanza fra nodi e applichiamo un algoritmo basato sulle “passerelle”, cioè archi che separano regioni dense (“isole”) all’interno del grafo. Discutiamo i risultati ottenuti su un insieme di dati composto da tweet.

This paper reports experiments on automatically detecting the gender of Twitter users, based on unstructured information found on their Twitter profile. A set of features previously proposed is evaluated on two datasets of English and Portuguese users, and their performance is assessed using several supervised and unsupervised approaches, including Naive Bayes variants, Logistic Regression, Support Vector Machines, Fuzzy c-Means clustering, and k-means. Results show that features perform well in both languages separately, but even best results were achieved when combining both languages. Supervised approaches reached 97.9 % accuracy, but Fuzzy c-Means also proved suitable for this task achieving 96.4 % accuracy. info:eu-repo/semantics/acceptedVersion

International audience; Automatic synonyms and semantically related word extraction is a challenging task, useful in many NLP applications such as question answering, search query expansion, text summarization, etc. While different studies addressed the task of word synonym extraction, only a few investigations tackled the problem of acquiring synonyms of multi-word terms (MWT) from specialized corpora. To extract pairs of synonyms of multi-word terms, we propose in this paper an unsupervised semi-compositional method that makes use of distributional semantics and exploit the compositional property shared by most MWT. We show that our method outperforms significantly the state-of-the-art.