Search

EuroVoc is a thesaurus maintained by the European Union Publication Office, used to describe and index legislative documents. The EuroVoc concepts are organized following a hierarchical structure, with 21 domains, 127 micro-thesauri terms, and more than 6,700 detailed descriptors. The large number of concepts in the EuroVoc thesaurus makes the manual classification of legal documents highly costly. In order to facilitate this classification work, we present two main contributions. The first one is the development of a hierarchical deep learning model to address the classification of legal documents according to the EuroVoc thesaurus. Instead of training a classifier for each hierarchy level, our model allows the simultaneous prediction of the three levels of the EuroVoc thesaurus. Our second contribution concerns the proposal of a new legal corpus for evaluating the classification of documents written in Portuguese. This corpus, named EUR-Lex PT, contains more than 220k documents, labeled under the three EuroVoc hierarchical levels. Comparative experiments with other state-of-the-art models indicate that our approach has competitive results, at the same time offering the ability to interpret predictions through attention weights.

This article reports on the 3rd ACM SIGSPATIAL Workshop on Geospatial Humanities, held in conjunction with the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. The article outlines the objectives of the workshop, and briefly describes the technical program.

Scholars in the humanities have a long tradition in the investigation of spatial theory and methods, as well as on the analysis of space and place within its multiple disciplines. Although technolo...

Deep neural networks have achieved state-of-the-art results in various vision and/or language tasks. Despite the use of large training datasets, most models are trained by iterating over single input-output pairs, discarding the remaining examples for the current prediction. In this work, we actively exploit the training data, using the information from nearest training examples to aid the prediction both during training and testing. Specifically, our approach uses the target of the most similar training example to initialize the memory state of an LSTM model, or to guide attention mechanisms. We apply this approach to image captioning and sentiment analysis, respectively through image and text retrieval. Results confirm the effectiveness of the proposed approach for the two tasks, on the widely used Flickr8 and IMDB datasets. Our code is publicly available at http://github.com/RitaRamo/retrieval-augmentation-nn., Accepted at IJCNN 2021

International audience; Geocoding aims to assign unambiguous locations (i.e., geographic coordinates) to place names (i.e., toponyms) referenced within documents (e.g., within spreadsheet tables or textual paragraphs). This task comes with multiple challenges, such as dealing with referent ambiguity (multiple places with a same name) or reference database completeness. In this work, we propose a geocoding approach based on modeling pairs of toponyms, which returns latitude-longitude coordinates. One of the input toponyms will be geocoded, and the second one is used as context to reduce ambiguities. The proposed approach is based on a deep neural network that uses Long Short-Term Memory (LSTM) units to produce representations from sequences of character n-grams. To train our model, we use toponym co-occurrences collected from different contexts, namely textual (i.e., co-occurrences of toponyms in Wikipedia articles) and geographical (i.e., inclusion and proximity of places based on Geonames data). Experiments based on multiple geographical areas of interest—France, United States, Great-Britain, Nigeria, Argentina and Japan—were conducted. Results show that models trained with co-occurrence data obtained a higher geocoding accuracy, and that proximity relations in combination with co-occurrences can help to obtain a slightly higher accuracy in geographical areas with fewer places in the data sources.

Our work aimed at experimentally assessing the benefits of model ensembling within the context of neural methods for passage re-ranking. Starting from relatively standard neural models, we use a previous technique named Fast Geometric Ensembling to generate multiple model instances from particular training schedules, then focusing or attention on different types of approaches for combining the results from the multiple model instances (e.g., averaging the ranking scores, using fusion methods from the IR literature, or using supervised learning-to-rank). Tests with the MS-MARCO dataset show that model ensembling can indeed benefit the ranking quality, particularly with supervised learning-to-rank although also with unsupervised rank aggregation.

While statistical information on socio-economic activities is widely available, the data are often collected or released only at a relatively aggregated level. In these aggregated forms, the data are useful for broad-scale assessments, although we often need to disaggregate the source data in order to provide more localized estimates, and in order to analyze correlations against geophysical variables. Spatial disaggregation techniques can be used in this context, to transform data from a set of source zones into a set of target zones, with different geometry and with a higher general level of spatial resolution. Still, few previous studies in the area have attempted to leverage state-of-the-art spatial disaggregation procedures in the context of socio-economic variables, instead focusing on applications related to population modeling. In this article, we report on experiments with a hybrid spatial disaggregation technique that combines state-of-the-art regression analysis procedures with the classic methods of dasymetric mapping and pycnophylactic interpolation. The hybrid procedure was used together with population density, land coverage, nighttime satellite imagery, and OpenStreetMap road density, as ancillary data to disaggregate different types of socio-economic indicators to a high-resolution grid. Our test specifically leveraged data relative to the Portuguese territory, resulting in the production of raster datasets with a resolution of 30 arc-seconds per cell. The article discusses the spatial disaggregation methodology and the quality of the obtained results under different experimental conditions.

Nowadays artificial intelligence (AI) is bringing tremendous opportunities and challenges to geospatial research. Big data enable computers to observe and learn the world from many different perspectives, while high performance machines support the development, training, and deployment of AI models within reasonable amount of time. Recent years have witnessed significant advances in the integration of geospatial study and AI in both academia and industry. There have already been many successful studies for both physical environment and human society. Focusing on modeling the physical nature, research has shown that deep learning can improve the representation of clouds that are smaller than the grid resolutions of climate models. Examining the human society, AI and natural language processing methods, such as word embeddings, help quantify changes in stereotypes and attitudes toward women and ethnic minorities over 100 years in the United States. There are also many other applications that effectively integrate AI with problems in geospatial studies, such as vehicle trajectory prediction, high-definition mapping and navigation, historical map digitizing, gazetteer conflation, geographic feature extraction, and place understanding. The 3nd International Workshop on AI for Geographic Knowledge Discovery (GeoAI 2019) builds on the success of the previous workshops in 2017 and 2018. GeoAI is bringing together geoscientists, computer scientists, engineers, entrepreneurs, and decision makers from academia, industry, and government to discuss the latest trends, successes, challenges, and opportunities in the field of artificial intelligence for data mining and geographic knowledge discovery.

Toponym resolution refers to the disambiguation of place names and other references to places present in textual documents, resolving them to unambiguous geographical identifiers (e.g., geographic coordinates of latitude and longitude). One of the major challenges in this task is that, usually, place names are highly ambiguous (e.g., there are several locations on the surface of the Earth that share the same name). In this paper, we propose to address the task through a recurrent neural network architecture with multiple inputs and outputs, specifically leveraging pre-trained contextual embeddings (ELMo) and bi-directional Long Short-Term Memory (LSTM) units, both commonly used for textual data modeling. The proposed model was tested on two datasets that were previously used to evaluate toponym resolution systems, namely the War of the Rebellion and the Local-Global Lexicon corpora. The obtained results outperform state-of-the-art results, confirming the superiority of the proposed method over other previous approaches.

EuroVoc is a thesaurus maintained by the European Union Publication Office, used to describe and index legislative documents. The Eurovoc concepts are organized following a hierarchical structure, with 21 domains, 127 micro-thesauri terms, and more than 6,700 detailed descriptors. The large number of concepts in the EuroVoc thesaurus makes the manual classification of legal documents highly costly. In order to facilitate this classification work, we present two main contributions. The first one is the development of a hierarchical deep learning model to address the classification of legal documents according to the EuroVoc thesaurus. Instead of training a classifier for each level, our model allows the simultaneous prediction of the three levels of the EuroVoc thesaurus. Our second contribution concerns the proposal of a new legal corpus for evaluating the classification of documents written in Portuguese. Our proposed corpus, named EUR-Lex PT, contains more than 220k documents, labeled under the three EuroVoc hierarchical levels. Comparative experiments with other state-of-the-art models indicate that our approach has competitive results, at the same time offering the ability to interpret predictions through attention weights.

Non-invasive medical imaging techniques, such as radiography or computed tomography, are extensively used in hospitals and clinics for the diagnosis of diverse injuries or diseases. However, the interpretation of these images, which often results in a free-text radiology report and/or a classification, requires specialized medical professionals, leading to high labor costs and waiting lists. Automatic inference of thoracic diseases from the results of chest radiography exams, e.g. for the purpose of indexing these documents, is still a challenging task, even if combining images with the free-text reports. Deep neural architectures can contribute to a more efficient indexing of radiology exams (e.g., associating the data to diagnostic codes), providing interpretable classification results that can guide the domain experts. This work proposes a novel multi-modal approach, combining a dual path convolutional neural network for processing images with a bidirectional recurrent neural network for processing text, enhanced with attention mechanisms and leveraging pre-trained clinical word embeddings. The experimental results show interesting patterns, e.g. validating the high performance of the individual components, and showing promising results for the multi-modal processing of radiology examination data, particularly when pre-training the components of the model with large pre-existing datasets (i.e., a 10% increase in terms of the average value for the areas under the receiver operating characteristic curves).

This survey article describes previous research addressing text-based document geocoding, i.e. the task of predicting the geospatial coordinates of latitude and longitude, that best correspond to an entire document, based on its textual contents. We describe (1) early document geocoding systems that use heuristics over place names mentioned in the text (e.g. names of cities and states), (2) probabilistic language modeling approaches, where generative models are built for different regions in the world (usually considering a discretization based on a rectangular grid) from the words occurring in a set of georeferenced training documents, which are then used to predict per-region probabilities for previously unseen test documents, (3) combinations of different models and heuristics, including clustering procedures, feature selection approaches, and/or language models built from different sources, and (4) recent approaches based on discriminative classification models.

Location is a major feature for describing the real world, appearing in many Web applications and search systems as a cross-cutting issue. Envisioning the establishment of an adequate venue for the location topic, LocWeb 2015 continued a workshop series at the intersection of geospatial search, information management, and Web architecture, with a main focus on location-aware information access. Contributions to the workshop highlighted various fields that benefit from a tailored approach towards location features. LocWeb 2015 showed an interdisciplinary mix of contributions, featuring one keynote and three full papers. In this report, we provide a brief overview of the workshop and its contributions.

Fake news are nowadays an issue of pressing concern, given their recent rise as a potential threat to high-quality journalism and well-informed public discourse. The Fake News Challenge (FNC-1) was organized in 2017 to encourage the development of machine learning-based classification systems for stance detection (i.e., for identifying whether a particular news article agrees, disagrees, discusses, or is unrelated to a particular news headline), thus helping in the detection and analysis of possible instances of fake news. This article presents a new approach to tackle this stance detection problem, based on the combination of string similarity features with a deep neural architecture that leverages ideas previously advanced in the context of learning efficient text representations, document classification, and natural language inference. Specifically, we use bi-directional Recurrent Neural Networks, together with max-pooling over the temporal/sequential dimension and neural attention, for representing (i) the headline, (ii) the first two sentences of the news article, and (iii) the entire news article. These representations are then combined/compared, complemented with similarity features inspired on other FNC-1 approaches, and passed to a final layer that predicts the stance of the article towards the headline. We also explore the use of external sources of information, specifically large datasets of sentence pairs originally proposed for training and evaluating natural language inference methods, in order to pre-train specific components of the neural network architecture (e.g., the RNNs used for encoding sentences). The obtained results attest to the effectiveness of the proposed ideas and show that our model, particularly when considering pre-training and the combination of neural representations together with similarity features, slightly outperforms the previous state-of-the-art., Comment: Accepted for publication in the special issue of the ACM Journal of Data and Information Quality (ACM JDIQ) on Combating Digital Misinformation and Disinformation

This monograph aims at analyzing narratives and accounts produced by a group of non-professional military interpreters from the Brazilian Army who were assigned to the United Nations Mission for the Stabilization of Haiti (MINUSTAH) to carry out tasks of interpretation and translation from Portuguese to English and vice versa, in an effort to understand whether they could be seen as a community of practice in the terms proposed by Wenger (1998). The main purpose of this study was to check how the group carried out its missions without being professionals of the areas of interpretation and translation, how their members dealt with their difficulties and what they learned from the experience they had. The methodology used was the collection of interviews from three volunteers who took part in the mission, with the objective of studying their accounts, and narratives using the Labovian theoretical framework, making a correlation with the ideas of Community of Practice. Finally, my findings suggest that the group was working within the characteristics of a community of practice and their professional experience could be used as a reference not only for the military, but also for the academic setting.

As cyberbullying becomes more and more frequent in social networks, automatically detecting it and pro-actively acting upon it becomes of the utmost importance. In this work, we study how a recent technique with proven success in similar tasks, Fuzzy Fingerprints, performs when detecting textual cyberbullying in social networks. Despite being commonly treated as binary classification task, we argue that this is in fact a retrieval problem where the only relevant performance is that of retrieving cyberbullying interactions. Experiments show that the Fuzzy Fingerprints slightly outperforms baseline classifiers when tested in a close to real life scenario, where cyberbullying instances are rarer than those without cyberbullying.

The field of Spatial Humanities has advanced substantially in the past years. The identification and extraction of toponyms and spatial information mentioned in historical text collections has allowed its use in innovative ways, making possible the application of spatial analysis and the mapping of these places with Geographic Information Systems. For instance, automated place name identification is nowadays possible with Named Entity Recognition (NER) systems. Statistical NER methods based on supervised learning, in particular, are highly successful with modern datasets. However, there are still major challenges to address when dealing with historical corpora. These challenges include language changes over time, spelling variations, transliterations, OCR errors, and sources written in multiple languages among others. In this article, considering a task of place name recognition over two collections of historical correspondence, we report an evaluation of five NER systems and an approach that combines these through a voting system. We found that although individual performance of each NER system was corpus dependent, the ensemble combination was able to achieve consistent measures of precision and recall, outperforming the individual NER systems. Additionally, the results showed that these NER system are not strongly dependent on pre-processing and translation to modern English.

With the prevalence of mobile computing systems and location based services, large amounts of spatio-temporal data are nowadays being collected, representing the mobility of people performing various activities. However, despite the increasing interest in the exploration of these data, there are still open challenges in various application contexts, e.g. related to visualisation and human–computer interaction. In order to support the extraction of useful and relevant information from the spatio-temporal and the thematic properties associated with human trajectories, it is crucial to develop and study adequate interactive visualisation techniques. In addition to the properties of the visualisations themselves, it is important to take into consideration the types of information present within the data and, more importantly, the types of tasks that a user might need to consider in order to achieve a given goal. The understanding of these factors may, in turn, simplify the development and the assessment of a given interactive visualisation. In this paper, we present and analyse the most relevant concepts associated to these topics. In particular, our analysis addresses the main properties associated with human trajectory data, the main types of visualisation tasks/objectives that the users may require in order to analyse that data and the high-level classes of techniques for visualising trajectory data. In addition, this paper also presents an overview on a user study, conducted in function of this analysis, to compare two classes of visualisation techniques, namely static maps and space-time cubes, regarding their adequacy in helping users completing basic visualisation tasks.

This article addresses the problem of disambiguating named entities, in text documents, towards entries in a knowledge base like Wikipedia. The proposed approach uses supervised learning to sort candidate knowledge base entries for each entity mentioned in a text, and then to classify the entry ranked in the first position as either the correct disambiguation or not. We present results with Portuguese and Spanish texts for a wide range of models and configuration options. Our experiments attest to the effectiveness of supervised learning methods in this specific task, showing that out-of-the-box algorithms and relatively simple features can achieve a high accuracy.

We address the assignment of ICD-10 codes for causes of death by analyzing free-text descriptions in death certificates, together with the associated autopsy reports and clinical bulletins, from the Portuguese Ministry of Health. We leverage a deep neural network that combines word embeddings, recurrent units, and neural attention, for the generation of intermediate representations of the textual contents. The neural network also explores the hierarchical nature of the input data, by building representations from the sequences of words within individual fields, which are then combined according to the sequences of fields that compose the inputs. Moreover, we explore innovative mechanisms for initializing the weights of the final nodes of the network, leveraging co-occurrences between classes together with the hierarchical structure of ICD-10. Experimental results attest to the contribution of the different neural network components. Our best model achieves accuracy scores over 89%, 81%, and 76%, respectively for ICD-10 chapters, blocks, and full-codes. Through examples, we also show that our method can produce interpretable results, useful for public health surveillance.

Several tasks related to geographical information retrieval and to the geographical information sciences involve toponym matching, that is, the problem of matching place names that share a common referent. In this article, we present the results of a wide-ranging evaluation on the performance of different string similarity metrics over the toponym matching task. We also report on experiments involving the usage of supervised machine learning for combining multiple similarity metrics, which has the natural advantage of avoiding the manual tuning of similarity thresholds. Experiments with a very large dataset show that the performance differences for the individual similarity metrics are relatively small, and that carefully tuning the similarity threshold is important for achieving good results. The methods based on supervised machine learning, particularly when considering ensembles of decision trees, can achieve good results on this task, significantly outperforming the individual similarity metrics.

The assignment of disease codes to clinical texts has a wide range of applications, including epidemiological studies or disease surveillance. We address the task of automatically assigning the ICD-10 codes for the underlying cause of death, from the free-text descriptions included in death certificates obtained from the Portuguese Ministry of Health. We specifically propose to leverage a deep neural network based on a two-level hierarchy of recurrent nodes together with attention mechanisms. The first level uses recurrent nodes for modeling the sequences of words given in individual fields of the death certificates, together with attention to weight the contribution of each word, producing intermediate representations for the contents of each field. The second level uses recurrent nodes to model a sequence of fields, using the representations produced by the first level and also leveraging attention in order to weight the contributions of the different fields. The paper reports on experiments with a dataset of 115,406 death certificates, presenting the results of an evaluation of the predictive accuracy of the proposed method, for different ICD-10 levels (i.e., chapter, block, or full code) and for particular causes of death. We also discuss how the neural attention mechanisms can help in interpreting the classification results.

This article presents an automated method for defining the boundaries of imprecise geographic regions, based on publicly available data. The method uses one-class support vector machines SVMs for interpolating from a set of point locations, which are assumed to lie in the region whose boundaries are to be defined, and leverages also on a combination of multiple Gaussian kernels, within the formalism of SVMs, to improve accuracy. The points that are used for model training correspond to geospatial coordinates associated with Flickr photos that are tagged with the name of the vague region to be defined. Besides considering latitude and longitude coordinates from Flickr photos, as done in a previous related work, each point location is also associated with a set of descriptive features, obtained from textual annotations and from publicly available raster datasets encoding population counts, terrain elevation, and/or land coverage information. The overall approach is evaluated by means of statistical classification measures, using regions whose boundaries are well defined i.e., the official boundaries for several European countries. Besides this formal evaluation, we also illustrate our results for several vague regions. Results show that our method performs better than a previous state-of-the-art approach i.e., we measured an improvement of 5.5% in terms of the F1 metric, which was based solely on interpolating from the geospatial coordinates of known points.

This paper presents a machine learning method for disambiguating place references in text. Solving this task can have important applications in the digital humanities and computational social sciences, by supporting the geospatial analysis of large document collections. We combine multiple features that capture the similarity between candidate disambiguations, the place references, and the context where the place references occur, in order to rank and choose from a set of candidate disambiguations, obtained from a knowledge base containing geospatial coordinates and textual descriptions for different places from all around the world. The proposed method was evaluated through English corpora used in previous work in this area, and also with a subset of the English Wikipedia. Experimental results demonstrate that the proposed method is indeed effective, showing that out-of-the-box learning algorithms and relatively simple features can obtain a high accuracy in this task.

This article reports on the 1st ACM SIGSPATIAL Workshop on Geospatial Humanities, held in conjunction with the 25th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. The article outlines the objectives of the workshop, and briefly describes the technical program.

High-resolution population grids built from historical census data can ease the analyses ofgeographical population changes, at the same time also facilitating the combination of populationdata with other GIS layers to perform analyses on a wide range of topics. This article reports onexperiments with a hybrid spatial disaggregation technique that combines the ideas of dasymetricmapping and pycnophylactic interpolation, using modern machine learning methods to combinedifferent types of ancillary variables, in order to disaggregate historical census data into a 200 mresolution grid. We specifically report on experiments related to the disaggregation of historicalpopulation counts from three different national censuses which took place around 1900, respectively inGreat Britain, Belgium, and the Netherlands. The obtained results indicate that the proposed methodis indeed highly accurate, outperforming simpler disaggregation schemes based on mass-preservingareal weighting or pycnophylactic interpolation. The best results were obtained using modernregression methods (i.e., gradient tree boosting or convolutional neural networks, depending on thecase study), which previously have only seldom been used for spatial disaggregation.

This paper describes Triplex-ST, a novel information extraction system for collecting spatio-temporal information from textual resources. Triplex-ST is based on a distantly supervised approach, which leverages rich linguistic annotations together with information in existing knowledge bases. In particular, we leverage triples associated with temporal and/or spatial contexts, e.g., as available from the YAGO knowledge base, so as to infer templates that capture new facts from previously unseen sentences.

Throughout the years, researchers have tried to understand dynamics and general patterns associated with human movement, e.g. in the context of urban planning, to improve the lives of citizens. To do this, it is necessary to properly analyse the spatio-temporal and the thematic properties of their trajectory data. Thematic maps, particularly 2D maps and 3D space-time cubes (STCs), are among the most common approaches to analyse and visualize these data. Previous research attests to the usefulness of these visualization techniques in different types of tasks. However, it is unlikely that the analysis of trajectories will be always limited to a specific type of task, thus, motivating additional studies to evaluate the dis/advantages of combining both techniques, within the same view/interaction context. In this paper, we address this specific challenge, by presenting a comparative study between three prototypes, one using a 2D map, one using a STC, and one combining both techniques, for the visualization of trajectories. Our results support previous studies' conclusions, by showing the advantages of 2D maps and STCs for spatial and spatio-temporal tasks, respectively. They also point out towards the advantages of using both techniques together, as users seem to prefer this alternative and are able to complete different types of tasks accurately, despite the increasing complexity of the visualization.

The deployment of smart metering in Brazil is still initial and various technological and social issues need to be discussed before the decision on how the vast meter park should be. The purpose of this study is to analyze smart meters on the perspective of each side interested in its adoption in the Brazilian electric sector. We analyze the comprehensive range of functions that a meter can have, highlighting remote communication, and power quality features according to the different scenarios of Brazillian distribution system. Finally, we suggest a smart metering model converging the existing demands and interests of the Brazilian electric sector.

This work presents the implementation, on computers with parallel architecture, of domain decomposition strategies for the coupled analysis of floating production systems (FPS). The strategies are associated to formulations for the coupling of the equations of motion of hull and lines, and to different algorithms for the solution of these equations. The strategies are grouped as “external” and “internal”. The former incorporates a “master-slave” scheme where the hull equations are solved in the “master” processor, and the equations of each line are solved in the “slave” processors. This scheme is associated to an external subcycling procedure, where the time step employed to integrate the hull equations is larger than the step for the lines. In the “internal” partition strategies, the FE mesh of each line is partitioned amongst the processors. Different strategies are presented, including an internal subcycling procedure associated to an explicit algorithm, using distinct time steps to integrate the equations of each partition (according to its physical characteristics); and an implicit domain decomposition method where completely arbitrary mesh partitions may be defined. Results of case studies are presented, to assess the performance of the parallel implementations described.

This vision paper argues that a geospatial knowledge base combining textual descriptions for concepts such as places, together with place types, semantic relations between concepts and, most importantly, polygonal geometries associated to the geospatial concepts, constitutes a valuable resource for researchers working on the computational modeling of spatial language. We describe a simple procedure for producing one such resource from existing open datasets, and discuss possible ways for moving beyond the current state-of-the-art within the general area of geospatial text mining, through studies supported by one such knowledge base.

Many natural language processing applications require information about the spatial locations of objects referenced in text, or spatial relations between these objects in space. For example, the phrase a book on the shelf contains information about the location of the object book, corresponding to a trajector, with respect to the object shelf, which in turn corresponds to a landmark. Spatial role labeling concerns with the task of automatically processing textual sentences and identifying objects of spatial scenes and relations between them. In this paper, we describe the application of modern machine learning methods to extract spatial roles and their relations, specifically by adapting a pre-existing system based on a convolutional neural network architecture that has been recently proposed for the more general task of semantic role labeling. We report on experiments with datasets from the SemEval challenges on spatial role labeling, showing that our method can achieve results in line with the current state-of-the-art. We therefore argue that that spatial role labeling can leverage on recent developments in semantic role labeling, requiring only minimal adaptations.

In this paper, we evaluate automated techniques, based on a hierarchical representation for the Earth's surface and leveraging SVM classifiers, for assigning geospatial coordinates to previously unseen documents, using only the raw text as input evidence. We report on experiments with Wikipedia documents in four different languages, and with two Twitter datasets from previous studies. We obtained state-of-the-art results, showing that document geocoding can be handled effectively with appropriate bag-of-words representations and with out-of-the-box supervised learning methods.

Two dimensional static maps and three dimensional space-time cubes are among the most studied techniques to visualize human movement data. Previous research suggests that both techniques are useful in different types of tasks. However, the analysis of trajectory data may not be focused in just one type of task, motivating further studies to quantify the advantages of combining both types of techniques. This paper describes our work-in-progress addressing this issue, proposing the combination of 2D maps and 3D space-time cubes for human trajectory visualization, and overviewing possible metrics for its evaluation.

This project arose from the absence in the market of a modular smartphone controlled robot capable of encouraging high school students to program and apply the physics and math's knowledge learned into it. Therefore this project's intention was to study the best way to develop a do-it-yourself (DIY) cost effective robot using only components off the shelf (COTS) and benefit from the omnipresence of smartphones. With the objective of making this robot attractive to anyone with low programming skills, it was important to make it configurable in an easy to understand language and a simple user interface, like the ones provided by Scratch and the MIT AppInventor2. The functional, physical and non-functional requirements for this robot and the free software developed are presented and validated attesting that this project was successfully completed.

This paper proposes a novel method that leverages georeferenced social media data, together with human assessments of particular words, to estimate population well-being across the U.S. territory. We specifically attempt to learn linear regression models that, by leveraging on simple features that essentially correspond to word counts in lexicons of emotionally-charged words, are capable of approximating a composite well-being index built through traditional surveying methods. Experiments with a large Twitter dataset collected within the year of 2012 attest for the feasibility of the proposed approach (i.e., we approximate the Gallup-Healthways composite well-being index with a mean absolute error of 0.91), and we then produced choropleth maps, either at a state- or at a county-level of detail, that show how well-being varies across the continental U.S. territory.

We present the INESC-ID system for the message polarity classification task of SemEval 2015. The proposed system does not make use of any hand-coded features or linguistic resources. It relies on projecting pre-trained structured skip-gram word embeddings into a small subspace. The word embeddings can be obtained from large amounts of Twitter data in unsupervised form. The sentiment analysis supervised training is thus reduced to finding the optimal projection which can be carried out efficiently despite the little data available. We analyze in detail the proposed approach and show that a competitive system can be attained with only a few configuration parameters.

Semi-supervised bootstrapping techniques for relationship extraction from text iteratively expand a set of initial seed relationships while limiting the semantic drift. We research bootstrapping for relationship extraction using word embeddings to find similar relationships. Experimental results show that relying on word embeddings achieves a better performance on the task of extracting four types of relationships from a collection of newswire documents when compared with a baseline using TFIDF to find similar relationships.

We present the approach followed by INESCID in the SemEval 2015 Twitter Sentiment Analysis challenge, subtask E. The goal was to determine the strength of the association of Twitter terms with positive sentiment. Using two labeled lexicons, we trained a regression model to predict the sentiment polarity and intensity of words and phrases. Terms were represented as word embeddings induced in an unsupervised fashion from a corpus of tweets. Our system attained the top ranking submission, attesting the general adequacy of the proposed approach.

This paper describes a system developed for the disorder identification subtask within task 14 of SemEval 2015. The developed system is based on a chain of two modules, one for recognition and another for normalization. The recognition module is based on an adapted version of the Stanford NER system to train CRF models in order to recognize disorder mentions. CRF models were build based on a novel encoding of entity spans as token classifications to also consider non-continuous entities, along with a rich set of features based on (i) domain lexicons and (ii) Brown clusters inferred from a large collection of clinical texts. For disorder normalization, we (i) generated a non ambiguous dictionary of abbreviations from the labelled files, using it together with (ii) an heuristic method based on similarity search and (iii) a comparison method based on the information content of each disorder. The system achieved an F-measure of 0.740 (the second best), with a precision of 0.779, a recall of 0.705.

Discovering knowledge from textual sources and subsequently expanding the coverage of knowledge bases like DBpedia or Freebase currently requires either extensive manual work or carefully designed information extractors. Information extractors capture triples from textual sentences. Each triple consists of a subject, a predicate/property, and an object. Triples can be mediated via verbs, nouns, adjectives, and appositions. We propose Triplex, an information extractor that complements previous efforts, concentrating on noun-mediated triples related to nouns, adjectives, and appositions. Triplex automatically constructs templates expressing noun-mediated triples from a bootstrapping set. The bootstrapping set is constructed without manual intervention by creating templates that include syntactic, semantic, and lexical constraints. We report on an automatic evaluation method to examine the output of information extractors both with and without the Triplex approach. Our experimental study indicates that Triplex is a promising approach for extracting noun-mediated triples.

With the increasing evolution of computer graphics, 3D visualizations have become more common and are nowadays seen as a promising way to represent complex types of information. In particular, space-time cubes (STC) have been proposed as an alternative to 2D maps for the visualization of spatio-temporal data, and they have become increasingly used to explore the dynamics and patterns of human movement. However, previous research has pointed out perceptual limitations that can condition the use of 3D views for decoding locations and spatial properties. We aim to address those issues by presenting a comparative study between three variants of the STC technique, with different methods to improve spatial awareness. Our results support that the use of a movable plane or an additional 2D map view improve users’ accuracy when performing common tasks, and are preferred over simpler, yet less cluttered approaches. Additionally, it also supports the possible advantages of combining 2D and 3D views for human trajectory visualization.

Relation extraction from textual documents is an important task in the context of information extraction. This task aims at identifying relations between pairs of named entities and assigning them a type. Relation extraction is often approached as a supervised classification problem, involving pre-processing steps such as text segmentation, entity recognition, and morphological and syntactic annotations. In previous studies, the way data is pre-processed differs among them, thus making the comparison of classification techniques for relation extraction unfair and inconclusive. Some of these classification techniques for relation extraction involve the use of kernels, which enable the comparison of complex structures. We propose a benchmark for the comparison of different kernels for relation extraction. Specifically, we propose the application of a common pre-processing stage, together with the use of an online learning algorithm to train Support Vector Machines with kernels designed for the classification of candidate pairs of related entities. We also report the results of the systematic experimental validation we have performed, using well known datasets in the area.

In this paper, we empirically evaluate an automated technique, based on a hierarchical representation for the Earth’s surface and leveraging linear classifiers, for assigning geospatial coordinates to previously unseen documents, using only the raw text as input evidence. We measured the results obtained with models based on Support Vector Machines, over collections of geo-referenced Wikipedia articles in four different languages, namely English, German, Spanish and Portuguese. The best performing models obtained state-of-the-art results, corresponding to an average prediction error of 83 Kilometers, and a median error of just 9 Kilometers, in the case of the English Wikipedia collection.

The LocWeb 2014 workshop continues a successful workshop series at the intersection of geospatial search, information management, and Web architecture with a focus towards location-aware information access. The workshop reflects a multitude of fields that demand and utilize location features, featuring presentations that look at the topic of location on the Web from an interdisciplinary perspective, including new approaches dealing with or utilizing geospatial information.

The increasing popularity of smartphones and position tracking devices fostered the interest on human trajectory data analysis, and enabled the collection of large amounts of movement data. However, there are still several open challenges regarding the visual and interactive analysis of these data, as more people, often non-experienced, start using these methods. This paper presents a comparative study between two common visualization techniques for trajectory data analysis, namely a space-time cube and a static map, taking into consideration high-level visualization and analysis tasks. The results obtained with two similar prototypes support the adequacy of both methods for simple analysis tasks, and reveal that static maps are more adequate for locate tasks, while space-time cubes provide a better support for association tasks.

This paper describes our participation in the message polarity classification task of SemEval 2014. We focused on exploiting unlabeled data to improve accuracy, combining features leveraging word representations with other, more common features, based on word tokens or lexicons. We analyse the contribution of the different features, concluding that unlabeled data yields significant improvements.

With the prevalence of mobile computing systems and location based services, the research interest on spatiotemporal data has significantly increased, as evidenced by the collection of huge amounts of movement data. Consequently, this type of data raises several issues, namely in the research area of geographic information visualization. Despite the existence of several visual analysis techniques for the exploration of movement data, it is still unclear how usable and useful these techniques are, how can they be improved, and for which situations are these techniques most suitable. In this paper, we present current open challenges on the visual analysis of movement data, and the Ph.D work in progress aiming to address these problems. Our work will explore several factors that may affect the users' performance, and, based on those factors we will propose a taxonomy and an evaluation framework covering different tasks and techniques.