The General Data Protection Regulation enforced on the 25th of may in 2018 is a response to the growing importance of IT in today’s society, accompanied by public demand for control over personal data. In contrast to the previous directive, the new regulation applies to personal data stored in an unstructured format, such as email, rather than solely structured data. Companies are now forced to accommodate to this change, among others, in order to be compliant. This study aims to provide a code of conduct for the processing of personal data in email as a measure for reaching compliance. Furthermore, this study investigates whether Named Entity Recognition (NER) can aid this process as a means of finding personal data in the form of names. A literature review of current research and recommendations was conducted for the code of conduct proposal. A NER system was constructed using a hybrid approach with Binary Logistic Regression, hand-crafted rules and gazetteers. The model was applied to a selection of emails, including attachments, obtained from a small consultancy company in the automotive industry. The proposed code of conduct consists of six items, applied to the consultancy firm. The NER-model demonstrated low ability to identify names and was therefore deemed insufficient for this task. / Dataskyddsförordningen började gälla den 25e maj 2018, och uppstod som ett svar på den okände betydelsen av IT i dagens samhälle samt allmänhetens krav på ökad kontroll över personuppgifter för den enskilde individen. Till skillnad från det tidigare direktivet, omfattar den nya förordningen även personuppgifter som är lagrad i ostrukturerad form, som till exempel e-post, snarare än endast i strukturerad form. Många företag tvingas därmed att anpassa sig efter detta, tillsammans med ett flertal andra nya krav, i syfte att efterfölja förordningen. Den här studien syftar till att lägga fram ett förslag på en uppförandekod för behandling av personuppgifter i e-post som ett verktyg för att nå medgörlighet. Utöver detta undersöks det om Named Entity Recognition (NER) kan användas som ett hjälpmedel vid identifiering av personuppgifter, mer specifikt namn. En litteraturstudie kring tidigare forskning och aktuella rekommendationer utfördes inför utformningen av uppförandekoden. Ett NER-system konstruerades med hjälp av Binär Logistisk Regression, handgjorda regler och ordlistor. Modellen applicerades på ett urval av e-postmeddelanden, med eventuella bilagor, som tillhandahölls från ett litet konsultbolag aktivt inom bilindustrin. Den rekommenderade uppförandekoden består av sex punkter, applicerade på konsultbolaget. NER-modellen påvisade en låg förmåga att identifiera namn och ansågs därför inte vara lämplig för den utsatta uppgiften.

Structured Product Labels (SPLs), the package inserts that accompany drugs governed by the Food and Drugs Administration (FDA), hold information about Adverse Drug Reactions (ADRs) that exists associated with drugs post-market. This information is valuable for actors working in the field of pharmacovigilance aiming to improve the safety of drugs. One such actor is Uppsala Monitoring Centre (UMC), a non-profit conducting pharmacovigilance research. In order to access the valuable information of the package inserts, UMC have constructed an SPL mining pipeline in order to mine SPLs for ADRs. This project aims to investigate new approaches to the solution to the Scan problem, the part of the pipeline responsible for extracting mentions of ADRs. The Scan problem is solved by approaching the problem as a Named Entity Recognition task, a subtask of Natural Language Processing. By using the transformer-based deep learning model BERT, with domain-specific pre-training, an F1-score of 0.8220 was achieved. Furthermore, the chosen model was used in an iteration of Active Learning in order to efficiently extend the available data pool with the most informative examples. Active Learning improved the F1-score to 0.8337. However, the Active Learning was benchmarked against a data set extended with random examples, showing similar improved scores, therefore this application of Active Learning could not be determined to be effective in this project.

[pt] A Extração de Informação é uma tarefa importante no domínio jurídico. Embora a presença de dados estruturados seja escassa, dados não estruturados na forma de documentos jurídicos, como sentenças, estão amplamente disponíveis. Se processados adequadamente, tais documentos podem fornecer informações valiosas sobre processos judiciais anteriores, permitindo uma melhor avaliação por profissionais do direito e apoiando aplicativos baseados em dados. Este estudo aborda a Extração de Informação no domínio jurídico, extraindo valor de sentenças relacionados a reclamações de consumidores. Mais especificamente, a extração de cláusulas categóricas é abordada através de classificação, onde seis modelos baseados em diferentes estruturas são analisados. Complementarmente, a extração de valores monetários relacionados a indenizações por danos morais é abordada por um modelo de Reconhecimento de Entidade Nomeada. Para avaliação, um conjunto de dados foi criado, contendo 964 sentenças anotados manualmente (escritas em português) emitidas por juízes de primeira instância. Os resultados mostram uma média de aproximadamente 97 por cento de acurácia na extração de cláusulas categóricas, e 98,9 por cento na aplicação de NER para a extração de indenizações por danos morais. / [en] Information Extraction is an important task in the legal domain. While the presence of structured and machine-processable data is scarce, unstructured data in the form of legal documents, such as legal opinions, is largely available. If properly processed, such documents can provide valuable information with regards to past lawsuits, allowing better assessment by legal professionals and supporting data-driven applications. This study addresses Information Extraction in the legal domain by extracting value from legal opinions related to consumer complaints. More specifically, the extraction of categorical provisions is addressed by classification, where six models based on different frameworks are analyzed. Moreover, the extraction of monetary values related to moral damage compensations is addressed by a Named Entity Recognition (NER) model. For evaluation, a dataset was constructed, containing 964 manually annotated legal opinions (written in Brazilian Portuguese) enacted by lower court judges. The results show an average of approximately 97 percent of accuracy when extracting categorical provisions, and 98.9 percent when applying NER for the extraction of moral damage compensations.

As part of our work on the NSF funded Integrated Digital Event Archiving and Library (IDEAL) project and the Global Event and Trend Archive Research (GETAR) project, we collected over 1.4 billion tweets using over 1,000 keywords, key phrases, mentions, or hashtags, starting from 2009. Since many tweets talk about events (with useful location information), such as natural disasters, emergencies, and accidents, it is important to geo-locate those tweets whenever possible. Due to possible location ambiguity, finding a tweet's location often is challenging. Many distinct places have the same geoname, e.g., "Greenville" matches 50 different locations in the U.S.A. Frequently, in tweets, explicit location information, like geonames mentioned, is insufficient, because tweets are often brief and incomplete. They have a small fraction of the full location information of an event due to the 140 character limitation. Location indicative words (LIWs) may include latent location information, for example, "Water main break near White House" does not have any geonames but it is related to a location "1600 Pennsylvania Ave NW, Washington, DC 20500 USA" indicated by the key phrase 'White House'. To disambiguate tweet locations, we first extracted geospatial named entities (geonames) and predicted implicit state (e.g., Virginia or California) information from entities using machine learning algorithms including Support Vector Machine (SVM), Naive Bayes (NB), and Random Forest (RF). Implicit state information helps reduce ambiguity. We also studied how location information of events is expressed in tweets and how latent location indicative information can help to geo-locate tweets. We then used a machine learning (ML) approach to predict the implicit state using geonames and LIWs. We conducted experiments with tweets (e.g., about potholes), and found significant improvement in disambiguating tweet locations using a ML algorithm along with the Stanford NER. Adding state information predicted by our classifiers increased the possibility to find the state-level geo-location unambiguously by up to 80%. We also studied over 6 million tweets (3 mid-size and 2 big-size collections about water main breaks, sinkholes, potholes, car crashes, and car accidents), covering 17 months. We found that up to 91.1% of tweets have at least one type of location information (geo-coordinates or geonames), or LIWs. We also demonstrated that in most cases adding LIWs helps geo-locate tweets with less ambiguity using a geo-coding API. Finally, we conducted additional experiments with the five different tweet collections, and found significant improvement in disambiguating tweet locations using a ML approach with geonames and all LIWs that are present in tweet texts as features. / Ph. D. / As part of our work on the projects “Integrated Digital Event Archiving and Library (IDEAL)” and “Global Event and Trend Archive Research (GETAR),” funded by NSF, we collected over 1.4 billion tweets using over 1,000 keywords, key phrases, mentions, or hashtags, starting from 2009. Since many tweets talk about events (with useful location information), such as natural disasters, emergencies, and accidents, it is important to geolocate those tweets whenever possible. Due to possible location ambiguity, finding a tweet’s location often is challenging. Many distinct places have the same geoname, e.g., “Greenville” matches 50 different locations in the U.S.A. Frequently, in tweets, explicit location information, like geonames mentioned, is insufficient, because tweets are often brief and incomplete. They have a small fraction of the full location information of an event due to the 140 character limitation. Location indicative words (LIWs) may include latent location information, for example, “Water main break near White House” does not have any geonames but it is related to a location “1600 Pennsylvania Ave NW, Washington, DC 20500 USA” indicated by the key phrase ‘White House’. To disambiguate tweet locations, we first extracted geonames, and then predicted implicit state (e.g., Virginia or California) information from entities using machine learning (ML) algorithms (wherein computers learn from examples what state is appropriate). Implicit state information helps reduce ambiguity. We also studied how location information of events is expressed in tweets and how latent location indicative information can help to geo-locate tweets. We then used a ML approach to predict the implicit state using geonames and LIWs. We conducted experiments with tweets (e.g., about potholes), and found significant improvement in disambiguating tweet locations using a ML algorithm along with the Stanford Named Entity Recognizer. Adding state information predicted by our classifiers increased the ability to find the state-level geo-location unambiguously by up to 80%. We also studied over 6 million tweets (in three mid-size and two big collections, about water main breaks, sinkholes, potholes, car crashes, and car accidents), covering 17 months. We found that up to 91.1% of tweets have at least one type of location information (geocoordinates or geonames), or LIWs. We also demonstrated that in most cases adding LIWs helps geo-locate tweets with less ambiguity using a geo-coding Web application (that converts addresses into geographic coordinates). Finally, we conducted additional experiments with the five different tweet collections, and found significant improvement in disambiguating tweet locations using a ML approach wherein the features considered are the geonames and all LIWs that are present in the tweet texts.

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal.

Cette thèse explique et présente notre démarche de la réalisation d’un système à base de règles de reconnaissance et de classification automatique des EN en arabe. C’est un travail qui implique deux disciplines : la linguistique et l’informatique. L’outil informatique et les règles la linguistiques s’accouplent pour donner naissance à une nouvelle discipline ; celle de « traitement automatique des langues », qui opère sur des niveaux différents (morphosyntaxique, syntaxique, sémantique, syntactico-sémantique etc.). Nous avons donc, dans ce qui nous concerne, mis en œuvre des informations et règles linguistiques nécessaires au service du logiciel informatique, qui doit être en mesure de les appliquer, pour extraire et classifier, par des annotations syntaxiques et/ou sémantiques, les différentes classes d’entités nommées.Ce travail de thèse s’inscrit donc dans un cadre général de traitement automatique des langues, mais plus particulièrement dans la continuité des travaux réalisés au niveau de l’analyse morphosyntaxique par la conception et la réalisation des bases des données lexicales SAMIA et ensuite DIINAR avec l’ensemble de résultats de recherches qui en découlent. C’est une tâche qui vise à l’enrichissement lexical par des entités nommées simples et complexes, et qui veut établir la transition de l’analyse morphosyntaxique vers l’analyse syntaxique, et syntatico-sémantique dans une visée plus générale de l’analyse du contenu textuel. Pour comprendre de quoi il s’agit, il nous était important de commencer par la définition de l’entité nommée. Et pour mener à bien notre démarche, nous avons distingué entre deux types principaux : pur nom propre et EN descriptive. Nous avons aussi établi une classification référentielle en se basant sur diverses classes et sous-classes qui constituent la référence de nos annotations sémantiques. Cependant, nous avons dû faire face à deux difficultés majeures : l’ambiguïté lexicale et les frontières des entités nommées complexes. Notre système adopte une approche à base de règles syntactico-sémantiques. Il est constitué, après le Niveau 0 d’analyse morphosyntaxique, de cinq niveaux de construction de patrons syntaxiques et syntactico-sémantiques basés sur les informations linguistique nécessaires (morphosyntaxiques, syntaxiques, sémantique, et syntactico-sémantique). Ce travail, après évaluation en utilisant deux corpus, a abouti à de très bons résultats en précision, en rappel et en F–mesure. Les résultats de notre système ont un apport intéressant dans différents application du traitement automatique des langues notamment les deux tâches de recherche et d’extraction d’informations. En effet, on les a concrètement exploités dans les deux applications (recherche et extraction d’informations). En plus de cette expérience unique, nous envisageons par la suite étendre notre système à l’extraction et la classification des phrases dans lesquelles, les entités classifiées, principalement les entités nommées et les verbes, jouent respectivement le rôle d’arguments et de prédicats. Un deuxième objectif consiste à l’enrichissement des différents types de ressources lexicales à l’instar des ontologies. / This thesis explains and presents our approach of rule-based system of arabic named entity recognition and classification. This work involves two disciplines : linguistics and computer science. Computer tools and linguistic rules are merged to give birth to a new discipline : Natural Languge Processsing, which operates in different levels (morphosyntactic, syntactic, semantic, syntactico-semantic…). So, in our particular case, we have put the necessary linguistic information and rules to software sevice. This later should be able to apply and implement them in order to recognise and classify, by syntactic and semantic annotations, the different named entity classes.This work of thesis is incorporated within the general domain of natural language processing, but it particularly falls within the scope of the continuity of the accomplished work in terms of morphosyntactic analysis and the realisation of lexical data bases of SAMIA and then DIINAR as well as the accompanying scientific recearch. This task aimes at lexical enrichement with simple and complex named entities and at establishing the transition from the morphological analysis into syntactic and syntactico-semantic analysis. The ultimate objective is text analysis. To understand what it is about, it was important to start with named entity definition. To carry out this task, we distinguished between two main named entity types : pur proper name and descriptive named entities. We have also established a referential classification on the basis of different classes and sub-classes which constitue the reference for our semantic annotations. Nevertheless, we are confronted with two major difficulties : lexical ambiguity and the frontiers of complex named entities. Our system adoptes a syntactico-semantic rule-based approach. After Level 0 of morpho-syntactic analysis, the system is made up of five levels of syntactic and syntactico-semantic patterns based on tne necessary linguisic information (i.e. morphosyntactic, syntactic, semantic and syntactico-semantic information).This work has obtained very good results in termes of precision, recall and F-measure. The output of our system has an interesting contribution in different applications of the natural language processing especially in both tasks of information retrieval and information extraction. In fact, we have concretely exploited our system output in both applications (information retrieval and information extraction). In addition to this unique experience, we envisage in the future work to extend our system into the sentence extraction and classification, in which classified entities, mainly named entities and verbs, play respectively the role of arguments and predicates. The second objective consists in the enrichment of different types of lexical resources such as ontologies.

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal

This thesis deals with mining of relevant information from documents and automatic splitting of multiple documents merged together. Moreover, it describes the design and implementation of software for data mining from documents and for automatic splitting of multiple documents. Methods for acquiring textual data from scanned documents, named entity recognition, document clustering, their supportive algorithms and metrics for automatic splitting of documents are described in this thesis. Furthermore, an algorithm of implemented software is explained and tools and techniques used by this software are described. Lastly, the success rate of the implemented software is evaluated. In conclusion, possible extensions and further development of this thesis are discussed at the end.

L’avènement d’internet suivi des réseaux sociaux a permis un accès facile et une diffusion rapide de l’information par toute personne disposant d’une connexion internet. L’une des conséquences néfastes de cela est la propagation de fausses informations appelées «fake news». Les fake news représentent aujourd’hui un enjeu majeur au regard de ces conséquences. De nombreuses personnes affirment encore aujourd’hui que sans la diffusion massive de fake news sur Hillary Clinton lors de la campagne présidentielle de 2016, Donald Trump n’aurait peut-être pas été le vainqueur de cette élection. Le sujet de ce mémoire concerne donc la détection automatique des fake news. De nos jours, il existe un grand nombre de travaux à ce sujet. La majorité des approches présentées se basent soit sur l’exploitation du contenu du texte d’entrée, soit sur le contexte social du texte ou encore sur un mélange entre ces deux types d’approches. Néanmoins, il existe très peu d’outils ou de systèmes efficaces qui détecte une fausse information dans la vie réelle, tout en incluant l’évolution de l’information au cours du temps. De plus, il y a un manque criant de systèmes conçues dans le but d’aider les utilisateurs des réseaux sociaux à adopter un comportement qui leur permettrait de détecter les fausses nouvelles. Afin d’atténuer ce problème, nous proposons un système appelé DS-Fake. À notre connaissance, ce système est le premier à inclure l’exploration de flux de données. Un flux de données est une séquence infinie et dénombrable d’éléments et est utilisée pour représenter des données rendues disponibles au fil du temps. DS-Fake explore à la fois l’entrée et le contenu d’un flux de données. L’entrée est une publication sur Twitter donnée au système afin qu’il puisse déterminer si le tweet est digne de confiance. Le flux de données est extrait à l’aide de techniques d’extraction du contenu de sites Web. Le contenu reçu par ce flux est lié à l’entrée en termes de sujets ou d’entités nommées mentionnées dans le texte d’entrée. DS-Fake aide également les utilisateurs à développer de bons réflexes face à toute information qui se propage sur les réseaux sociaux. DS-Fake attribue un score de crédibilité aux utilisateurs des réseaux sociaux. Ce score décrit la probabilité qu’un utilisateur puisse publier de fausses informations. La plupart des systèmes utilisent des caractéristiques comme le nombre de followers, la localisation, l’emploi, etc. Seuls quelques systèmes utilisent l’historique des publications précédentes d’un utilisateur afin d’attribuer un score. Pour déterminer ce score, la majorité des systèmes utilisent la moyenne. DS-Fake renvoie un pourcentage de confiance qui détermine la probabilité que l’entrée soit fiable. Contrairement au petit nombre de systèmes qui utilisent l’historique des publications en ne prenant pas en compte que les tweets précédents d’un utilisateur, DS-Fake calcule le score de crédibilité sur la base des tweets précédents de tous les utilisateurs. Nous avons renommé le score de crédibilité par score de légitimité. Ce dernier est basé sur la technique de la moyenne Bayésienne. Cette façon de calculer le score permet d’atténuer l’impact des résultats des publications précédentes en fonction du nombre de publications dans l’historique. Un utilisateur donné ayant un plus grand nombre de tweets dans son historique qu’un autre utilisateur, même si les tweets des deux sont tous vrais, le premier utilisateur est plus crédible que le second. Son score de légitimité sera donc plus élevé. À notre connaissance, ce travail est le premier qui utilise la moyenne Bayésienne basée sur l’historique de tweets de toutes les sources pour attribuer un score à chaque source. De plus, les modules de DS-Fake ont la capacité d’encapsuler le résultat de deux tâches, à savoir la similarité de texte et l’inférence en langage naturel hl(en anglais Natural Language Inference). Ce type de modèle qui combine ces deux tâches de TAL est également nouveau pour la problématique de la détection des fake news. DS-Fake surpasse en termes de performance toutes les approches de l’état de l’art qui ont utilisé FakeNewsNet et qui se sont basées sur diverses métriques. Il y a très peu d’ensembles de données complets avec une variété d’attributs, ce qui constitue un des défis de la recherche sur les fausses nouvelles. Shu et al. ont introduit en 2018 l’ensemble de données FakeNewsNet pour résoudre ce problème. Le score de légitimité et les tweets récupérés ajoutent des attributs à l’ensemble de données FakeNewsNet. / The advent of the internet, followed by online social networks, has allowed easy access and rapid propagation of information by anyone with an internet connection. One of the harmful consequences of this is the spread of false information, which is well-known by the term "fake news". Fake news represent a major challenge due to their consequences. Some people still affirm that without the massive spread of fake news about Hillary Clinton during the 2016 presidential campaign, Donald Trump would not have been the winner of the 2016 United States presidential election. The subject of this thesis concerns the automatic detection of fake news. Nowadays, there is a lot of research on this subject. The vast majority of the approaches presented in these works are based either on the exploitation of the input text content or the social context of the text or even on a mixture of these two types of approaches. Nevertheless, there are only a few practical tools or systems that detect false information in real life, and that includes the evolution of information over time. Moreover, no system yet offers an explanation to help social network users adopt a behaviour that will allow them to detect fake news. In order to mitigate this problem, we propose a system called DS-Fake. To the best of our knowledge, this system is the first to include data stream mining. A data stream is a sequence of elements used to represent data elements over time. This system explores both the input and the contents of a data stream. The input is a post on Twitter given to the system that determines if the tweet can be trusted. The data stream is extracted using web scraping techniques. The content received by this flow is related to the input in terms of topics or named entities mentioned in the input text. This system also helps users develop good reflexes when faced with any information that spreads on social networks. DS-Fake assigns a credibility score to users of social networks. This score describes how likely a user can publish false information. Most of the systems use features like the number of followers, the localization, the job title, etc. Only a few systems use the history of a user’s previous publications to assign a score. To determine this score, most systems use the average. DS-Fake returns a percentage of confidence that determines how likely the input is reliable. Unlike the small number of systems that use the publication history by taking into account only the previous tweets of a user, DS-Fake calculates the credibility score based on the previous tweets of all users. We renamed the credibility score legitimacy score. The latter is based on the Bayesian averaging technique. This way of calculating the score allows attenuating the impact of the results from previous posts according to the number of posts in the history. A user who has more tweets in his history than another user, even if the tweets of both are all true, the first user is more credible than the second. His legitimacy score will therefore be higher. To our knowledge, this work is the first that uses the Bayesian average based on the post history of all sources to assign a score to each source. DS-Fake modules have the ability to encapsulate the output of two tasks, namely text similarity and natural language inference. This type of model that combines these two NLP tasks is also new for the problem of fake news detection. There are very few complete datasets with a variety of attributes, which is one of the challenges of fake news research. Shu et al. introduce in 2018 the FakeNewsNet dataset to tackle this issue. Our work uses and enriches this dataset. The legitimacy score and the retrieved tweets from named entities mentioned in the input texts add features to the FakeNewsNet dataset. DS-Fake outperforms all state-of-the-art approaches that have used FakeNewsNet and that are based on various metrics.

This master's thesis describes Semantic Enrichment Component (SEC), that searches entities (e.g., persons or places) in the input text document and returns information about them. The goals of this component are to create a single interface for named entity recognition tools, to enable parallel document processing, to save memory while using the knowledge base, and to speed up access to its content. To achieve these goals, the output of the named entity recognition tools in the text was specified, the tool for storing the preprocessed knowledge base into the shared memory was implemented, and the client-server scheme was used to create the component.