SammanfattningI  denna  studie  utforskas  kapaciteterna  och  begränsningarna  hos  avancerade  stora språkmodeller (SSM), med särskilt fokus på ChatGPT-4 och Google Bard. Studien inleds med att ge en historisk bakgrund till artificiell intelligens och hur denna utveckling har lett fram till skapandet av dessa modeller. Därefter genomförs en kritisk analys av deras prestanda i språkbehandling och problemlösning. Genom att evaluera deras effektivitet i hanteringen av nyhetsinnehåll och sociala medier, samt i utförandet av kreativa uppgifter som pussel, belyses deras förmåga inom språklig bearbetning samt de utmaningar de möter i att förstå nyanser och utöva kreativt tänkande.I denna studie framkom det att SSM har en avancerad förmåga att förstå och reagera på komplexa språkstrukturer. Denna förmåga är dock inte utan begränsningar, speciellt när det kommer till uppgifter som kräver en noggrann bedömning för att skilja mellan sanning och osanning. Denna observation lyfter fram en kritisk aspekt av SSM:ernas nuvarande kapacitet, de är effektiva inom många områden, men möter fortfarande utmaningar i att hantera de finare nyanserna i mänskligt språk och tänkande. Studiens resultat betonar även vikten av mänsklig tillsyn vid användning av artificiell intelligens (AI), vilket pekar på behovet av att ha realistiska förväntningar på AI:s kapacitet och betonar vidare betydelsen av en ansvarsfull utveckling  av  AI,  där  en  noggrann  uppmärksamhet  kring etiska  aspekter  är  central.  En kombination av mänsklig intelligens och AI föreslås som en lösning för att hantera komplexa utmaningar, vilket bidrar till en fördjupad förståelse av avancerade språkmodellers dynamik och deras roll inom AI:s bredare utveckling och tillämpning.

Les modèles de réponse aux questions ont montré des résultats impressionnants sur plusieurs ensembles de données et tâches de réponse aux questions. Cependant, lorsqu'ils sont testés sur des ensembles de données hors domaine, la performance diminue. Afin de contourner l'annotation manuelle des données d'entraînement du nouveau domaine, des paires de questions-réponses peuvent être générées synthétiquement à partir de données non annotées. Dans ce travail, nous nous intéressons à la génération de données synthétiques et nous testons différentes méthodes de traitement du langage naturel pour les deux étapes de création d'ensembles de données : génération de questions et génération de réponses. Nous utilisons les ensembles de données générés pour entraîner les modèles UnifiedQA et Bert-QA et nous les testons sur SCIQ, un ensemble de données hors domaine sur la physique, la chimie et la biologie pour la tâche de question-réponse à choix multiples, ainsi que sur HotpotQA, TriviaQA, NatQ et SearchQA, quatre ensembles de données hors domaine pour la tâche de question-réponse. Cette procédure nous permet d'évaluer et de comparer les méthodes basées sur des règles avec les méthodes de réseaux neuronaux. Nous montrons que les méthodes basées sur des règles produisent des résultats supérieurs pour la tâche de question-réponse à choix multiple, mais que les méthodes de réseaux neuronaux produisent généralement des meilleurs résultats pour la tâche de question-réponse. Par contre, nous observons aussi qu'occasionnellement, les méthodes basées sur des règles peuvent compléter les méthodes de réseaux neuronaux et produire des résultats compétitifs lorsqu'on entraîne Bert-QA avec les bases de données synthétiques provenant des deux méthodes. / Question Answering models have shown impressive results in several question answering datasets and tasks. However, when tested on out-of-domain datasets, the performance decreases. In order to circumvent manually annotating training data from the new domain, question-answer pairs can be generated synthetically from unnanotated data. In this work, we are interested in the generation of synthetic data and we test different Natural Language Processing methods for the two steps of dataset creation: question/answer generation. We use the generated datasets to train QA models UnifiedQA and Bert-QA and we test it on SCIQ, an out-of-domain dataset about physics, chemistry, and biology for MCQA, and on HotpotQA, TriviaQA, NatQ and SearchQA, four out-of-domain datasets for QA. This procedure allows us to evaluate and compare rule-based methods with neural network methods. We show that rule-based methods yield superior results for the multiple-choice question-answering task, but neural network methods generally produce better results for the question-answering task. However, we also observe that occasionally, rule-based methods can complement neural network methods and produce competitive results when training Bert-QA with synthetic databases derived from both methods.

Tesis por compendio / [ES] En la actualidad, el mundo político tiene tanto o más impacto en la sociedad que ésta en el mundo político. Los líderes o representantes de partidos políticos hacen uso de su poder en los medios de comunicación, para modificar posiciones ideológicas y llegar al pueblo con el objetivo de ganar popularidad en las elecciones gubernamentales.A través de un lenguaje engañoso, los textos políticos pueden contener sesgos partidistas y sociales que minan la percepción de la realidad. Como resultado, los seguidores de una ideología, o miembros de una categoría social, se sienten amenazados por otros grupos sociales o ideológicos, o los perciben como competencia, derivándose así una polarización política con agresiones físicas y verbales. La comunidad científica del Procesamiento del Lenguaje Natural (NLP, según sus siglas en inglés) contribuye cada día a detectar discursos de odio, insultos, mensajes ofensivos, e información falsa entre otras tareas computacionales que colindan con ciencias sociales. Sin embargo, para abordar tales tareas, es necesario hacer frente a diversos problemas entre los que se encuentran la dificultad de tener textos etiquetados, las limitaciones de no trabajar con un equipo interdisciplinario, y los desafíos que entraña la necesidad de soluciones interpretables por el ser humano. Esta tesis se enfoca en la detección de sesgos partidistas y sesgos sociales, tomando como casos de estudio el hiperpartidismo y los estereotipos sobre inmigrantes. Para ello, se propone un modelo basado en una técnica de enmascaramiento de textos capaz de detectar lenguaje engañoso incluso en temas controversiales, siendo capaz de capturar patrones del contenido y el estilo de escritura. Además, abordamos el problema usando modelos basados en BERT, conocidos por su efectividad al capturar patrones sintácticos y semánticos sobre las mismas representaciones de textos. Ambos enfoques, la técnica de enmascaramiento y los modelos basados en BERT, se comparan en términos de desempeño y explicabilidad en la detección de hiperpartidismo en noticias políticas y estereotipos sobre inmigrantes. Para la identificación de estos últimos, se propone una nueva taxonomía con fundamentos teóricos en sicología social, y con la que se etiquetan textos extraídos de intervenciones partidistas llevadas a cabo en el Parlamento español. Los resultados muestran que los enfoques propuestos contribuyen al estudio del hiperpartidismo, así como a identif i car cuándo los ciudadanos y políticos enmarcan a los inmigrantes en una imagen de víctima, recurso económico, o amenaza. Finalmente, en esta investigación interdisciplinaria se demuestra que los estereotipos sobre inmigrantes son usados como estrategia retórica en contextos políticos. / [CA] Avui, el món polític té tant o més impacte en la societat que la societat en el món polític. Els líders polítics, o representants dels partits polítics, fan servir el seu poder als mitjans de comunicació per modif i car posicions ideològiques i arribar al poble per tal de guanyar popularitat a les eleccions governamentals. Mitjançant un llenguatge enganyós, els textos polítics poden contenir biaixos partidistes i socials que soscaven la percepció de la realitat. Com a resultat, augmenta la polarització política nociva perquè els seguidors d'una ideologia, o els membres d'una categoria social, veuen els altres grups com una amenaça o competència, que acaba en agressions verbals i físiques amb resultats desafortunats. La comunitat de Processament del llenguatge natural (PNL) té cada dia noves aportacions amb enfocaments que ajuden a detectar discursos d'odi, insults, missatges ofensius i informació falsa, entre altres tasques computacionals relacionades amb les ciències socials. No obstant això, molts obstacles impedeixen eradicar aquests problemes, com ara la dif i cultat de tenir textos anotats, les limitacions dels enfocaments no interdisciplinaris i el repte afegit per la necessitat de solucions interpretables. Aquesta tesi se centra en la detecció de biaixos partidistes i socials, prenent com a cas pràctic l'hiperpartidisme i els estereotips sobre els immigrants. Proposem un model basat en una tècnica d'emmascarament que permet detectar llenguatge enganyós en temes polèmics i no polèmics, capturant pa-trons relacionats amb l'estil i el contingut. A més, abordem el problema avaluant models basats en BERT, coneguts per ser efectius per capturar patrons semàntics i sintàctics en la mateixa representació. Comparem aquests dos enfocaments (la tècnica d'emmascarament i els models basats en BERT) en termes de rendiment i les seves solucions explicables en la detecció de l'hiperpartidisme en les notícies polítiques i els estereotips d'immigrants. Per tal d'identificar els estereotips dels immigrants, proposem una nova tax-onomia recolzada per la teoria de la psicologia social i anotem un conjunt de dades de les intervencions partidistes al Parlament espanyol. Els resultats mostren que els nostres models poden ajudar a estudiar l'hiperpartidisme i identif i car diferents marcs en què els ciutadans i els polítics perceben els immigrants com a víctimes, recursos econòmics o amenaces. Finalment, aquesta investigació interdisciplinària demostra que els estereotips dels immigrants s'utilitzen com a estratègia retòrica en contextos polítics. / [EN] Today, the political world has as much or more impact on society than society has on the political world. Political leaders, or representatives of political parties, use their power in the media to modify ideological positions and reach the people in order to gain popularity in government elections. Through deceptive language, political texts may contain partisan and social biases that undermine the perception of reality. As a result, harmful political polarization increases because the followers of an ideology, or members of a social category, see other groups as a threat or competition, ending in verbal and physical aggression with unfortunate outcomes. The Natural Language Processing (NLP) community has new contri-butions every day with approaches that help detect hate speech, insults, of f ensive messages, and false information, among other computational tasks related to social sciences. However, many obstacles prevent eradicating these problems, such as the dif f i culty of having annotated texts, the limitations of non-interdisciplinary approaches, and the challenge added by the necessity of interpretable solutions. This thesis focuses on the detection of partisan and social biases, tak-ing hyperpartisanship and stereotypes about immigrants as case studies. We propose a model based on a masking technique that can detect deceptive language in controversial and non-controversial topics, capturing patterns related to style and content. Moreover, we address the problem by evalu-ating BERT-based models, known to be ef f ective at capturing semantic and syntactic patterns in the same representation. We compare these two approaches (the masking technique and the BERT-based models) in terms of their performance and the explainability of their decisions in the detection of hyperpartisanship in political news and immigrant stereotypes. In order to identify immigrant stereotypes, we propose a new taxonomy supported by social psychology theory and annotate a dataset from partisan interventions in the Spanish parliament. Results show that our models can help study hyperpartisanship and identify dif f erent frames in which citizens and politicians perceive immigrants as victims, economic resources, or threat. Finally, this interdisciplinary research proves that immigrant stereotypes are used as a rhetorical strategy in political contexts. / This PhD thesis was funded by the MISMIS-FAKEnHATE research project (PGC2018-096212-B-C31) of the Spanish Ministry of Science and Innovation. / Sánchez Junquera, JJ. (2022). Detecting Deception, Partisan, and Social Biases [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/185784 / Compendio

Classic literature is rich, be it linguistically, historically, or culturally, making it valuable for future studies. Consequently, this project chose a set of 48 classic books to conduct a stylometric analysis on the defined set of books, adopting an approach used by a related work to divide the books into text segments, quantify the resulting text segments, and analyze the books using the quantified values to understand the linguistic attributes of the books. Apart from the latter, this project conducted different classification tasks for other objectives. In one respect, the study used the quantified values of the text segments of the books for classification tasks using advanced models like LightGBM and TabNet to assess the application of this approach in authorship attribution. From another perspective, the study utilized a State-Of-The-Art model, namely, RoBERTa for classification tasks using the segmented texts of the books instead to evaluate the performance of the model in authorship attribution. The results uncovered the characteristics of the books to a reasonable degree. Regarding the authorship attribution tasks, the results suggest that segmenting and quantifying text using stylometric analysis and supervised machine learning algorithms is practical in such tasks. This approach, while showing promise, may still require further improvements to achieve optimal performance. Lastly, RoBERTa demonstrated high performance in authorship attribution tasks.

Many, if not all people have online social accounts (OSAs) on an online community (OC) such as Facebook (Meta), Twitter (X), Instagram (Meta), Mastodon, Nostr. OCs enable quick and easy interaction with friends, family, and even online communities to share information about. There is also a dark side to Ocs, where users with malicious intent join OC platforms with the purpose of criminal activities such as spreading fake news/information, cyberbullying, propaganda, phishing, stealing, and unjust enrichment. These criminal activities are especially concerning when harming minors. Detection and mitigation are needed to protect and help OCs and stop these criminals from harming others. Many solutions exist; however, they are typically focused on a single category of malicious intent detection rather than an all-encompassing solution. To answer this challenge, we propose the first steps of a framework for analyzing and identifying malicious intent in OCs that we refer to as malicious mntent detection framework (MIDF). MIDF is an extensible proof-of-concept that uses machine learning techniques to enable detection and mitigation. The framework will first be used to detect malicious users using solely relationships and then can be leveraged to create a suite of malicious intent vector detection models, including phishing, propaganda, scams, cyberbullying, racism, spam, and bots for open-source online social networks, such as Mastodon, and Nostr.

[ES] El triaje de los incidentes de urgencias y emergencias extrahospitalarias representa un reto difícil, debido a las limitaciones temporales y a la incertidumbre. Además, errores en este proceso pueden tener graves consecuencias para los pacientes. Por lo tanto, cualquier herramienta o estrategia novedosa que mejore estos procesos ofrece un valor sustancial en términos de atención al paciente y gestión global de los incidentes. La hipótesis en la que se basa esta tesis es que el Aprendizaje Automático, concretamente el Aprendizaje Profundo, puede mejorar estos procesos proporcionando estimaciones de la gravedad de los incidentes, mediante el análisis de millones de datos derivados de llamadas de emergencia de la Comunitat Valenciana (España) que abarcan desde 2009 hasta 2019. Por tanto, esta tesis profundiza en el diseño y desarrollo de modelos basados en Aprendizaje Profundo Multitarea que aprovechan los datos multimodales asociados a eventos de urgencias y emergencias extrahospitalarias. Nuestro objetivo principal era predecir si el incidente suponía una situación de riesgo vital, la demora admisible de la respuesta y si era competencia del sistema de emergencias o de atención primaria. Utilizando datos disponibles entre 2009 y 2012, se observaron mejoras sustanciales en las métricas macro F1, con ganancias del 12.5% para la clasificación de riesgo vital, del 17.5% para la demora en la respuesta y del 5.1% para la clasificación por jurisdicción, en comparación con el protocolo interno de triaje de la Comunidad Valenciana. Sin embargo, los sistemas, los protocolos de triaje y las prácticas operativas evolucionan de forma natural con el tiempo. Los modelos que mostraron un rendimiento excelente con el conjunto de datos inicial de 2009 a 2012 no demostraron la misma eficacia cuando se evaluaron con datos posteriores que abarcaban de 2014 a 2019. Estos últimos habían sufrido modificaciones en comparación con los anteriores, que dieron lugar a variaciones en las distribuciones de probabilidad, caracterizadas e investigadas meticulosamente en esta tesis. Continuando con nuestra investigación, nos centramos en la incorporación de técnicas de Aprendizaje Continuo Profundo en nuestros desarrollos. Gracias a ello, pudimos mitigar sustancialmente los efectos adversos consecuencia de los cambios distribucionales sobre el rendimiento. Los resultados indican que, si bien las fluctuaciones de rendimiento no se eliminan por completo, pueden mantenerse dentro de un rango manejable. En particular, con respecto a la métrica F1, cuando las variaciones distribucionales son ligeras o moderadas, el comportamiento se mantiene estable, sin variar más de un 2.5%. Además, nuestra tesis demuestra la viabilidad de construir herramientas auxiliares que permitan a los operadores interactuar con estos complejos modelos. En consecuencia, sin interrumpir el flujo de trabajo de los profesionales, se hace posible proporcionar retroalimentación mediante predicciones de probabilidad para cada clase de etiqueta de gravedad y tomar las medidas pertinentes. Por último, los resultados de esta tesis tienen implicaciones directas en la gestión de las urgencias y emergencias extrahospitalarias en la Comunidad Valenciana, al integrarse el modelo final resultante en los centros de atención de llamadas. Este modelo utilizará los datos proporcionados por los operadores telefónicos para calcular automáticamente las predicciones de gravedad, que luego se compararán con las generadas por el protocolo de triaje interno. Cualquier disparidad entre estas predicciones desencadenará la derivación del incidente a un coordinador médico, que supervisará su tratamiento. Por lo tanto, nuestra tesis, además de realizar importantes contribuciones al campo de la Investigación en Aprendizaje Automático Biomédico, también conlleva implicaciones sustanciales para mejorar la gestión de las urgencias y emergencias extrahospitalarias en el contexto de la Comunidad Valenciana. / [CA] El triatge dels incidents d'urgències i emergències extrahospitalàries representa un repte difícil, a causa de les limitacions temporals i de la incertesa. A més, els errors en aquest procés poden tindre greus conseqüències per als pacients. Per tant, qualsevol eina o estratègia innovadora que millore aquests processos ofereix un valor substancial en termes d'atenció al pacient i gestió global dels incidents. La hipòtesi en què es basa aquesta tesi és que l'Aprenentatge Automàtic, concretament l'Aprenentatge Profund, pot millorar significativament aquests processos proporcionant estimacions de la gravetat dels incidents, mitjançant l'anàlisi de milions de dades derivades de trucades d'emergència de la Comunitat Valenciana (Espanya) que abasten des de 2009 fins a 2019. Per tant, aquesta tesi aprofundeix en el disseny i desenvolupament de models basats en Aprenentatge Profund Multitasca que aprofiten dades multimodals d'incidents mèdics d'urgències i emergències extrahospitalàries. El nostre objectiu principal era predir si l'incident suposava una situació de risc vital, la demora admissible de la resposta i si era competència del sistema d'emergències o d'atenció primària. Utilitzant dades disponibles entre 2009 i 2012, es van observar millores substancials en les mètriques macro F1, amb guanys del 12.5% per a la classificació de risc vital, del 17.5% per a la demora en la resposta i del 5.1% per a la classificació per jurisdicció, en comparació amb el protocol intern de triatge de la Comunitat Valenciana. Tanmateix, els protocols de triatge i les pràctiques operatives evolucionen de forma natural amb el temps. Els models que van mostrar un rendiment excel·lent amb el conjunt de dades inicial de 2009 a 2012 no van demostrar la mateixa eficàcia quan es van avaluar amb dades posteriors que abastaven de 2014 a 2019. Aquestes últimes havien sofert modificacions en comparació amb les anteriors, que van donar lloc a variacions en les distribucions de probabilitat, caracteritzades i investigades minuciosament en aquesta tesi. Continuant amb la nostra investigació, ens vam centrar en la incorporació de tècniques d'Aprenentatge Continu als nostres desenvolupaments. Gràcies a això, vam poder mitigar substancialment els efectes adversos sobre el rendiment conseqüència dels canvis distribucionals. Els resultats indiquen que, si bé les fluctuacions de rendiment no s'eliminen completament al llarg del temps, poden mantenir-se dins d'un rang manejable. En particular, respecte a la mètrica F1, quan les variacions distribucionals són lleugeres o moderades, el comportament es manté estable, sense variar més d'un 2.5%. A més, la nostra tesi demostra la viabilitat de construir eines auxiliars que permeten als operadors interactuar amb aquests models complexos. En conseqüència, sense interrompre el flux de treball dels professionals, es fa possible proporcionar retroalimentació mitjançant prediccions de probabilitat per a cada classe d'etiqueta de gravetat i prendre les mesures pertinents. Finalment, els resultats d'aquesta tesi tenen implicacions directes en la gestió de les urgències i emergències extrahospitalàries a la Comunitat Valenciana, al integrar-se el model final resultant als centres d'atenció de telefonades. Aquest model utilitzarà les dades proporcionades pels operadors telefònics per calcular automàticament les prediccions de gravetat, que després es compararan amb les generades pel protocol de triatge intern. Qualsevol disparitat entre aquestes prediccions desencadenarà la derivació de l'incident a un coordinador mèdic, que supervisarà el seu tractament. Per tant, és evident que la nostra tesi, a més de realitzar importants contribucions al camp de la Investigació en Aprenentatge Automàtic Biomèdic, també comporta implicacions substancials per a millorar la gestió de les urgències i emergències extrahospitalàries en el context de la Comunitat Valenciana. / [EN] Triage for out-of-hospital emergency incidents represents a tough challenge, primarily due to time constraints and uncertainty. Furthermore, errors in this process can have severe consequences for patients. Therefore, any novel tool or strategy that enhances these processes can offer substantial value in terms of patient care and overall management of out-of-hospital emergency medical incidents. The hypothesis upon which this thesis is based is that Machine Learning, specifically Deep Learning, can improve these processes by providing estimations of the severity of incidents, by analyzing millions of data derived from emergency calls from the Valencian Region (Spain) spanning from 2009 to 2019. Hence, this thesis delves into designing and developing Deep Multitask Learning models that leverage multimodal out-of-hospital emergency medical data. Our primary objective was to predict whether the incident posed a life-threatening situation, the admissible response delay, and whether it fell under the jurisdiction of the emergency system or primary care. Using data available from 2009 to 2012, the results obtained were promising. We observed substantial improvements in macro F1-scores, with gains of 12.5% for life-threatening classification, 17.5% for response delay, and 5.1% for jurisdiction classification, compared to the in-house triage protocol of the Valencian Region. However, systems, dispatch protocols, and operational practices naturally evolve over time. Models that exhibited excellent performance with the initial dataset from 2009 to 2012 did not demonstrate the same efficacy when evaluated on data spanning from 2014 to 2019. This later dataset had undergone modifications compared to the earlier one, which led to dataset shifts, which we have meticulously characterized and investigated in this thesis. Continuing our research, we incorporated Deep Continual Learning techniques in our developments. As a result, we could substantially mitigate the adverse performance effects consequence of dataset shifts. The results indicate that, while performance fluctuations are not completely eliminated, they can be kept within a manageable range. In particular, with respect to the F1-score, when distributional variations fall within the light to moderate range, the performance remains stable, not varying by more than 2.5%. Furthermore, our thesis demonstrates the feasibility of building auxiliary tools that enable dispatchers to interact with these complex deep models. Consequently, without disrupting professionals' workflow, it becomes possible to provide feedback through probability predictions for each severity label class and take appropriate actions based on these predictions. Finally, the outcomes of this thesis hold direct implications for the management of out-of-hospital emergency medical incidents in the Valencian Region. The final model resulting from our research is slated for integration into the emergency medical dispatch centers of the Valencian Region. This model will utilize data provided by dispatchers to automatically compute severity predictions, which will then be compared with those generated by the in-house triage protocol. Any disparities between these predictions will trigger the referral of the incident to a physician coordinator, who will oversee its handling. Therefore, it is evident that our thesis, in addition to making significant contributions to the field of Biomedical Machine Learning Research, also carries substantial implications for enhancing the management of out-of-hospital emergencies in the context of the Valencian Region. / Ferri Borredà, P. (2024). Deep Continual Multimodal Multitask Models for Out-of-Hospital Emergency Medical Call Incidents Triage Support in the Presence of Dataset Shifts [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/203192

This thesis is a compendium of scientific works and engineering specifications that have been contributed to a large community of stakeholders to be copied, adapted, mixed, built upon and exploited in any way possible to achieve a common goal: Integrating Natural Language Processing (NLP) and Language Resources Using Linked Data The explosion of information technology in the last two decades has led to a substantial growth in quantity, diversity and complexity of web-accessible linguistic data. These resources become even more useful when linked with each other and the last few years have seen the emergence of numerous approaches in various disciplines concerned with linguistic resources and NLP tools. It is the challenge of our time to store, interlink and exploit this wealth of data accumulated in more than half a century of computational linguistics, of empirical, corpus-based study of language, and of computational lexicography in all its heterogeneity. The vision of the Giant Global Graph (GGG) was conceived by Tim Berners-Lee aiming at connecting all data on the Web and allowing to discover new relations between this openly-accessible data. This vision has been pursued by the Linked Open Data (LOD) community, where the cloud of published datasets comprises 295 data repositories and more than 30 billion RDF triples (as of September 2011). RDF is based on globally unique and accessible URIs and it was specifically designed to establish links between such URIs (or resources). This is captured in the Linked Data paradigm that postulates four rules: (1) Referred entities should be designated by URIs, (2) these URIs should be resolvable over HTTP, (3) data should be represented by means of standards such as RDF, (4) and a resource should include links to other resources. Although it is difficult to precisely identify the reasons for the success of the LOD effort, advocates generally argue that open licenses as well as open access are key enablers for the growth of such a network as they provide a strong incentive for collaboration and contribution by third parties. In his keynote at BNCOD 2011, Chris Bizer argued that with RDF the overall data integration effort can be “split between data publishers, third parties, and the data consumer”, a claim that can be substantiated by observing the evolution of many large data sets constituting the LOD cloud. As written in the acknowledgement section, parts of this thesis has received numerous feedback from other scientists, practitioners and industry in many different ways. The main contributions of this thesis are summarized here: Part I – Introduction and Background. During his keynote at the Language Resource and Evaluation Conference in 2012, Sören Auer stressed the decentralized, collaborative, interlinked and interoperable nature of the Web of Data. The keynote provides strong evidence that Semantic Web technologies such as Linked Data are on its way to become main stream for the representation of language resources. The jointly written companion publication for the keynote was later extended as a book chapter in The People’s Web Meets NLP and serves as the basis for “Introduction” and “Background”, outlining some stages of the Linked Data publication and refinement chain. Both chapters stress the importance of open licenses and open access as an enabler for collaboration, the ability to interlink data on the Web as a key feature of RDF as well as provide a discussion about scalability issues and decentralization. Furthermore, we elaborate on how conceptual interoperability can be achieved by (1) re-using vocabularies, (2) agile ontology development, (3) meetings to refine and adapt ontologies and (4) tool support to enrich ontologies and match schemata. Part II - Language Resources as Linked Data. “Linked Data in Linguistics” and “NLP & DBpedia, an Upward Knowledge Acquisition Spiral” summarize the results of the Linked Data in Linguistics (LDL) Workshop in 2012 and the NLP & DBpedia Workshop in 2013 and give a preview of the MLOD special issue. In total, five proceedings – three published at CEUR (OKCon 2011, WoLE 2012, NLP & DBpedia 2013), one Springer book (Linked Data in Linguistics, LDL 2012) and one journal special issue (Multilingual Linked Open Data, MLOD to appear) – have been (co-)edited to create incentives for scientists to convert and publish Linked Data and thus to contribute open and/or linguistic data to the LOD cloud. Based on the disseminated call for papers, 152 authors contributed one or more accepted submissions to our venues and 120 reviewers were involved in peer-reviewing. “DBpedia as a Multilingual Language Resource” and “Leveraging the Crowdsourcing of Lexical Resources for Bootstrapping a Linguistic Linked Data Cloud” contain this thesis’ contribution to the DBpedia Project in order to further increase the size and inter-linkage of the LOD Cloud with lexical-semantic resources. Our contribution comprises extracted data from Wiktionary (an online, collaborative dictionary similar to Wikipedia) in more than four languages (now six) as well as language-specific versions of DBpedia, including a quality assessment of inter-language links between Wikipedia editions and internationalized content negotiation rules for Linked Data. In particular the work described in created the foundation for a DBpedia Internationalisation Committee with members from over 15 different languages with the common goal to push DBpedia as a free and open multilingual language resource. Part III - The NLP Interchange Format (NIF). “NIF 2.0 Core Specification”, “NIF 2.0 Resources and Architecture” and “Evaluation and Related Work” constitute one of the main contribution of this thesis. The NLP Interchange Format (NIF) is an RDF/OWL-based format that aims to achieve interoperability between Natural Language Processing (NLP) tools, language resources and annotations. The core specification is included in and describes which URI schemes and RDF vocabularies must be used for (parts of) natural language texts and annotations in order to create an RDF/OWL-based interoperability layer with NIF built upon Unicode Code Points in Normal Form C. In , classes and properties of the NIF Core Ontology are described to formally define the relations between text, substrings and their URI schemes. contains the evaluation of NIF. In a questionnaire, we asked questions to 13 developers using NIF. UIMA, GATE and Stanbol are extensible NLP frameworks and NIF was not yet able to provide off-the-shelf NLP domain ontologies for all possible domains, but only for the plugins used in this study. After inspecting the software, the developers agreed however that NIF is adequate enough to provide a generic RDF output based on NIF using literal objects for annotations. All developers were able to map the internal data structure to NIF URIs to serialize RDF output (Adequacy). The development effort in hours (ranging between 3 and 40 hours) as well as the number of code lines (ranging between 110 and 445) suggest, that the implementation of NIF wrappers is easy and fast for an average developer. Furthermore the evaluation contains a comparison to other formats and an evaluation of the available URI schemes for web annotation. In order to collect input from the wide group of stakeholders, a total of 16 presentations were given with extensive discussions and feedback, which has lead to a constant improvement of NIF from 2010 until 2013. After the release of NIF (Version 1.0) in November 2011, a total of 32 vocabulary employments and implementations for different NLP tools and converters were reported (8 by the (co-)authors, including Wiki-link corpus, 13 by people participating in our survey and 11 more, of which we have heard). Several roll-out meetings and tutorials were held (e.g. in Leipzig and Prague in 2013) and are planned (e.g. at LREC 2014). Part IV - The NLP Interchange Format in Use. “Use Cases and Applications for NIF” and “Publication of Corpora using NIF” describe 8 concrete instances where NIF has been successfully used. One major contribution in is the usage of NIF as the recommended RDF mapping in the Internationalization Tag Set (ITS) 2.0 W3C standard and the conversion algorithms from ITS to NIF and back. One outcome of the discussions in the standardization meetings and telephone conferences for ITS 2.0 resulted in the conclusion there was no alternative RDF format or vocabulary other than NIF with the required features to fulfill the working group charter. Five further uses of NIF are described for the Ontology of Linguistic Annotations (OLiA), the RDFaCE tool, the Tiger Corpus Navigator, the OntosFeeder and visualisations of NIF using the RelFinder tool. These 8 instances provide an implemented proof-of-concept of the features of NIF. starts with describing the conversion and hosting of the huge Google Wikilinks corpus with 40 million annotations for 3 million web sites. The resulting RDF dump contains 477 million triples in a 5.6 GB compressed dump file in turtle syntax. describes how NIF can be used to publish extracted facts from news feeds in the RDFLiveNews tool as Linked Data. Part V - Conclusions. provides lessons learned for NIF, conclusions and an outlook on future work. Most of the contributions are already summarized above. One particular aspect worth mentioning is the increasing number of NIF-formated corpora for Named Entity Recognition (NER) that have come into existence after the publication of the main NIF paper Integrating NLP using Linked Data at ISWC 2013. These include the corpora converted by Steinmetz, Knuth and Sack for the NLP & DBpedia workshop and an OpenNLP-based CoNLL converter by Brümmer. Furthermore, we are aware of three LREC 2014 submissions that leverage NIF: NIF4OGGD - NLP Interchange Format for Open German Governmental Data, N^3 – A Collection of Datasets for Named Entity Recognition and Disambiguation in the NLP Interchange Format and Global Intelligent Content: Active Curation of Language Resources using Linked Data as well as an early implementation of a GATE-based NER/NEL evaluation framework by Dojchinovski and Kliegr. Further funding for the maintenance, interlinking and publication of Linguistic Linked Data as well as support and improvements of NIF is available via the expiring LOD2 EU project, as well as the CSA EU project called LIDER, which started in November 2013. Based on the evidence of successful adoption presented in this thesis, we can expect a decent to high chance of reaching critical mass of Linked Data technology as well as the NIF standard in the field of Natural Language Processing and Language Resources.:CONTENTS i introduction and background 1 1 introduction 3 1.1 Natural Language Processing . . . . . . . . . . . . . . . 3 1.2 Open licenses, open access and collaboration . . . . . . 5 1.3 Linked Data in Linguistics . . . . . . . . . . . . . . . . . 6 1.4 NLP for and by the Semantic Web – the NLP Inter- change Format (NIF) . . . . . . . . . . . . . . . . . . . . 8 1.5 Requirements for NLP Integration . . . . . . . . . . . . 10 1.6 Overview and Contributions . . . . . . . . . . . . . . . 11 2 background 15 2.1 The Working Group on Open Data in Linguistics (OWLG) 15 2.1.1 The Open Knowledge Foundation . . . . . . . . 15 2.1.2 Goals of the Open Linguistics Working Group . 16 2.1.3 Open linguistics resources, problems and chal- lenges . . . . . . . . . . . . . . . . . . . . . . . . 17 2.1.4 Recent activities and on-going developments . . 18 2.2 Technological Background . . . . . . . . . . . . . . . . . 18 2.3 RDF as a data model . . . . . . . . . . . . . . . . . . . . 21 2.4 Performance and scalability . . . . . . . . . . . . . . . . 22 2.5 Conceptual interoperability . . . . . . . . . . . . . . . . 22 ii language resources as linked data 25 3 linked data in linguistics 27 3.1 Lexical Resources . . . . . . . . . . . . . . . . . . . . . . 29 3.2 Linguistic Corpora . . . . . . . . . . . . . . . . . . . . . 30 3.3 Linguistic Knowledgebases . . . . . . . . . . . . . . . . 31 3.4 Towards a Linguistic Linked Open Data Cloud . . . . . 32 3.5 State of the Linguistic Linked Open Data Cloud in 2012 33 3.6 Querying linked resources in the LLOD . . . . . . . . . 36 3.6.1 Enriching metadata repositories with linguistic features (Glottolog → OLiA) . . . . . . . . . . . 36 3.6.2 Enriching lexical-semantic resources with lin- guistic information (DBpedia (→ POWLA) → OLiA) . . . . . . . . . . . . . . . . . . . . . . . . 38 4 DBpedia as a multilingual language resource: the case of the greek dbpedia edition. 39 4.1 Current state of the internationalization effort . . . . . 40 4.2 Language-specific design of DBpedia resource identifiers 41 4.3 Inter-DBpedia linking . . . . . . . . . . . . . . . . . . . 42 4.4 Outlook on DBpedia Internationalization . . . . . . . . 44 5 leveraging the crowdsourcing of lexical resources for bootstrapping a linguistic linked data cloud 47 5.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . 48 5.2 Problem Description . . . . . . . . . . . . . . . . . . . . 50 5.2.1 Processing Wiki Syntax . . . . . . . . . . . . . . 50 5.2.2 Wiktionary . . . . . . . . . . . . . . . . . . . . . . 52 5.2.3 Wiki-scale Data Extraction . . . . . . . . . . . . . 53 5.3 Design and Implementation . . . . . . . . . . . . . . . . 54 5.3.1 Extraction Templates . . . . . . . . . . . . . . . . 56 5.3.2 Algorithm . . . . . . . . . . . . . . . . . . . . . . 56 5.3.3 Language Mapping . . . . . . . . . . . . . . . . . 58 5.3.4 Schema Mediation by Annotation with lemon . 58 5.4 Resulting Data . . . . . . . . . . . . . . . . . . . . . . . . 58 5.5 Lessons Learned . . . . . . . . . . . . . . . . . . . . . . . 60 5.6 Discussion and Future Work . . . . . . . . . . . . . . . 60 5.6.1 Next Steps . . . . . . . . . . . . . . . . . . . . . . 61 5.6.2 Open Research Questions . . . . . . . . . . . . . 61 6 nlp & dbpedia, an upward knowledge acquisition spiral 63 6.1 Knowledge acquisition and structuring . . . . . . . . . 64 6.2 Representation of knowledge . . . . . . . . . . . . . . . 65 6.3 NLP tasks and applications . . . . . . . . . . . . . . . . 65 6.3.1 Named Entity Recognition . . . . . . . . . . . . 66 6.3.2 Relation extraction . . . . . . . . . . . . . . . . . 67 6.3.3 Question Answering over Linked Data . . . . . 67 6.4 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 6.4.1 Gold and silver standards . . . . . . . . . . . . . 69 6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 iii the nlp interchange format (nif) 73 7 nif 2.0 core specification 75 7.1 Conformance checklist . . . . . . . . . . . . . . . . . . . 75 7.2 Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 7.2.1 Definition of Strings . . . . . . . . . . . . . . . . 78 7.2.2 Representation of Document Content with the nif:Context Class . . . . . . . . . . . . . . . . . . 80 7.3 Extension of NIF . . . . . . . . . . . . . . . . . . . . . . 82 7.3.1 Part of Speech Tagging with OLiA . . . . . . . . 83 7.3.2 Named Entity Recognition with ITS 2.0, DBpe- dia and NERD . . . . . . . . . . . . . . . . . . . 84 7.3.3 lemon and Wiktionary2RDF . . . . . . . . . . . 86 8 nif 2.0 resources and architecture 89 8.1 NIF Core Ontology . . . . . . . . . . . . . . . . . . . . . 89 8.1.1 Logical Modules . . . . . . . . . . . . . . . . . . 90 8.2 Workflows . . . . . . . . . . . . . . . . . . . . . . . . . . 91 8.2.1 Access via REST Services . . . . . . . . . . . . . 92 8.2.2 NIF Combinator Demo . . . . . . . . . . . . . . 92 8.3 Granularity Profiles . . . . . . . . . . . . . . . . . . . . . 93 8.4 Further URI Schemes for NIF . . . . . . . . . . . . . . . 95 8.4.1 Context-Hash-based URIs . . . . . . . . . . . . . 99 9 evaluation and related work 101 9.1 Questionnaire and Developers Study for NIF 1.0 . . . . 101 9.2 Qualitative Comparison with other Frameworks and Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 9.3 URI Stability Evaluation . . . . . . . . . . . . . . . . . . 103 9.4 Related URI Schemes . . . . . . . . . . . . . . . . . . . . 104 iv the nlp interchange format in use 109 10 use cases and applications for nif 111 10.1 Internationalization Tag Set 2.0 . . . . . . . . . . . . . . 111 10.1.1 ITS2NIF and NIF2ITS conversion . . . . . . . . . 112 10.2 OLiA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 10.3 RDFaCE . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 10.4 Tiger Corpus Navigator . . . . . . . . . . . . . . . . . . 121 10.4.1 Tools and Resources . . . . . . . . . . . . . . . . 122 10.4.2 NLP2RDF in 2010 . . . . . . . . . . . . . . . . . . 123 10.4.3 Linguistic Ontologies . . . . . . . . . . . . . . . . 124 10.4.4 Implementation . . . . . . . . . . . . . . . . . . . 125 10.4.5 Evaluation . . . . . . . . . . . . . . . . . . . . . . 126 10.4.6 Related Work and Outlook . . . . . . . . . . . . 129 10.5 OntosFeeder – a Versatile Semantic Context Provider for Web Content Authoring . . . . . . . . . . . . . . . . 131 10.5.1 Feature Description and User Interface Walk- through . . . . . . . . . . . . . . . . . . . . . . . 132 10.5.2 Architecture . . . . . . . . . . . . . . . . . . . . . 134 10.5.3 Embedding Metadata . . . . . . . . . . . . . . . 135 10.5.4 Related Work and Summary . . . . . . . . . . . 135 10.6 RelFinder: Revealing Relationships in RDF Knowledge Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 10.6.1 Implementation . . . . . . . . . . . . . . . . . . . 137 10.6.2 Disambiguation . . . . . . . . . . . . . . . . . . . 138 10.6.3 Searching for Relationships . . . . . . . . . . . . 139 10.6.4 Graph Visualization . . . . . . . . . . . . . . . . 140 10.6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . 141 11 publication of corpora using nif 143 11.1 Wikilinks Corpus . . . . . . . . . . . . . . . . . . . . . . 143 11.1.1 Description of the corpus . . . . . . . . . . . . . 143 11.1.2 Quantitative Analysis with Google Wikilinks Cor- pus . . . . . . . . . . . . . . . . . . . . . . . . . . 144 11.2 RDFLiveNews . . . . . . . . . . . . . . . . . . . . . . . . 144 11.2.1 Overview . . . . . . . . . . . . . . . . . . . . . . 145 11.2.2 Mapping to RDF and Publication on the Web of Data . . . . . . . . . . . . . . . . . . . . . . . . . 146 v conclusions 149 12 lessons learned, conclusions and future work 151 12.1 Lessons Learned for NIF . . . . . . . . . . . . . . . . . . 151 12.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 151 12.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 153

Diagnostic Captioning is described as “the automatic generation of a diagnostic text from a set of medical images of a patient collected during an examination” [59] and it can assist inexperienced doctors and radiologists to reduce clinical errors or help experienced professionals increase their productivity. In this context, tools that would help medical doctors produce higher quality reports in less time could be of high interest for medical imaging departments, as well as significantly impact deep learning research within the biomedical domain, which makes it particularly interesting for people involved in industry and researchers all along. In this work, we attempted to develop Diagnostic Captioning systems, based on novel Deep Learning approaches, to investigate to what extent Neural Networks are capable of performing medical image tagging, as well as automatically generating a diagnostic text from a set of medical images. Towards this objective, the first step is concept detection, which boils down to predicting the relevant tags for X-RAY images, whereas the ultimate goal is caption generation. To this end, we further participated in ImageCLEFmedical 2022 evaluation campaign, addressing both the concept detection and the caption prediction tasks by developing baselines based on Deep Neural Networks; including image encoders, classifiers and text generators; in order to get a quantitative measure of my proposed architectures’ performance [28]. My contribution to the evaluation campaign, as part of this work and on behalf of NeuralDynamicsLab¹ group at KTH Royal Institute of Technology, within the school of Electrical Engineering and Computer Science, ranked 4th in the former and 5th in the latter task [55, 68] among 12 groups included within the top-10 best performing submissions in both tasks. / Diagnostisk textning avser automatisk generering från en diagnostisk text från en uppsättning medicinska bilder av en patient som samlats in under en undersökning och den kan hjälpa oerfarna läkare och radiologer, minska kliniska fel eller hjälpa erfarna yrkesmän att producera diagnostiska rapporter snabbare [59]. Därför kan verktyg som skulle hjälpa läkare och radiologer att producera rapporter av högre kvalitet på kortare tid vara av stort intresse för medicinska bildbehandlingsavdelningar, såväl som leda till inverkan på forskning om djupinlärning, vilket gör den domänen särskilt intressant för personer som är involverade i den biomedicinska industrin och djupinlärningsforskare. I detta arbete var mitt huvudmål att utveckla system för diagnostisk textning, med hjälp av nya tillvägagångssätt som används inom djupinlärning, för att undersöka i vilken utsträckning automatisk generering av en diagnostisk text från en uppsättning medi-cinska bilder är möjlig. Mot detta mål är det första steget konceptdetektering som går ut på att förutsäga relevanta taggar för röntgenbilder, medan slutmålet är bildtextgenerering. Jag deltog i ImageCLEF Medical 2022-utvärderingskampanjen, där jag deltog med att ta itu med både konceptdetektering och bildtextförutsägelse för att få ett kvantitativt mått på prestandan för mina föreslagna arkitekturer [28]. Mitt bidrag, där jag representerade forskargruppen NeuralDynamicsLab² , där jag arbetade som ledande forskningsingenjör, placerade sig på 4:e plats i den förra och 5:e i den senare uppgiften [55, 68] bland 12 grupper som ingår bland de 10 bästa bidragen i båda uppgifterna.

Models based around the transformer architecture have become one of the most prominent for solving a multitude of natural language processing (NLP)tasks since its introduction in 2017. However, much research related to the transformer model has focused primarily on achieving high performance and many problems remain unsolved. Two of the most prominent currently are the lack of high performing non-English pre-trained models, and the limited number of words most trained models can incorporate for their context. Solving these problems would make NLP models more suitable for real-world applications, improving information retrieval, reading comprehension, and more. All previous research has focused on incorporating long-context for English language models. This thesis investigates the cross-lingual transferability between languages when only training for long-context in English. Training long-context models in English only could make long-context in low-resource languages, such as Swedish, more accessible since it is hard to find such data in most languages and costly to train for each language. This could become an efficient method for creating long-context models in other languages without the need for such data in all languages or pre-training from scratch. We extend the models’ context using the training scheme of the Longformer architecture and fine-tune on a question-answering task in several languages. Our evaluation could not satisfactorily confirm nor deny if transferring long-term context is possible for low-resource languages. We believe that using datasets that require long-context reasoning, such as a multilingual TriviaQAdataset, could demonstrate our hypothesis’s validity.

This essay investigates what may constitute a hypnotic digital artefact from a design standpoint. This essay is meant to help designers who want to create hypnotic digital artefacts in the shape of a game, or researchers who wants to further this field. With a case study analysing the game Cultist Simulator, this essay observes applications from this essay’s frameworks: NLP, Procedural Rhetorics, Flow, Trance, and Ericksonian Hypnosis. The case study serves to demonstrate how a larger scale reflection of intrinsic cross over points between hypnosis and the video game medium may take place within state-of-the-art discourse. This essay fulfils its design-aid purpose by charting factors that can be put in place to facilitate a trance and a hypnosis in a game, in a design table summarising design methods discussed. The means that may put a player’s mind in abeyance are posited here regarding how this may influence the game experience, including induction techniques, where suggestions are provided in how these might translate to a game format. Through its frameworks and case study, hypnotic content generation is put in focus, where this essay finds that games utilising metaphors and depicting inner spaces carry significance in this pursuit. It also finds that mirroring communication of the unconscious, such as adhering to rules of a dream state, and acknowledging the unconscious’ uses and capacities, has potential in this pursuit. Importantly, the essay includes a discussion on Cultist Simulator’s decadent aesthetics and its role in leading a player towards an alternate state of consciousness.