Usage of language models in an in-context learning environment has been adapted for a wide range of tasks. Recent works have showcased the impact of pretraining data on the in-context performance of language models. In this work, we experiment with numbers having high and low frequencies in the pretraining data to understand the impact of term frequencies on the model's performance. We also experiment with random and adversarial demonstrations to understand the pretraining bias present in the model. Through these experiments, we showcase the importance of pretraining frequencies of the numbers present in the demonstrations and explain how highly frequent terms can be used in the demonstrations to achieve better task performance. Moreover, we also show the impact of pretraining bias on the model's performance and explain how the model overcomes this bias with more demonstrations. / Master of Science / Recent works focus on understanding and improving the arithmetic capabilities of the state-of-the-art (SOTA) systems in the domain of Natural Language Processing (NLP). This work focuses on designing and performing novel experiments to analyze the impact of training data on the performance of such systems. Through these experiments, this work showcases interesting properties of the SOTA systems which will promote future research to understand them better as well as help in creating better downstream applications.

This research was concerned with furthering theoretical and practical understanding of student learning at university through a longitudinal, cross-sectional, in-depth study of first year students in a specific learning context, namely Educational Psychology. The main aim of the study was to investigate ways of assisting students to be effective learners. The particular role that affect played in learning and the relationship between learning behaviour and learning outcomes, was explored. A Conceptual Model of student learning incorporating student cognition, metacognition, motivation, affect and academic performance in a specific social and cultural context, underpinned the study. The study documented the design, implementation and evaluation - from both the students' and teacher's perspectives - of an in-context learning support program for first year students, using both quantitative and qualitative methodologies.The program was based on a theoretical framework which integrated cognitive, behavioural and social learning perspectives and focussed on increasing students' repertoire of learning strategies, promoting their higher level thinking and understanding, developing their metacognitive skills and managing their affect. It included an emphasis on student goal setting and time management, reading and writing strategies, learning for tests and exams, self-management, reflecting on and evaluating learning, and dealing with test anxiety.The main findings of the study were that providing in-context learning support was associated with positive changes in students' learning strategy use, motivational orientations, and affective reactions. Students valued teacher support and instructional strategies that promoted active learning. The instructor found that providing learning support was more challenging and rewarding than teaching content alone. The role of context - ++ / in particular, assessment tasks - in learning, was highlighted. The implications for teaching and learning were examined and the Conceptual Model was further refined. The research resulted in a more holistic and integrated perspective on learning support provision and on the role of cognitive, metacognitive, motivational and affective factors, and academic performance, in student learning.

This thesis investigates the vocabulary learning behavior of Romanian high school students in a digital context. The research identifies the vocabulary learning strategies used by EFL high school students and focuses on how the choice of vocabulary learning strategies varies across four independent variables: students' age, gender, academic profile (math-ICT, humanities, science and economic-technical) and language program (intensive English, bilingual, normal). These variables are hypothesized to influence learners' vocabulary behavior. Furthermore, the study examines the technology enhanced tools (computer and mobile assisted language learning tools) used by these students in their vocabulary learning as well as their attitudes towards using technology in vocabulary learning. Likewise, the study analyzes how students' choice of technology enhanced tools and their attitudes towards them vary across the four independent variables. The study is a mixed methods investigation with 1,239 participants (60% female, 40% male, aged 14-19 years old) learning English as a foreign language in nine Romanian secondary schools. Of the 1,239 participants who filled in the self-reported questionnaire, 43 also participated in focus group discussions prior to the administration of the questionnaire. The quantitative data were analyzed using descriptive and inferential statistics procedures whereas the qualitative data were analyzed thematically. The results from both phases were integrated in the results chapter. The main findings indicated that Romanian high school students prefer social strategies, followed by determination, metacognitive, cognitive and memory strategies. However, the usage of the strategies in these categories is medium towards low. As for individual vocabulary learning strategies, the participants reported that the impact of a new word, English media, guessing from context, associating the word with a picture and using cognates are frequently used strategies. The results also indicated that students' use of vocabulary learning strategies varies across the four independent variables. As far as the use of digital tools for vocabulary learning, the findings indicated that the students in this particular cultural context use few available digital tools with a preference for online dictionaries, games and social networking web sites. The results showed that overall Romanian students are not very familiar with computer and mobile assisted language learning tools, their attitudes towards the use of digital tools for vocabulary learning are neutral and they mostly associate the use of personal devices with their personal space, suggesting that they may not want to embed learning in their everyday activities. The results enrich existing knowledge of vocabulary learning strategies in a Romanian cultural context and they also give us an insight into how high school students use computer and mobile assisted language tools in their vocabulary learning. Implications for theory and practice are also discussed.

This thesis examines the potential of a learning outcome-led model of curriculum design to influence how students perceive learning in education studies within a modular context of a new university. It identifies and compares the conceptions of learning held by students and lecturers on traditional and outcome-led modules, and it explores and specifies the design factors which shape these conceptions. The issue is located within the interpretivist paradigm for the research seeks understanding which derives from the perceptions, attitudes and beliefs that students and their lecturers hold about learning in a given context. But the methodology employed is not wholly consistent with this paradigm, for a qualitative approach is complemented by the use of factor analysis techniques to facilitate the identification of the design features which influence how students perceive learning. The approach is thus eclectic drawing on quantitative methods to examine what is essentially qualitative data. An innovative model of learning outcome-led design is proposed, implemented and modified as a result of the research. The learner is placed at the centre of the learning experience which is defined as incorporating three domains: the teaching context; the assessment régime; and the directed learning undertaken by students outside of taught sessions. The model incorporates a trichotomy of outcomes which define the subject -specific, the transferable skills and the generic academic outcomes which influence directly both the content and process of learning, and which successful students are expected to achieve on completion of a module. The findings show that five design features influence how students perceive learning: the clarity of expectations; congruence between the content and process of each domain of the learning experience; direction in respect to the learning activities which should be undertaken in each domain to achieve the outcomes; and the content and process of the teaching context. The data suggest that a much higher profile should be given to metacognitive skills in curriculum development in HE because how students perceive both the process and the content of learning profoundly influences their conception of learning and, consistent with the underpinning theory, how they approach learning and therefore ultimately the kind of outcomes they achieve. The research leads to recommendations for the modification of the three models of learning in context; Ramsden (1988), Biggs (1990b) and Prosser (1995), which are presented and analysed in the thesis. The findings suggest that the learning experience should be redefined to specify the three domains - the teaching context, assessment régime and directed learning - and that clarity of expectations, metacognitive skills and congruence between the content and process of learning in each of the domains should be articulated as directly influencing students' conceptions of learning. The models should also seek to indicate that learning outcomes influence how students perceive learning, and that therefore they feature both at the starting point and as the end product of a contextualised learning process. The findings relating to students' conceptions of learning show that the study of outcome-led modules has resulted in a much greater degree of congruence between how lecturers and students perceive learning in a given module and that fewer students studying outcome-led modules hold a quantitative conception of learning. This suggests that the outcome-led model does have the potential to improve teaching and learning and consequently that there is an educational rationale for curriculum development premised on this model.

Making information stored in databases more accessible to users inexperienced in structured query language (SQL) by converting natural language to SQL queries has long been a prominent research area in both the database and natural language processing (NLP) communities. There have been numerous approaches proposed for this task, such as encoder-decoder frameworks, semantic grammars, and more recently with the use of large language models (LLMs). When training LLMs to successfully generate SQL queries from natural language questions there are three notable methods used, pretraining, transfer learning and in-context learning (ICL). ICL is particularly advantageous in scenarios where the hardware at hand is limited, time is of concern and large amounts of task specific labled data is nonexistent. This study seeks to evaluate two strategies in ICL, namely zero-shot and few-shot scenarios using the Mistral-7B-Instruct LLM. Evaluation of the few-shot scenarios was conducted using two techniques, random selection and Jaccard Similarity. The zero-shot scenarios served as a baseline for the few-shot scenarios to overcome, which ended as anticipated, with the few-shot scenarios using Jaccard similarity outperforming the other two methods, followed by few-shot scenarios using random selection coming in at second best, and the zero-shot scenarios performing the worst. Evaluation results acquired based on execution accuracy and exact matching accuracy confirm that leveraging similarity in demonstrating examples when prompting the LLM will enhance the models knowledge about the database schema and table names which is used during the inference phase leadning to more accurately generated SQL queries than leveraging diversity in demonstrating examples.

Cette thèse traite des problèmes de modélisation, reconnaissance, localisation et utilisation du contexte pour la manipulation d'objets par un robot. Le processus de modélisation se divise en quatre composantes : le système réel, les données capteurs, les propriétés à reproduire et le modèle. En spécifiant chacune des ces composantes, il est possible de définir un processus de modélisation adapté au problème présent, la manipulation d'objets par un robot. Cette analyse mène à l'adoption des descripteurs de texture locaux pour la modélisation. La modélisation basée sur des descripteurs de texture locaux a été abordé dans de nombreux travaux traitant de structure par le mouvement (SfM) ou de cartographie et localisation simultanée (SLAM). Les méthodes existantes incluent Bundler, Roboearth et 123DCatch. Pourtant, aucune de ces méthodes n'a recueilli le consensus. En effet, l'implémentation d'une approche similaire montre que ces outils sont difficiles d'utilisation même pour des utilisateurs experts et qu'ils produisent des modèles d'une haute complexité. Cette complexité est utile pour fournir un modèle robuste aux variations de point de vue. Il existe deux façons pour un modèle d'être robuste : avec le paradigme des vues multiple ou celui des descripteurs forts. Dans le paradigme des vues multiples, le modèle est construit à partir d'un grand nombre de points de vue de l'objet. Le paradigme des descripteurs forts compte sur des descripteurs résistants aux changements de points de vue. Les expériences réalisées montrent que des descripteurs forts permettent d'utiliser un faible nombre de vues, ce qui résulte en un modèle simple. Ces modèles simples n'incluent pas tout les point de vus existants mais les angles morts peuvent être compensés par le fait que le robot est mobile et peut adopter plusieurs points de vue. En se basant sur des modèles simples, il est possible de définir des méthodes de modélisation basées sur des images seules, qui peuvent être récupérées depuis Internet. A titre d'illustration, à partir d'un nom de produit, il est possible de récupérer des manières totalement automatiques des images depuis des magasins en ligne et de modéliser puis localiser les objets désirés. Même avec une modélisation plus simple, dans des cas réel ou de nombreux objets doivent être pris en compte, il se pose des problèmes de stockage et traitement d'une telle masse de données. Cela se décompose en un problème de complexité, il faut traiter de nombreux modèles rapidement, et un problème d'ambiguïté, des modèles peuvent se ressembler. L'impact de ces deux problèmes peut être réduit en utilisant l'information contextuelle. Le contexte est toute information non issue des l'objet lui même et qui aide a la reconnaissance. Ici deux types de contexte sont abordés : le lieu et les objets environnants. Certains objets se trouvent dans certains endroits particuliers. En connaissant ces liens lieu/objet, il est possible de réduire la liste des objets candidats pouvant apparaître dans un lieu donné. Par ailleurs l'apprentissage du lien lieu/objet peut être fait automatiquement par un robot en modélisant puis explorant un environnement. L'information appris peut alors être fusionnée avec l'information visuelle courante pour améliorer la reconnaissance. Dans les cas des objets environnants, un objet peut souvent apparaître au cotés d'autres objets, par exemple une souris et un clavier. En connaissant la fréquence d'apparition d'un objet avec d'autres objets, il est possible de réduire la liste des candidats lors de la reconnaissance. L'utilisation d'un Réseau de Markov Logique est particulièrement adaptée à la fusion de ce type de données. Cette thèse montre la synergie de la robotique et du contexte pour la modélisation, reconnaissance et localisation d'objets. / This Thesis addresses the modeling, recognition, localization and use of context for objects manipulation by a robot. We start by presenting the modeling process and its components: the real system, the sensors' data, the properties to reproduce and the model. We show how, by specifying each of them, one can define a modeling process adapted to the problem at hand, namely object manipulation by a robot. This analysis leads us to the adoption of local textured descriptors for object modeling. Modeling with local textured descriptors is not a new concept, it is the subject of many Structure from Motion (SfM) or Simultaneous Localization and Mapping (SLAM) works. Existing methods include bundler, roboearth modeler and 123DCatch. Still, no method has gained widespread adoption. By implementing a similar approach, we show that they are hard to use even for expert users and produce highly complex models. Such complex techniques are necessary to guaranty the robustness of the model to view point change. There are two ways to handle the problem: the multiple views paradigm and the robust features paradigm. The multiple views paradigm advocate in favor of using a large number of views of the object. The robust feature paradigm relies on robust features able to resist large view point changes. We present a set of experiments to provide an insight into the right balance between both. By varying the number of views and using different features we show that small and fast models can provide robustness to view point changes up to bounded blind spots which can be handled by robotic means. We propose four different methods to build simple models from images only, with as little a priori information as possible. The first one applies to planar or piecewise planar objects and relies on homographies for localization. The second approach is applicable to objects with simple geometry, such as cylinders or spheres, but requires many measures on the object. The third method requires the use of a calibrated 3D sensor but no additional information. The fourth technique doesn't need a priori information at all. We apply this last method to autonomous grocery objects modeling. From images automatically retrieved from a grocery store website, we build a model which allows recognition and localization for tracking. Even using light models, real situations ask for numerous object models to be stored and processed. This poses the problems of complexity, processing multiple models quickly, and ambiguity, distinguishing similar objects. We propose to solve both problems by using contextual information. Contextual information is any information helping the recognition which is not directly provided by sensors. We focus on two contextual cues: the place and the surrounding objects. Some objects are mainly found in some particular places. By knowing the current place, one can restrict the number of possible identities for a given object. We propose a method to autonomously explore a previously labeled environment and establish a correspondence between objects and places. Then this information can be used in a cascade combining simple visual descriptors and context. This experiment shows that, for some objects, recognition can be achieved with as few as two simple features and the location as context. The objects surrounding a given object can also be used as context. Objects like a keyboard, a mouse and a monitor are often close together. We use qualitative spatial descriptors to describe the position of objects with respect to their neighbors. Using a Markov Logic Network, we learn patterns in objects disposition. This information can then be used to recognize an object when surrounding objects are already identified. This Thesis stresses the good match between robotics, context and objects recognition.

Cette thèse s’intéresse à la problématique de l’adaptation automatique de services dans ledomaine de l’intelligence ambiante. L’étude de la littérature montre que la sensibilité aucontexte est devenue un élément central pour la conception et la mise en place de servicesadaptatifs. Cependant, sa prise en compte se limite généralement à des descriptionsélémentaires de situations ou à des modèles prédéfinis. Afin de permettre une adaptation auxchangements d’habitudes des utilisateurs, à la dynamique de l’environnement et àl’hétérogénéité des sources de perception, nous proposons des mécanismes de découverte decontexte et de situations déclencheurs d’adaptation. Ces mécanismes s’appuient sur destechniques de fouille de données et sont intégrés au sein d’une architecture d’adaptationautomatique de services. Ces travaux ont été réalisés et appliqués à des projets d’intelligenceambiante pour de l’assistance à des personnes et plus particulièrement dans le cadre du projetITEA- MIDAS. / This thesis addresses the problem of dynamic adaptation of services in the context of ambientintelligence applications. Literature study shows how context-awareness plays a central rolein the design and implementation of adaptive services. However, its use is still limited toelementary descriptions and predefined situational models. Dynamic adaptation should becapable of following user habits to yield dynamic answers to environmental change, and tosupport heterogeneous sources of context. To this end, we propose mechanisms to discovercontexts and situations that trigger adaptation. These mechanisms rely on data miningtechniques, and are integrated within an architecture for dynamic adaptation of services. Thiswork was carried out and applied to ambient intelligence projects for the elderly, providingsupport and assistance in their daily lives, particularly in the context of the ITEA-MIDASproject.

<p dir="ltr">Semantic pairs, which consist of related entities or concepts, serve as the foundation for comprehending the meaning of language in both written and spoken forms. These pairs enable to grasp the nuances of relationships between words, phrases, or ideas, forming the basis for more advanced language tasks like entity recognition, sentiment analysis, machine translation, and question answering. They allow to infer causality, identify hierarchies, and connect ideas within a text, ultimately enhancing the depth and accuracy of automated language processing.</p><p dir="ltr">Nevertheless, the task of extracting semantic pairs from sentences poses a significant challenge, necessitating the relevance of syntactic dependency patterns (SDPs). Thankfully, semantic relationships exhibit adherence to distinct SDPs when connecting pairs of entities. Recognizing this fact underscores the critical importance of extracting these SDPs, particularly for specific semantic relationships like hyponym-hypernym, meronym-holonym, and cause-effect associations. The automated extraction of such SDPs carries substantial advantages for various downstream applications, including entity extraction, ontology development, and question answering. Unfortunately, this pivotal facet of pattern extraction has remained relatively overlooked by researchers in the domains of natural language processing (NLP) and information retrieval.</p><p dir="ltr">To address this gap, I introduce an attention-based supervised deep learning model, ASPER. ASPER is designed to extract SDPs that denote semantic relationships between entities within a given sentential context. I rigorously evaluate the performance of ASPER across three distinct semantic relations: hyponym-hypernym, cause-effect, and meronym-holonym, utilizing six datasets. My experimental findings demonstrate ASPER's ability to automatically identify an array of SDPs that mirror the presence of these semantic relationships within sentences, outperforming existing pattern extraction methods by a substantial margin.</p><p dir="ltr">Second, I want to use the SDPs to extract semantic pairs from sentences. I choose to extract cause-effect entities from medical literature. This task is instrumental in compiling various causality relationships, such as those between diseases and symptoms, medications and side effects, and genes and diseases. Existing solutions excel in sentences where cause and effect phrases are straightforward, such as named entities, single-word nouns, or short noun phrases. However, in the complex landscape of medical literature, cause and effect expressions often extend over several words, stumping existing methods, resulting in incomplete extractions that provide low-quality, non-informative, and at times, conflicting information. To overcome this challenge, I introduce an innovative unsupervised method for extracting cause and effect phrases, PatternCausality tailored explicitly for medical literature. PatternCausality employs a set of cause-effect dependency patterns as templates to identify the key terms within cause and effect phrases. It then utilizes a novel phrase extraction technique to produce comprehensive and meaningful cause and effect expressions from sentences. Experiments conducted on a dataset constructed from PubMed articles reveal that PatternCausality significantly outperforms existing methods, achieving a remarkable order of magnitude improvement in the F-score metric over the best-performing alternatives. I also develop various PatternCausality variants that utilize diverse phrase extraction methods, all of which surpass existing approaches. PatternCausality and its variants exhibit notable performance improvements in extracting cause and effect entities in a domain-neutral benchmark dataset, wherein cause and effect entities are confined to single-word nouns or noun phrases of one to two words.</p><p dir="ltr">Nevertheless, PatternCausality operates within an unsupervised framework and relies heavily on SDPs, motivating me to explore the development of a supervised approach. Although SDPs play a pivotal role in semantic relation extraction, pattern-based methodologies remain unsupervised, and the multitude of potential patterns within a language can be overwhelming. Furthermore, patterns do not consistently capture the broader context of a sentence, leading to the extraction of false-positive semantic pairs. As an illustration, consider the hyponym-hypernym pattern <i>the w of u</i> which can correctly extract semantic pairs for a sentence like <i>the village of Aasu</i> but fails to do so for the phrase <i>the moment of impact</i>. The root cause of this limitation lies in the pattern's inability to capture the nuanced meaning of words and phrases in a sentence and their contextual significance. These observations have spurred my exploration of a third model, DepBERT which constitutes a dependency-aware supervised transformer model. DepBERT's primary contribution lies in introducing the underlying dependency structure of sentences to a language model with the aim of enhancing token classification performance. To achieve this, I must first reframe the task of semantic pair extraction as a token classification problem. The DepBERT model can harness both the tree-like structure of dependency patterns and the masked language architecture of transformers, marking a significant milestone, as most large language models (LLMs) predominantly focus on semantics and word co-occurrence while neglecting the crucial role of dependency architecture.</p><p dir="ltr">In summary, my overarching contributions in this thesis are threefold. First, I validate the significance of the dependency architecture within various components of sentences and publish SDPs that incorporate these dependency relationships. Subsequently, I employ these SDPs in a practical medical domain to extract vital cause-effect pairs from sentences. Finally, my third contribution distinguishes this thesis by integrating dependency relations into a deep learning model, enhancing the understanding of language and the extraction of valuable semantic associations.</p>

<p dir="ltr">We propose a fully online procedure to conduct statistical inference with adaptively collected data. The low-rank structure of the model parameter and the adaptivity nature of the data collection process make this task challenging: standard low-rank estimators are biased and cannot be obtained in a sequential manner while existing inference approaches in sequential decision-making algorithms fail to account for the low-rankness and are also biased. To tackle the challenges previously outlined, we first develop an online low-rank estimation process employing Stochastic Gradient Descent with noisy observations. Subsequently, to facilitate statistical inference using the online low-rank estimator, we introduced a novel online debiasing technique designed to address both sources of bias simultaneously. This method yields an unbiased estimator suitable for parameter inference. Finally, we developed an inferential framework capable of establishing an online estimator for performing inference on the optimal policy value. In theory, we establish the asymptotic normality of the proposed online debiased estimators and prove the validity of the constructed confidence intervals for both inference tasks. Our inference results are built upon a newly developed low-rank stochastic gradient descent estimator and its non-asymptotic convergence result, which is also of independent interest.</p>

<p dir="ltr">Current artificial intelligence (AI) systems are proficient at tasks in a closed-world setting where the rules are often rigid. However, in real-world applications, the environment is usually open and dynamic. In this work, we investigate the effects of such dynamic environments on AI systems and develop ways to mitigate those effects. Central to our exploration is the concept of \textit{novelties}. Novelties encompass structural changes, unanticipated events, and environmental shifts that can confound traditional AI systems. We categorize novelties based on their representation, anticipation, and impact on agents, laying the groundwork for systematic detection and adaptation strategies. We explore novelties in the context of stochastic games. Decision-making in stochastic games exercises many aspects of the same reasoning capabilities needed by AI agents acting in the real world. A multi-agent stochastic game allows for infinitely many ways to introduce novelty. We propose an extension of the deep reinforcement learning (DRL) paradigm to develop agents that can detect and adapt to novelties in these environments. To address the sample efficiency challenge in DRL, we introduce a hybrid approach that combines fixed-policy methods with traditional DRL techniques, offering enhanced performance in complex decision-making tasks. We present a novel method for detecting anticipated novelties in multi-agent games, leveraging information theory to discern patterns indicative of collusion among players. Finally, we introduce DABLER, a pioneering deep reinforcement learning architecture that dynamically adapts to changing environmental conditions through broad learning approaches and environment recognition. Our findings underscore the importance of developing AI systems equipped to navigate the uncertainties of the open world, offering promising pathways for advancing AI research and application in real-world settings.</p>