In recent years, deep learning based sequence modeling (neural sequence modeling) techniques have made substantial progress in many tasks, including information retrieval, question answering, information extraction, machine translation, etc. Benefiting from the highly scalable attention-based Transformer architecture and enormous open access online data, large-scale pre-trained language models have shown great modeling and generalization capacity for sequential data. However, not all domains benefit equally from the rapid development of neural sequence modeling. Domains like healthcare and software engineering have vast amounts of sequential data containing rich knowledge, yet remain under-explored due to a number of challenges: 1) the distribution of the sequences in specific domains is different from the general domain; 2) the effective comprehension of domain-specific data usually relies on domain knowledge; and 3) the labelled data is usually scarce and expensive to get in domain-specific settings. In this thesis, we focus on the research problem of applying neural sequence modeling methods to address both common and domain-specific challenges from the healthcare and software engineering domains. We systematically investigate neural-based machine learning approaches to address the above challenges in three research directions: 1) learning with long sequences, 2) learning from domain knowledge and 3) learning under limited supervision. Our work can also potentially benefit more domains with large amounts of sequential data. / Doctor of Philosophy / In the last few years, computer programs that learn and understand human languages (an area called machine learning for natural language processing) have significantly improved. These advances are visible in various areas such as retrieving information, answering questions, extracting key details from texts, and translating between languages. A key to these successes has been the use of a type of neural network structure known as a "Transformer", which can process and learn from lots of information found online. However, these successes are not uniform across all areas. Two fields, healthcare and software engineering, still present unique challenges despite having a wealth of information. Some of these challenges include the different types of information in these fields, the need for specific expertise to understand this information, and the shortage of labeled data, which is crucial for training machine learning models. In this thesis, we focus on the use of machine learning for natural language processing methods to solve these challenges in the healthcare and software engineering fields. Our research investigates learning with long documents, learning from domain-specific expertise, and learning when there's a shortage of labeled data. The insights and techniques from our work could potentially be applied to other fields that also have a lot of sequential data.

Le code source offre un domaine d'application passionnant des méthodes d'apprentissage en profondeur, englobant des tâches telles que la synthèse, la réparation et l'analyse de programmes, ainsi que des tâches à l'intersection du code et du langage naturel. Bien que les modèles d’apprentissage profond pour le code, en particulier les grands modèles de langage, aient récemment connu un succès significatif, ils peuvent avoir du mal à se généraliser à du code invisible. Cela peut conduire à des inexactitudes, en particulier lorsque vous travaillez avec des référentiels contenant des logiciels propriétaires ou du code en cours de travail. L'objectif principal de cette thèse est d'exploiter efficacement les signaux utiles du contexte disponible afin d'améliorer les performances des modèles de code d'apprentissage profond pour une tâche donnée. En incorporant ces indices contextuels, les capacités de généralisation du modèle sont amplifiées, fournissant des informations supplémentaires non évidentes à partir de l'entrée d'origine et orientant son attention vers des détails essentiels. De plus, l'utilisation d'indices contextuels facilite l'adaptation aux nouvelles tâches et améliore les performances des tâches existantes en effectuant des prédictions plus contextuelles. Pour y parvenir, nous présentons un cadre général comprenant deux étapes : (a) l'amélioration du contexte, qui implique l'enrichissement de l'entrée avec un contexte de support obtenu grâce à l'identification et à la sélection d'indices contextuels pertinents, et (b) la prédiction à l'aide du contexte amélioré, où nous exploitez le contexte de support combiné aux entrées pour faire des prédictions précises. La thèse présente quatre articles qui proposent diverses approches pour ces étapes. Le premier article divise le problème standard de la programmation par exemples en deux étapes: (a) trouver des programmes qui satisfont des exemples individuels (solutions par exemple) et, (b) combiner ces solutions par exemple en tirant parti de leurs états d'exécution de programme pour trouver un programme qui satisfait tous les exemples donnés. Le deuxième article propose une approche pour sélectionner des informations ciblées à partir du fichier actuel et les utiliser pour adapter le modèle de complétion de code à un contexte local jamais vu précédemment. Le troisième article s'appuie sur le deuxième article en tirant parti des indices contextuels de l'ensemble du répertoire de code à l'aide d'un ensemble de requêtes ({\it prompts}) proposées suggérant l'emplacement et le contenu du contexte particulièrement utile à extraire du répertoire. Nous proposons un cadre pour sélectionner la requête la plus pertinente, qui est ensuite utilisée pour demander à un modèle de langage de code de générer des prédictions pour le reste de la ligne de code suivant un curseur positionné dans un fichier. Le quatrième article prolonge le troisième article en proposant un cadre qui apprend à combiner plusieurs contextes divers à partir du répertoire. Nous montrons que la formation de modèles de language de code plus petits de cette manière fonctionne mieux ou à égalité avec des modèles beaucoup plus grands qui n'utilisent pas le contexte du répertoire de code. / Source code provides an exciting application area of deep learning methods, encompassing tasks like program synthesis, repair, and analysis, as well as tasks at the intersection of code and natural language. Although deep learning models for code, particularly large language models, have recently seen significant success, they can face challenges in generalizing to unseen code. This can lead to inaccuracies especially when working with repositories that contain proprietary software or work-in-progress code. The main focus of this thesis is to effectively harness useful signals from the available context such that it can improve the performance of the deep learning models of code at the given task. By incorporating these contextual cues, the model's generalization capabilities are amplified, providing additional insights not evident from the original input and directing its focus toward essential details. Furthermore, the use of contextual cues aids in adapting to new tasks and boosts performance on existing ones by making more context-aware predictions. To achieve this, we present a general framework comprising two stages: (a) Context Enhancement, which involves enriching the input with support context obtained through the identification and selection of relevant contextual cues, and (b) Prediction using the Enhanced Context, where we leverage the support context combined with the input to make accurate predictions. The thesis presents four articles that propose diverse approaches for these stages. The first article breaks the standard problem of programming by examples into two stages: (a) finding programs that satisfy individual examples (per-example solutions) and, (b) combining these per-example solutions by leveraging their program execution states to find a program that satisfies all given examples. The second article proposes an approach for selecting targeted information from the current file and using it to adapt the code completion model to an unseen, local context. The third article builds upon the second article by leveraging contextual cues from the entire code repository using a set of prompt proposals that govern the location and content of the context that should be taken from the repository. We propose a framework to select the most relevant prompt proposal context which is then used to prompt a large language model of code to generate predictions for the tokens in the rest of the line following the cursor in a file. The fourth article extends the third article by proposing a framework that learns to combine multiple diverse contexts from the repository. We show that training smaller models of code this way performs better or at par with significantly larger models that are not trained with repository context.