Neural network models such as Transformer-based BERT, mBERT and RoBERTa are achieving impressive performance (Devlin et al., 2019; Lewis et al., 2020; Liu et al., 2019; Raffel et al., 2020; Y. Sun et al., 2019), but we still know little about their inner working due to the complex technique like multi-head self-attention they implement. Attention is commonly taken as a crucial way to explain the model outputs, but there are studies argue that attention may not provide faithful and reliable explanations in recent years (Jain and Wallace, 2019; Pruthi et al., 2020; Serrano and Smith, 2019; Wiegreffe and Pinter, 2019). Bastings and Filippova (2020) then propose that saliency may give better model interpretations since it is designed to find which token contributes to the prediction, i.e. the exact goal of explanation.  In this thesis, we investigate the extent to which syntactic structure is reflected in BERT, mBERT and RoBERTa trained on English and Chinese by using a gradient-based saliency method introduced by Simonyan et al. (2014). We examine the dependencies that our models and baselines predict.  We find that our models can predict some dependencies, especially those that have shorter mean distance and more fixed position of heads and dependents, even though all our models can handle global dependencies in theory. Besides, BERT usually has higher overall accuracy on connecting dependents to their corresponding heads, followed by mBERT and RoBERTa. Yet all the three model in fact have similar results on individual relations. Moreover, models trained on English have better performances than models trained on Chinese, possibly because of the flexibility of Chinese language.

There is an increasing interest in using deep neural networks in various downstream natural language processing tasks. Such models are commonly used as black boxes, meaning that their decision-making is difficult to interpret. In order to build trust in models, it is crucial to analyse their inner workings which lead to predictions. The need to interpret natural language processing models has induced research on linguistically-informed interpretability. This field revolves around choosing specific linguistic phenomena and inspecting models' capability to capture them without being explicitly trained for it.  This thesis project contributes to the field by assessing the ability of BERT-style models to learn subject number in Lithuanian and English. The experiments revolve around designing diagnostic classifiers which are used to determine if the models are capable of learning this particular linguistic phenomenon. The results show that BERT-style models are capable of implicitly learning the number of a subject both in Lithuanian and English. However, this seems to be harder in Lithuanian, as diagnostic classifiers show a lower accuracy. The study observes that the accuracy of logistic regression diagnostic classifiers fluctuates to a large extent. Fully connected neural network classifiers outperform logistic regression classifiers.

For the vast majority of languages, Natural Language Processing (NLP) tools are either absent entirely, or leave much to be desired in their final performance. Despite having nearly 4 million speakers, one such low-resource language is Galician. In an effort to expand available NLP resources, this project sought to construct a dataset for Semantic Role Labeling (SRL) and produce a baseline for future research to use in comparisons. SRL is a task which has shown success in amplifying the final output for various NLP systems, including Machine Translation and other interactive language models. This project was successful in that fact and produced 24 SRL models and two SRL datasets; one Galician and one Spanish. mBERT and XLM-R were chosen as the baseline architectures; additional models were first pre-trained on the SRL task in a language other than the target to measure the effects of transfer-learning. Scores are reported on a scale of 0.0-1.0. The best performing Galician SRL model achieved an f1 score of 0.74, introducing a baseline for future Galician SRL systems. The best performing Spanish SRL model achieved an f1 score of 0.83, outperforming the baseline set by the 2009 CoNLL Shared Task by 0.025. A pre-processing method, verbal indexing, was also introduced which allowed for increased performance in the SRL parsing of highly complex sentences; effects were amplified in scenarios where the model was both pre-trained and fine-tuned on datasets utilizing the method, but still visible even when only used during fine-tuning. / För de allra flesta språken saknas språkteknologiska verktyg (NLP) helt, eller för dem de var i finns tillgängliga är dessa verktygs prestanda minst sagt, sämre än medelmåttig. Trots sina nästan 4 miljoner talare, är galiciska ett språk med brist på tillräckliga resurser. I ett försök att utöka tillgängliga NLP-resurser för språket, konstruerades i detta projekt en uppsättning data för så kallat Semantic Role Labeling (SRL) som sedan användes för att utveckla grundläggande SRL-modeller att falla tillbaka på och jämföra  med i framtida forskning. SRL är en uppgift som har visat framgång när det gäller att förstärka slutresultatet för olika NLP-system, inklusive maskinöversättning och andra interaktiva språkmodeller. I detta avseende visade detta projekt på framgång och som del av det utvecklades 24 SRL-modeller och två SRL-datauppsåttningar; en galicisk och en spansk. mBERT och XLM-R valdes som baslinjearkitekturer; ytterligare modeller tränades först på en SRL-uppgift på ett språk annat än målspråket för att mäta effekterna av överföringsinlärning (Transfer Learning) Poäng redovisas på en skala från 0.0-1.0. Den galiciska SRL-modellen med bäst prestanda uppnådde ett f1-poäng på 0.74, vilket introducerar en baslinje för framtida galiciska SRL-system. Den bästa spanska SRL-modellen uppnådde ett f1-poäng på 0.83, vilket överträffade baslinjen +0.025 som sattes under CoNLL Shared Task 2009. I detta projekt introduceras även en ny metod för behandling av lingvistisk data, så kallad verbalindexering, som ökade prestandan av mycket komplexa meningar. Denna prestandaökning först märktes ytterligare i de scenarier och är en modell både förtränats och finjusterats på uppsättningar data som behandlats med metoden, men visade även på märkbara förbättringar då en modell endast genomgått finjustering. / Para la gran mayoría de los idiomas, las herramientas de procesamiento del lenguaje natural (NLP) están completamente ausentes o dejan mucho que desear en su desempeño final. A pesar de tener casi 4 millones de hablantes, el gallego continúa siendo un idioma de bajos recursos. En un esfuerzo por expandir los recursos de NLP disponibles, el objetivo de este proyecto fue construir un conjunto de datos para el Etiquetado de Roles Semánticos (SRL) y producir una referencia para que futuras investigaciones puedan utilizar en sus comparaciones. SRL es una tarea que ha tenido éxito en la amplificación del resultado final de varios sistemas NLP, incluida la traducción automática, y otros modelos de lenguaje interactivo. Este proyecto fue exitoso en ese hecho y produjo 24 modelos SRL y dos conjuntos de datos SRL; uno en gallego y otro en español. Se eligieron mBERT y XLM-R como las arquitecturas de referencia; previamente se entrenaron modelos adicionales en la tarea SRL en un idioma distinto al idioma de destino para medir los efectos del aprendizaje por transferencia. Las puntuaciones se informan en una escala de 0.0 a 1.0. El modelo SRL gallego con mejor rendimiento logró una puntuación de f1 de 0.74, introduciendo un objetivo de referencia para los futuros sistemas SRL gallegos. El modelo español de SRL con mejor rendimiento logró una puntuación de f1 de 0.83, superando la línea base establecida por la Tarea Compartida CoNLL de 2009 en 0.025. También se introdujo un método de preprocesamiento, indexación verbal, que permitió un mayor rendimiento en el análisis SRL de oraciones muy complejas; los efectos se amplificaron cuando el modelo primero se entrenó y luego se ajustó con los conjuntos de datos que utilizaban el método, pero los efectos aún fueron visibles incluso cuando se lo utilizó solo durante el ajuste.

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.