In this work, we explore the possibility of using word alignment in parallel corpus to project language annotations such as Part-of-Speech tags and dependency relation from high-resource languages to low-resource languages. We use a parallel corpus of Bible translations, including 1,444 translations in 986 languages, and a well-developed parser is used to annotate source languages (English, French, German, and Czech). The annotations are projected to low-resource languages based on the word alignment results. Then we design and refine the process of detecting verbs and the subjects/objects linked to this verb, find and count the word orders. We used data from The World Atlas of Language Structures (WALS) to check if our program gives satisfactory results, including some Central African languages with different word orders in positive and negative clauses. And our method gives acceptable results. We explain our results and propose some languages with different word orders in positive and negative clauses. After looking up grammar books, we ensure one language out of three has this feature. Also, some possible ways to improve the performance of this method are described.

In this thesis, different aspects concerning how to make synthetic talking faces more expressive have been studied. How can we collect data for the studies, how is the lip articulation affected by expressive speech, can the recorded data be used interchangeably in different face models, can we use eye movements in the agent for communicative purposes? The work of this thesis includes studies of these questions and also an experiment using a talking head as a complement to a targeted audio device, in order to increase the intelligibility of the speech. The data collection described in the first paper resulted in two multimodal speech corpora. In the following analysis of the recorded data it could be stated that expressive modes strongly affect the speech articulation, although further studies are needed in order to acquire more quantitative results and to cover more phonemes and expressions as well as to be able to generalise the results to more than one individual. When switching the files containing facial animation parameters (FAPs) between different face models (as well as research sites), some problematic issues were encountered despite the fact that both face models were created according to the MPEG-4 standard. The evaluation test of the implemented emotional expressions showed that best recognition results were obtained when the face model and FAP-file originated from the same site. The perception experiment where a synthetic talking head was combined with a targeted audio, parametric loudspeaker showed that the virtual face augmented the intelligibility of speech, especially when the sound beam was directed slightly to the side of the listener i. e. at lower sound intesities. In the experiment with eye gaze in a virtual talking head, the possibility of achieving mutual gaze with the observer was assessed. The results indicated that it is possible, but also pointed at some design features in the face model that need to be altered in order to achieve a better control of the perceived gaze direction. / QC 20101126

When advertisements are placed on web pages, the context in which the advertisements are presented is important. For example, manufacturers of kitchen knives may not want their advertisement to appear in a news article about a knife-wielding murderer. The purpose of the current work is to explore the ability of pre-trained language models on text classification tasks for determining whether the content of a given article is brand-safe, that is, suitable for brand advertising. A Norwegian-language news dataset containing 3600 news items was manually labelled with negative topics. Five pre-trained BERT language models were tested, including one multilingual BERT and four language models pre-trained specifically on Norwegian. Different training settings and fine-tuning methods were also tested for two best-performing models. It was found that more structurally complex language models and language models trained on corpora that were large or had larger vocabularies performed better on the text classification task during testing. However, the performance of smaller models is also acceptable if there is a trade-off between the better performance and the time and processing power required. As far as training and fine-tuning settings are concerned, this work found that for news texts, the initial part of the articles, which often contain the most information, is the optimal choice of parts as input to the model BERT. Another achievement and contribution of this work was the manual tagging of a Norwegian news dataset on negative topics.        This thesis also points to some possible directions for future work, such as experimenting with different label granularity, experimenting with multilingual controlled training, and training with few samples.

In this thesis, the application of transformer-based models on the real-world task of identifying patients as implant bearing is investigated. The task is approached as a classification task and five transformer-based models relying on the BERT architecture are implemented, along with a Support Vector Machine (SVM) as a baseline for comparison. The models are fine-tuned with Swedish medical texts, i.e. patients’ medical histories. The five transformer-based models in question makes use of two pre-trained BERT models, one released by the National Library of Sweden and a second one using the same pre-trained model but which has also been further pre-trained on domain specific language. These are in turn fine-tuned using five different types of architectures. These are: (1) a typical BERT model, (2) GAN-BERT, (3) RoBERT, (4) chunkBERT, (5) a frequency based optimized BERT. The final classifier, an SVM baseline, is trained using TF-IDF as the feature space. The data used in the thesis comes from a subset of an unreleased corpus from four Swedish clinics that cover a span of five years. The subset contains electronic medical records of patients belonging to the radiology, and cardiology clinics. Four training sets were created, respectively containing 100, 200, 300, and 903 labelled records. The test set, containing 300 labelled samples, was also created from said subset. The labels upon which the models are trained are created by labelling the patients as implant bearing based on the amount of implant terms each patient history contain. The results are promising, and show favourable performance when classifying the patient histories. Models trained on 903 and 300 samples are able to outperform the baseline, and at their peak, BERT, chunkBERT and the frequency based optimized BERT achieves an F1-measure of 0.97. When trained using 100 and 200 labelled records all of the transformerbased models are outperformed by the baseline, except for the semi-supervised GAN-BERT which is able to achieve competitive scores with 200 records. There is not a clear delineation between using the pre-trained BERT or the BERT model that has additional pre-training on domain specific language. However, it is believed that further research could shed additional light on the subject since the results are inconclusive. / Patient-Safe Magnetic Resonance Imaging Examination by AI-based Medical Screening

This report researches a method for creating knowledge graphs, a specific way of structuring information, using distributional semantic models. Two different algorithms for selecting graph edges and two different algorithms for labelling edges are tried, and variations of those are evaluated. We perform experiments comparing our knowledge graphs with existing manually constructed knowledge graphs of high quality, with respect to graph structure and edge labels. We find that the algorithms usually produces graphs with a structure similar to that of manually constructed knowledge graphs, as long as the data set is sufficiently large and general, and that the similarity of edge labels to manually chosen edge labels vary widely depending on input.

This thesis investigates the enhancement of legal contract Natural Language Inference (NLI) classification through supervised fine-tuning on general domain NLI, in the case of ContractNLI and Span NLI BERT (Koreeda and Manning, 2021), a multi-task document information extraction dataset and framework. Annotated datasets of a specific professional domain are scarce due to the high time and labour cost required to create them. Since NLI is a simple yet effective task in inducing and evaluating natural language understanding (NLU) abilities in language models, there is potential in leveraging abundant general domain NLI datasets to aid information extraction and classification for legal contracts. This work evaluates the impact of transfer learning from Adversarial NLI (Nie et al.,2020) from the general domain to ContractNLI, via sequential and mixed batch fine-tuning. The study also extends its investigation to the effects of the model’s evidence identification component on NLI, by fine-tuning on the Contract Understanding Atticus Dataset (Hendrycks et al., 2021). The results highlight the benefits of fine-tuning with general domain NLI data, particularly for hypotheses in the target task with balanced entailment and contradiction training examples. In addition, the study demonstrates the reciprocal relationship between evidence identification and NLI classification, where improvements in the former enhance the accuracy of the latter. With NLI being more commonly applied to information extraction settings in specialised domains, this work sheds light on the potential impacts of existing general domain NLI resources in stepping up classification performance in specific domains.

Metaphoric Paraphrase generation is a relatively new and unexplored Natural Language Generation task. The aim of the task is to develop computational systems that paraphrase literal sentences into cogent metaphoric ones. Challenges in the field include representation of common sense knowledge and ensuring meaning retention when dealing with phrases that are dissimilar in their literal sense. This thesis will deal with the specific task of paraphrasing literal adjective phrases into metaphoric noun phrases, taking into consideration the preceding context of the adjective phrase. Two different systems were developed as part of this study. The systems are identical, apart from the fact that one is endowed with a knowledge representation based on Conceptual Metaphor Theory. The paraphrases generated by the systems, along with paraphrases written by a native speaker of English, were scored on the parameters of meaning retention and creativity by a crowd-sourced panel. Both systems were able to generate cogent metaphoric paraphrases, although fairly unreliably compared to the human. The system endowed with Conceptual Metaphor Theory knowledge got a lower average meaning retention score and a higher average creativity score than the system without Conceptual Metaphor Theory knowledge representation. In addition to that it was found that less similarity in sentence embeddings of literal sentences and metaphoric paraphrases of them correlates with a higher level of perceived meaning retention and a lower perceived creativity of the metaphoric paraphrase. It was also found that less difference in GPT-2 log probability between literal sentences and metaphoric paraphrases of them correlates with humans evaluating the paraphrases as less creative.

In this thesis, we explore the impact of M-BERT and different transfer sizes on the choice of different transfer languages in dependency parsing. In order to investigate our research questions, we conduct a series of experiments on the treebanks in Universal Dependencies with UUParser.     The main conclusions and contributions of this study are as follows:   First, we train a variety of languages in several different scripts with M-BERT being added into the parsing framework, which is one of the most state-of-the-art deep learning models based on the Transformer architecture. In general, we get advancing results with M-BERT compared with the randomly initialized embedding in UUParser.    Second, since it is a common way to choose a source language, which is 'close' to the target language in cross-lingual parsing, we try to explore what 'close' languages actually are, as there is not a definition for 'close'. In our study, we explore how strongly the parsing results are correlated with the different linguistic distances between the source and target languages. The relevant data is queried from URIEL Database. We find that the parsing performance is more dependent on inventory, syntactic and featural distance than on the geographic, genetic and phonological distance in zero-shot experiments. In the few-shot prediction, the parsing accuracy shows stronger correlation with inventory and syntactic distance than with others.     Third, we vary the training sizes in few-shot experiments with M-BERT being added to see how the parsing results are influenced. We find that it is very obvious that few-shot experiments outperform zero-shot experiments. With the source sizes being cut, all parsing scores decrease. However, we do not see a linear drop of the results.

Text normalization plays a crucial role in text-to-speech systems by ensuring that the input text is in an appropriate format and consists of standardized words prior to grapheme-to-phoneme conversion for text-to-speech. The aim of this study was to assess the performance of five text normalization systems based on different methods. These text normalization systems were evaluated on the English Google text normalization dataset. The evaluation was based on the similarity between the ground truth and normalized outputs from each text normalization system. Since multiple ground truth issues occurred during the evaluation, the original similarity scores needed to be manually re-scored. The re-scoring was employed on a sample data semi-randomly extracted from the evaluation dataset. According to the results, the accuracy of these text normalization systems  can be ranked as follows: the Duplex system, the Hybrid system, the VT system, the RS system, and the WFST system. For the two rule-based systems from ReadSpeaker, the VT system performed slightly better than the RS system, with a slight difference in the original similarity score. By analyzing the error patterns produced during the normalization process, the study provided valuable insights into the strengths and limitations of these systems. The findings of this study contribute to the refinement of internal rules, leading to improved accuracy and effectiveness of text normalization in text-to-speech applications.

This paper describes the process of building a machine translation lexicon for use in the train and transport domain with the machine translation system MATS. The lexicon will consist of a Swedish part, an English part and links between them and is derived from a Trados translation memory which is split into a training(90%) part and a testing(10%) part. The task is carried out mainly by using existing word linking software and recycling previous machine translation lexicons from other domains. In order to do this, a method is developed where focus lies on automation by means of both existing and self developed software, in combination with manual interaction. The domain specific lexicon is then extended with a domain neutral core lexicon and a less domain neutral general lexicon. The different lexicons are automatically and manually evaluated through machine translation on the test corpus. The automatic evaluation of the largest lexicon yielded a NEVA score of 0.255 and a BLEU score of 0.190. The manual evaluation saw 34% of the segments correctly translated, 37%, although not correct, perfectly understandable and 29% difficult to understand.