In this paper, we present our methodology for supervised stance classification of sparse and imbalanced social media data. We test our framework on a manually labeled dataset of 5700 messages about immigration in the Swedish language posted on the Flashback forum, a controversial online discussion platform. Our proposed approach currently achieves a macro- averaged F1-score of 0.72 for test data on a two-class problem compared against 0.27 for a baseline four-class model. Since effective classification of imbalanced and sparse textual data in under-resourced languages presents certain methodological challenges, our study contributes to a discussion on the best pathways to achieve highest model performance given the character of the data and unavailability of large training datasets for this task. Moreover, this work exemplifies the application of ML methodology to social media data, which can be particularly relevant for social scientists working in this area and interested in leveraging the possibilities of machine learning in their research field. This methodology and the obtained results provide a foundation for further in-depth analyses of social media texts in the Swedish language following a data-driven approach.

In this paper, we present our methodology for supervised stance classification of sparse and imbalanced social media data. We test our framework on a manually labeled dataset of 5700 messages about immigration in the Swedish language posted on the Flashback forum, a controversial online discussion platform. Our proposed approach currently achieves a macro- averaged F1-score of 0.72 for test data on a two-class problem compared against 0.27 for a baseline four-class model. Since effective classification of imbalanced and sparse textual data in under-resourced languages presents certain methodological challenges, our study contributes to a discussion on the best pathways to achieve highest model performance given the character of the data and unavailability of large training datasets for this task. Moreover, this work exemplifies the application of ML methodology to social media data, which can be particularly relevant for social scientists working in this area and interested in leveraging the possibilities of machine learning in their research field. This methodology and the obtained results provide a foundation for further in-depth analyses of social media texts in the Swedish language following a data-driven approach.

In this work, we explore the performance of supervised stance classification methods for social media texts in under-resourced languages and using limited amounts of labeled data. In particular, we focus specifically on the possibilities and limitations of the application of classic machine learning versus deep learning in social sciences. To achieve this goal, we use a training dataset of 5.7K messages posted on Flashback Forum, a Swedish discussion platform, further supplemented with the previously published ABSAbank-Imm annotated dataset, and evaluate the performance of various model parameters and configurations to achieve the best training results given the character of the data. Our experiments indicate that classic machine learning models achieve results that are on par or even outperform those of neural networks and, thus, could be given priority when considering machine learning approaches for similar knowledge domains, tasks, and data. At the same time, the modern pre-trained language models provide useful and convenient pipelines for obtaining vectorized data representations that can be combined with classic machine learning algorithms. We discuss the implications of their use in such scenarios and outline the directions for further research.

In this work, we explore the performance of supervised stance classification methods for social media texts in under-resourced languages and using limited amounts of labeled data. In particular, we focus specifically on the possibilities and limitations of the application of classic machine learning versus deep learning in social sciences. To achieve this goal, we use a training dataset of 5.7K messages posted on Flashback Forum, a Swedish discussion platform, further supplemented with the previously published ABSAbank-Imm annotated dataset, and evaluate the performance of various model parameters and configurations to achieve the best training results given the character of the data. Our experiments indicate that classic machine learning models achieve results that are on par or even outperform those of neural networks and, thus, could be given priority when considering machine learning approaches for similar knowledge domains, tasks, and data. At the same time, the modern pre-trained language models provide useful and convenient pipelines for obtaining vectorized data representations that can be combined with classic machine learning algorithms. We discuss the implications of their use in such scenarios and outline the directions for further research.

An overview of the most common ontological resources and methods of their construction and application is given. For purposes of scientific research we analyzed the characteristics of ontologies in the public domain and corpus containing criminal context. Additionally, we have recently developed a Flask-based web application that generates ontologies using the Anytree library.

The authors also developed a multilingual basic ontology called "Illegal Web content" based on a corpus of texts in criminal context in English, Ukrainian, Kazakh and Russian languages. The development of this ontology was motivated by the need for effective analysis and prevention of criminal activities based on textual information disseminated on the internet. The newly developed web application allows users to create ontologies by importing text files in different languages, and then automatically generates an ontology based on the text. The application is user-friendly, and allows users to customize the ontology by adding or removing nodes, changing the labels of nodes and edges, and setting the weight of edges.

Overall, the development of the "Illegal Web content" ontology and the web application represents a significant contribution to the field of ontology development and text processing for criminal investigation and prevention. The main characteristics of the Web application, including its ease of use and customizability, make it a valuable tool for researchers and practitioners alike.

Communication with customers is a major part of customer experience as well as a great source of data mining. More businesses are engaging with consumers via text messages. Before 2020, 39% of businesses already use some form of text messaging to communicate with their consumers. Many more were expected to adopt the technology after 2020[1]. Email response rates are merely 8%, compared to a response rate of 45% for text messaging[2]. A significant portion of this communication involves customer enquiries or support messages sent in both directions.

According to estimates, more than 80% of today’s data is stored in an unstructured format (suchas text, image, audio, or video) [3], with a significant portion of it being stated in ambiguous natural language. When analyzing such data, qualitative data analysis techniques are usually employed. In order to facilitate the automated examination of huge corpora of textual material, researchers have turned to natural language processing techniques[4].

Under the light of shared statistics above, Billogram[5] has decided that support messages between creditors and recipients can be mined for predictive maintenance purposes, such as early identification of an outlier like a bug, defect, or wrongly built feature. As one sentence goal definition, Billogram is looking for an answer to ”why are people reaching out to begin with?” This thesis project discusses implementing unsupervised clustering of support messages by benefiting from natural language processing methods as well as performance metrics of results to answer Billogram’s question. The research also contains intent recognition of clustered messages in two different ways, one automatic and one semi-manual, the results have been discussed and compared.

LDA and manual intent assignment approach of the first research has 100 topics and a 0.293 coherence score. On the other hand, the second approach produced 158 clusters with UMAP and HDBSCAN while intent recognition was automatic. Creating clusters will help identifying issues which can be subjects of increased focus, automation, or even down-prioritizing. Therefore, this research lands in the predictive maintenance[9] area. This study, which will get better over time with more iterations in the company, also contains the preliminary work for ”labeling” or ”describing”clusters and their intents.

In this thesis, we explore lyrics generation as a subbranch of computational creativity.

The dataset is comprised of 16,858 Chinese songs from 160 singers. Compared to the baseline which uses a character-level Recurrent Neural Network (RNN), we apply a sequence to sequence model with a bidirectional encoder and a decoder with attention mechanism. In addition, we introduce new standards to evaluate the quality of generated lyrics: lyricism,

accuracy, singability and coherence. Two quality tests are included: human judgment and selecting which lyrics are written by human lyricists.

In the evaluation, the average scores for our model perform better than the baseline in all standards. The average probability of a human judging lyrics generated by our model as created by a human is 17.23%.

In this work, we explore the usefulness of contextualized embeddings from language models on lexical semantic change (LSC) detection. With diachronic corpora spanning two time periods, we construct word embeddings for a selected set of target words, aiming at detecting potential LSC of each target word across time. We explore different systems of embeddings to cover three topics: contextualized vs static word embeddings, token- vs type-based embeddings, and multilingual vs monolingual language models.

We use a multilingual dataset covering three languages (English, German, Swedish) and explore each system of embedding with two subtasks, a binary classification task and a ranking task. We compare the performance of different systems of embeddings, and seek to answer our research questions through discussion and analysis of experimental results.

We show that contextualized word embeddings are on par with static word embeddings in the classification task. Our results also show that it is more beneficial to use the contextualized embeddings from a multilingual model than from a language specific model in most cases. We present that token-based setting is strong for static embeddings, and type-based setting for contextual embeddings, especially for the ranking task.

We provide some explanation for the results we achieve, and propose improvements that can be made to our experiments for future work.

We describe the Uppsala NLP submission to SemEval-2021 Task 2 on multilingual and cross-lingual word-in-context disambiguation. We explore the usefulness of three pre-trained multilingual language models, XLM-RoBERTa (XLMR), Multilingual BERT (mBERT) and multilingual distilled BERT (mDistilBERT). We compare these three models in two setups, fine-tuning and as feature extractors. In the second case we also experiment with using dependency-based information. We find that fine-tuning is better than feature extraction. XLMR performs better than mBERT in the cross-lingual setting both with fine-tuning and feature extraction, whereas these two models give a similar performance in the multilingual setting. mDistilBERT performs poorly with fine-tuning but gives similar results to the other models when used as a feature extractor. We submitted our two best systems, fine-tuned with XLMR and mBERT.

Since the advent of automatic evaluation, tasks within Natural Language Processing (NLP), including Machine Translation, have been able to better utilize both time and labor resources. Later, multilingual pre-trained models (MLMs)have uplifted many languages’ capacity to participate in NLP research. Contextualized representations generated from these MLMs are both influential towards several downstream tasks and have inspired practitioners to better make sense of them. We propose the adoption of BERTScore, coupled with contrastive learning, for machine translation evaluation in lieu of BLEU - the industry leading metric. While BERTScore computes a similarity score for each token in a candidate and reference sentence, it does away with exact matches in favor of computing token similarity using contextual embeddings. We improve BERTScore via contrastive learning-based fine-tuning on MLMs. We use contrastive learning to improve BERTScore across different language pairs in both high and low resource settings (English-Hausa, English-Chinese), across three models (XLM-R, mBERT, and LaBSE) and across three domains (news,religious, combined). We also investigated both the effects of pairing relatively linguistically similar low-resource languages (Somali-Hausa), and data size on BERTScore and the corresponding Pearson correlation to human judgments. We found that reducing the distance between cross-lingual embeddings via contrastive learning leads to BERTScore having a substantially greater correlation to system-level human evaluation than BLEU for mBERT and LaBSE in all language pairs in multiple domains.

In this thesis, we mainly investigate the influence of using unsupervised morphological segmentation as features on the dependency parsing of morphologically rich languages such as Finnish, Estonian, Hungarian, Turkish, Uyghur, and Kazakh. Studying the morphology of these languages is of great importance for the dependency parsing of morphologically rich languages since dependency relations in a sentence of these languages mostly rely on morphemes rather than word order. In order to investigate our research questions, we have conducted a large number of parsing experiments both on MaltParser and UDPipe. We have generated the supervised morphology and the predicted POS tags from UDPipe, and obtained the unsupervised morphological segmentation from Morfessor, and have converted the unsupervised morphological segmentation into features and added them to the UD treebanks of each language. We have also investigated the different ways of converting the unsupervised segmentation into features and studied the result of each method. We have reported the Labeled Attachment Score (LAS) for all of our experimental results.

The main finding of this study is that dependency parsing of some languages can be improved simply by providing unsupervised morphology during parsing if there is no manually annotated or supervised morphology available for such languages. After adding unsupervised morphological information with predicted POS tags, we get improvement of 4.9%, 6.0%, 8.7%, 3.3%, 3.7%, and 12.0% on the test set of Turkish, Uyghur, Kazakh, Finnish, Estonian, and Hungarian respectively on MaltParser, and the parsing accuracies have been improved by 2.7%, 4.1%, 8.2%, 2.4%, 1.6%, and 2.6% on the test set of Turkish, Uyghur, Kazakh, Finnish, Estonian, and Hungarian respectively on UDPipe when comparing the results from the models which do not use any morphological information during parsing.

Topic models for focused analysis aim to capture topics within the limiting scope of a targeted aspect (which could be thought of as some inner topic within a certain domain). To serve their analytic purposes, topics are expected to be semantically-coherent and closely aligned with human intuition – this in itself poses a major challenge for the more common topic modeling algorithms which, in a broader sense, perform a full analysis that covers all aspects and themes within a collection of texts. The paper attempts to construct a viable focused-analysis topic model which learns topics from Twitter data written in a closely related group of non-standardized varieties of Arabic widely spoken in the Levant region (i.e Levantine Arabic). Results are compared to a baseline model as well as another targeted topic model designed precisely to serve the purpose of focused analysis. The model is capable of adequately capturing topics containing terms which fall within the scope of the targeted aspect when judged overall. Nevertheless, it fails to produce human-friendly and semantically-coherent topics as several topics contained a number of intruding terms while others contained terms, while still relevant to the targeted aspect, thrown together seemingly at random.

The acoustic characteristics of so-called 'dist' tones, commonly used in singing rock music, are analyzed in a case study. In an initial experiment a professional rock singer produced examples of 'dist' tones. The tones were found to contain aperiodicity, SPL at 0.3 m varied between 90 and 96 dB, and subglottal pressure varied in the range of 20-43 cm H₂O, a doubling yielding, on average, an SPL increase of 2.3 dB. In a second experiment, the associated vocal fold vibration patterns were recorded by digital high-speed imaging of the same singer. Inverse filtering of the simultaneously recorded audio signal showed that the aperiodicity was caused by a low frequency modulation of the flow glottogram pulse amplitude. This modulation was produced by an aperiodic or periodic vibration of the supraglottic mucosa. This vibration reduced the pulse amplitude by obstructing the airway for some of the pulses produced by the apparently periodically vibrating vocal folds. The supraglottic mucosa vibration can be assumed to be driven by the high airflow produced by the elevated subglottal pressure.

Using statistical approaches beside the traditional methods of natural language processing could significantly improve both the quality and performance of several natural language processing (NLP) tasks. The effective usage of these approaches is subject to the availability of the informative, accurate and detailed corpora on which the learners are trained. This article introduces a bootstrapping method for developing annotated corpora based on a complex and rich linguistically motivated elementary structure called supertag. To this end, a hybrid method for supertagging is proposed that combines both of the generative and discriminative methods of supertagging. The method was applied on a subset of Wall Street Journal (WSJ) in order to annotate its sentences with a set of linguistically motivated elementary structures of the English XTAG grammar that is using a lexicalised tree-adjoining grammar formalism. The empirical results confirm that the bootstrapping method provides a satisfactory way for annotating the English sentences with the mentioned structures. The experiments show that the method could automatically annotate about 20% of WSJ with the accuracy of F-measure about 80% of which is particularly 12% higher than the F-measure of the XTAG Treebank automatically generated from the approach proposed by Basirat and Faili [(2013). Bridge the gap between statistical and hand-crafted grammars. Computer Speech and Language, 27, 1085-1104].

The world harbours a diversity of some 6,500 mutually unintelligible languages. As has been increasingly observed by linguists, many minority languages are becoming endangered and will be lost forever if not documented. The increased urgency has led to the development of several global endangerment databases and a more fine-grained understanding of the language endangerment progression as well as its possible reversal. In the present paper, we explore the terminological correlates of this development as found in the descriptive linguistic literature, using a corpus of over 10,000 digitized grammatical descriptions. Comparing this with existing endangerment databases, we find that simply counting terms related to endangerment does signal endangerment, but the degree of endangerment is more difficult to assess from grammatical descriptions. The label endangered seems to be an umbrella term that covers different situations ranging from moribund languages with less than ten speakers to minority languages with several thousand speakers. For many languages considered endangered in existing databases, explicit terms to this effect cannot be found in their descriptions. The discrepancy is due to incompleteness of the searchterm set, gaps in the literature, and projected rather than observed information in the databases. Our explorations illustrate the potential for database curation assisted by computational searches both to maintain accuracy of the databases and to investigate assumed language endangerment. Future work includes a larger cloud of search terms, usage of term frequencies, and prescreening of descriptive literature for the existence of a relevant section. From the perspective of descriptive linguistics, this study calls for a more careful correlation between the language endangerment indexes, as developed in the global endangerment databases, and the treatment of the endangerment status of individual languages in descriptive grammars.

This thesis explores the foundations of text classification, using both empirical and deductive methods, with a focus on author identification and syntactic methods. We strive for a thorough theoretical understanding of what affects the effectiveness of classification in general. 

To begin with, we systematically investigate the effects of some parameters on the accuracy of author identification. How is the accuracy affected by the number of candidate authors, and the amount of data per candidate? Are there differences in how methods react to the changes in parameters? Using the same techniques, we see indications that methods previously thought to be topic-independent might not be so, but that syntactic methods may be the best option for avoiding topic dependence. This means that previous studies may have overestimated the power of lexical methods. We also briefly look for ways of spotting which particular features might be the most effective for classification. Apart from author identification, we apply similar methods to identifying properties of the author, including age and gender, and attempt to estimate the number of distinct authors in a text sample. In all cases, the techniques are proven viable if not overwhelmingly accurate, and we see that lexical and syntactic methods give very similar results. 

In the final parts, we see some results of automata theory that can be of use for syntactic analysis and classification. First, we generalise a known algorithm for finding a list of the best-ranked strings according to a weighted automaton, to doing the same with trees and a tree automaton. This result can be of use for speeding up parsing, which often runs in several steps, where each step needs several trees from the previous as input. Second, we use a compressed version of deterministic finite automata, known as failure automata, and prove that finding the optimal compression is NP-complete, but that there are efficient algorithms for finding good approximations. Third, we find and prove the derivatives of regular expressions with cuts. Derivatives are an operation on expressions to calculate the remaining expression after reading a given symbol, and cuts are an extension to regular expressions found in many programming languages. Together, these findings may be able to improve on the syntactic analysis which we have seen is a valuable tool for text classification.

Derivatives of regular expressions are an operation which for a given expression produces an expression for what remains after a specific symbol has been read. This can be used as a step in transforming an expression into a finite string automaton. Cuts are an extension of the ordinary regular expressions; the cut operator is essentially a concatenation without backtracking, formalising a behaviour found in many programming languages. Just as for concatenation, we can also define an iterated cut operator. We show and derive expressions for the derivatives of regular expressions with cuts and iterated cuts.

In order to ensure smooth turn-taking between conversational participants, interlocutors must have ways of providing information to one another about whether they have finished speaking or intend to continue. The current work investigates Swedish speakers’ use of hand gestures in conjunction with turn change or turn hold in unrestricted, spontaneous speech. As has been reported by other researchers, we find that speakers’ gestures end before the end of speech in cases of turn change, while they may extend well beyond the end of a given speech chunk in the case of turn hold. We investigate the degree to which prosodic cues and gesture cues to turn transition in Swedish face-to-face conversation are complementary versus functioning additively. The co-occurrence of acoustic prosodic features and gesture at potential turn boundaries gives strong support for considering hand gestures as part of the prosodic system, particularly in the context of discourse-level information such as maintaining smooth turn transition.

In a number of languages, intonational patterns at prosodic boundaries are considered to be relevant for turn transition or turn hold. A perception experiment tested the influence of fundamental frequency (F0) peak height and rising final contours on Swedish listeners’ judgment about whether a speaker wanted to hold the turn. While F0 peak height, as has been previously shown, did influence listeners’ judgments, the end height of rising pitch tails apparently did not influence listeners’ judgments about whether a speaker planned to continue talking, even though they showed sensitivity to the differences in a discrimination task. The differences in responses in the tasks, as well as the difference from results found for other languages, may indicate that listeners used comparative prominence to guide their judgments, rather than intonation playing a direct role in the turn-transition system.

In many cases of turn transition in conversation, a new speaker may respond to phonetic cues from the end of the prior turn, including variation in prosodic features such as pitch and final lengthening. Although consistent pitch and lengthening features are well-established for some languages at potential points of turn transition, this is not necessarily the case for Swedish. The current study uses a two-alternative forced choice task to investigate how variation in pitch contour and lengthening at the ends of syntactically complete turns can influence listeners’ expectations of turn hold or turn transition. Both lengthening and pitch contour features were found to influence listeners’ judgments about whether turn transition would occur, with shorter length and higher final pitch peaks associated with turn hold. Furthermore, listeners were more certain about their judgments when asked about turn-hold rather than turn-change, suggesting an imbalance in the strength of turn-hold versus turn-transition cues.

Speakers and listeners have been shown to use phonetic cues to help them in tracking the ongoing structure of conversational interaction, but fragmentation between qualitative and quantitative research means that the forms and functions of these cues have been given varying characterizations. The current study explores prosodic variation in contrastive structures in conversational data, using a combined methodology adopting aspects from both qualitative (conversation analysis) and quantitative (experimental phonetics/phonology) approaches. Statistical and conversation-analytical methods used together reveal relationships between prosodic variation and interactional function, such as variations in pitch range across adjacent turns being linked to the presence of 'stepwise' topic changes.

A professional impersonator has been studied when training his voice tomimic two target speakers. A three-fold investigation has been conducted; acomputer-based speaker verification system was used, phonetic-acousticmeasurements were made and a perception test was conducted. Our ideabehind using this type of system is to measure how close to the target voice aprofessional impersonation might be able to reach and to relate this tophonetic-acoustic analyses of the mimic speech and human perception. Thesignificantly increased verification scores and the phonetic-acoustic analysesshow that the impersonator really changes his natural voice and speech in hisimitations. The results of the perception test show that there is no, or only asmall, correlation between the verification system and the listeners whenestimating the voice imitations and how close they are to one of the targetspeakers.

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. 

Large-scale pre-trained Vision Language Models (VLMs) have proven effective for zero-shot classification. Despite the success, most traditional VLMs-based methods are restricted by the assumption of partial source supervision or ideal target vocabularies, which rarely satisfy the open-world scenario. In this paper, we aim at a more challenging setting, Realistic Zero-Shot Classification, which assumes no annotation but instead a broad vocabulary. To address the new problem, we propose the Self Structural Semantic Alignment (S3A) framework, which extracts the structural semantic information from unlabeled data while simultaneously selflearning. Our S3A framework adopts a unique Cluster-Vote-Prompt-Realign (CVPR) algorithm, which iteratively groups unlabeled data to derive structural semantics for pseudo-supervision. Our CVPR algorithm includes iterative clustering on images, voting within each cluster to identify initial class candidates from the vocabulary, generating discriminative prompts with large language models to discern confusing candidates, and realigning images and the vocabulary as structural semantic alignment. Finally, we propose to self-train the CLIP image encoder with both individual and structural semantic alignment through a teacher-student learning strategy. Our comprehensive experiments across various generic and fine-grained benchmarks demonstrate that the S3A method substantially improves over existing VLMs-based approaches, achieving a more than 15% accuracy improvement over CLIP on average. Our codes, models, and prompts are publicly released at https://github.com/shengeatamath/S3A.

This thesis aims to perform named entity recognition for English social media texts. Named Entity Recognition (NER) is applied in many NLP tasks as an important preprocessing procedure. Social media texts contain lots of real-time data and therefore serve as a valuable source for information extraction. Nevertheless, NER for social media texts is a rather challenging task due to the noisy context. Traditional approaches to deal with this task use hand-crafted features but prove to be both time-consuming and very task-specific. As a result, they fail to deliver satisfactory performance. The goal of this thesis is to tackle this task by automatically identifying and annotating the named entities with multiple types with the help of neural network methods. In this thesis, we experiment with three different word embeddings and character embedding neural network architectures that combine long short- term memory (LSTM), bidirectional LSTM (BI-LSTM) and conditional random field (CRF) to get the best result. The data and evaluation tool comes from the previous shared tasks on Noisy User-generated Text (W- NUT) in 2017. We achieve the best F1 score 42.44 using BI-LSTM-CRF with character-level representation extracted by a BI-LSTM, and pre-trained word embeddings trained by GloVe. We also find out that the results could be improved with larger training data sets.

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.

Multilingual natural language processing (NLP) is increasingly recognized for its potential in processing diverse text-type data, including those from social media, reviews, and technical reports. Multilingual language models like mBERT and XLM-RoBERTa (XLM-R) play a pivotal role in multilingual NLP. Notwithstanding their capabilities, the performance of these models largely relies on the availability of annotated training data. This thesis employs the multilingual pre-trained model XLM-R to examine its efficacy in sequence labelling to open-ended questions on democracy across multilingual surveys. Traditional annotation practices have been labour-intensive and time-consuming, with limited automation attempts. Previous studies often translated multilingual data into English, bypassing the challenges and nuances of native languages. Our study explores automatic multilingual annotation at the token level for democracy survey responses in five languages: Hungarian, Italian, Polish, Russian, and Spanish. The results reveal promising F1 scores, indicating the feasibility of using multilingual models for such tasks. However, the performance of these models is closely tied to the quality and nature of the training set. This research paves the way for future experiments and model adjustments, underscoring the importance of refining training data and optimizing model techniques for enhanced classification accuracy.

With podcasts being a fast growing audio-only form of media, an effective way of promoting different podcast shows becomes more and more vital to all the stakeholders concerned, including the podcast creators, the podcast streaming platforms, and the podcast listeners. This thesis investigates the relatively little studied topic of automatic podcast trailer generation, with the purpose of en- hancing the overall visibility and publicity of different podcast contents and gen- erating more user engagement in podcast listening. This thesis takes a hotspot- based approach, by specifically defining the vague concept of “hotspot” and designing different appropriate methods for hotspot detection. Different meth- ods are analyzed and compared, and the best methods are selected. The selected methods are then used to construct an automatic podcast trailer generation sys- tem, which consists of four major components and one schema to coordinate the components. The system can take a random podcast episode audio as input and generate an around 1 minute long trailer for it. This thesis also proposes two human-based podcast trailer evaluation approaches, and the evaluation results show that the proposed system outperforms the baseline with a large margin and achieves promising results in terms of both aesthetics and functionality.

The visual exploration of large and complex network structures remains a challenge for many application fields. Moreover, a growing number of real world networks are multivariate and often interconnected with each other. Entities in a network may have relationships with elements of other related data sets, which do not necessarily have to be networks themselves, and these relationships may be defined by attributes that can vary greatly. In this work, we propose a comprehensive visual analytics approach that supports researchers to specify and subsequently explore attribute-based relationships across networks, text documents, and derived secondary data. Our approach provides an individual search functionality based on keywords and semantically similar terms over the entire text corpus to find related network nodes. For examining these nodes in the interconnected network views, we introduce a new interaction technique, called Hub2Go, which facilitates the navigation by guiding the user to the information of interest. To showcase our system, we use a large text corpus collected from research papers listed in the IEEE VIS publications dataset that consists of 2752 documents over a period of 25 years. Here, we analyze relationships between various heterogeneous networks, a Bag-of-Words index, and a word similarity matrix, all derived from the initial corpus and metadata. 

In this paper, we present our process to establish a PICO and a sentiment annotated corpus of clinical trial publications. PICO stands for Population, Intervention, Comparison and Outcome - these four classes can be used for more advanced and specific search queries. For example, a physician can determine how well a drug works only in the subgroup of children. Additionally to the PICO extraction, we conducted a sentiment annotation, where the sentiment refers to whether the conclusion of a trial was positive, negative or neutral. We created both corpora with the help of medical experts and non-experts as annotators.

Foreign words, such as names, locations or sometimes entire phrases, are a problem for any system that is meant to convert graphemes to phonemes (g2p; i.e.converting written text into phonetic transcription). In this thesis, we investigate both rule-based and neural methods of phonetically transcribing English words found in Icelandic text, taking into account the rules and constraints of how foreign phonemes can be mapped into Icelandic phonology.

We implement a rule-based system by compiling grammars into finite-state transducers. In deciding on which rules to include, and evaluating their coverage, we use a list of the most frequently-found English words in a corpus of Icelandic text. The output of the rule-based system is then manually evaluated and corrected (when needed) and subsequently used as data to train a simple bidirectional LSTM g2p model. We train models both with and without length and stress labels included in the gold annotated data.

Although the scores for neither model are close to the state-of-the-art for either Icelandic or English, both our rule-based system and LSTM model show promising initial results and improve on the baseline of simply using an Icelandic g2p model, rule-based or neural, on English words. We find that the greater flexibility of the LSTM model seems to give it an advantage over our rule-based system when it comes to modeling certain phenomena. Most notable is the LSTM’s ability to more accurately transcribe relations between graphemes and phonemes for English vowel sounds.

Given there does not exist much previous work on g2p transcription specifically handling English words within the Icelandic phonological constraints and it remains an unsolved task, our findings present a foundation for the development of further research, and contribute to improving g2p systems for Icelandic as a whole.

Textual duplicates can be hard to detect as they differ in words but have similar semantic meaning. At Etteplan, a technical documentation company, they have many writers that accidentally re-write existing instructions explaining procedures. These "duplicates" clutter the database.

This is not desired because it is duplicate work. The condition of the database will only deteriorate as the company expands. This thesis attempts to map where the problem is worst, and also how to calculate how many duplicates there are.

The corpus is small, but written in a controlled natural language called Simplified Technical English. The method uses document embeddings from doc2vec and clustering by use of HDBSCAN* and validation using Density-Based Clustering Validation index (DBCV), to chart the problems. A survey was sent out to try to determine a threshold value of when documents stop being duplicates, and then using this value, a theoretical duplicate count was calculated.

This paper explores how interfaces that fully uses our ability to communicate through the visual, auditory, and tactile senses, may enhance mobile interaction. The first step is to look beyond the desktop. We do not need to reinvent computing, but we need to see that mobile interaction does not benefit from desktop metaphors alone. The next step is to look at what we have at hand, and as we will see, mobile devices are already quite apt for multimodal interaction. The question is how we can coordinate information communicated through several senses in a way that enhances interaction. By mapping information over communication circuit, semiotic representation, and sense applied for interaction; a framework for multimodal interaction is outlined that can offer some guidance to integration. By exemplifying how a wide range of research prototypes fit into the framework today, it is shown how interfaces communicating through several modalities may enhance mobile interaction tomorrow.

We introduce the Cambridge Law Corpus (CLC), a corpus for legal AI research. It consists of over 250 000 court cases from the UK. Most cases are from the 21st century, but the corpus includes cases as old as the 16th century. This paper presents the first release of the corpus, containing the raw text and meta-data. Together with the corpus, we provide annotations on case outcomes for 638 cases, done by legal experts. Using our annotated data, we have trained and evaluated case outcome extraction with GPT-3, GPT-4 and RoBERTa models to provide benchmarks. We include an extensive legal and ethical discussion to address the potentially sensitive nature of this material. As a consequence, the corpus will only be released for research purposes under certain restrictions.

Current methods for word alignment require considerable amounts of parallel text to deliver accurate results, a requirement which is met only for a small minority of the world’s approximately 7,000 languages. We show that by jointly performing word alignment and annotation transfer in a novel Bayesian model, alignment accuracy can be improved for language pairs where annotations are available for only one of the languages—a finding which could facilitate the study and processing of a vast number of low-resource languages. We also present an evaluation where our method is used to perform single-source and multi-source part-of-speech transfer with 22 translations of the same text in four different languages. This allows us to quantify the considerable variation in accuracy depending on the specific source text(s) used, even with different translations into the same language.

This study discusses the structure of Swedish compounds within the framework of Construction Grammar, and applies the result to Word Sense Disambiguation of compound components. A construction-based approach is shown to achieve significantly better results than a set of baselines.

In this thesis I explore Bayesian models for word alignment, how they can be improved through joint annotation transfer, and how they can be extended to parallel texts in more than two languages. In addition to these general methodological developments, I apply the algorithms to problems from sign language research and linguistic typology.

In the first part of the thesis, I show how Bayesian alignment models estimated with Gibbs sampling are more accurate than previous methods for a range of different languages, particularly for languages with few digital resources available—which is unfortunately the state of the vast majority of languages today. Furthermore, I explore how different variations to the models and learning algorithms affect alignment accuracy.

Then, I show how part-of-speech annotation transfer can be performed jointly with word alignment to improve word alignment accuracy. I apply these models to help annotate the Swedish Sign Language Corpus (SSLC) with part-of-speech tags, and to investigate patterns of polysemy across the languages of the world.

Finally, I present a model for multilingual word alignment which learns an intermediate representation of the text. This model is then used with a massively parallel corpus containing translations of the New Testament, to explore word order features in 1001 languages.