Recent advances in large neural network-style models have demonstrated great performance in various applications, such as image generation, question answering, and audio classification. However, these deep and high-capacity models require a large amount of labeled data to function properly, rendering them inapplicable in many real-world scenarios. This dissertation focuses on the development and evaluation of advanced machine learning algorithms to solve the following research questions: (1) How to learn novel classes with limited labeled data, (2) How to adapt a large pre-trained model to the target domain if only unlabeled data is available, (3) How to boost the performance of the few-shot learning model with unlabeled data, and (4) How to utilize limited labeled data to learn new classes without the training data in the same domain. First, we study few-shot learning in text classification tasks. Meta-learning is becoming a popular approach for addressing few-shot text classification and has achieved state-of-the-art performance. However, the performance of existing approaches heavily depends on the interclass variance of the support set. To address this problem, we propose a TART network for few-shot text classification. The model enhances the generalization by transforming the class prototypes to per-class fixed reference points in task-adaptive metric spaces. In addition, we design a novel discriminative reference regularization to maximize divergence between transformed prototypes in task-adaptive metric spaces to improve performance further. In the second problem we focus on self-learning in cross-lingual transfer task. Our goal here is to develop a framework that can make the pretrained cross-lingual model continue learning the knowledge with large amount of unlabeled data. Existing self-learning methods in crosslingual transfer tasks suffer from the large number of incorrectly pseudo-labeled samples used in the training phase. We first design an uncertainty-aware cross-lingual transfer framework with pseudo-partial-labels. We also propose a novel pseudo-partial-label estimation method that considers prediction confidences and the limitation to the number of candidate classes. Next, to boost the performance of the few-shot learning model with unlabeled data, we propose a semi-supervised approach for few-shot semantic segmentation task. Existing solutions for few-shot semantic segmentation cannot easily be applied to utilize image-level weak annotations. We propose a class-prototype augmentation method to enrich the prototype representation by utilizing a few image-level annotations, achieving superior performance in one-/multi-way and weak annotation settings. We also design a robust strategy with softmasked average pooling to handle the noise in image-level annotations, which considers the prediction uncertainty and employs the task-specific threshold to mask the distraction. Finally, we study the cross-domain few-shot learning in the semantic segmentation task. Most existing few-shot segmentation methods consider a setting where base classes are drawn from the same domain as the new classes. Nevertheless, gathering enough training data for meta-learning is either unattainable or impractical in many applications. We extend few-shot semantic segmentation to a new task, called Cross-Domain Few-Shot Semantic Segmentation (CD-FSS), which aims to generalize the meta-knowledge from domains with sufficient training labels to low-resource domains. Then, we establish a new benchmark for the CD-FSS task and evaluate both representative few-shot segmentation methods and transfer learning based methods on the proposed benchmark. We then propose a novel Pyramid-AnchorTransformation based few-shot segmentation network (PATNet), in which domain-specific features are transformed into domain-agnostic ones for downstream segmentation modules to fast adapt to unseen domains. / Doctor of Philosophy / Nowadays, deep learning techniques play a crucial role in our everyday existence. In addition, they are crucial to the success of many e-commerce and local businesses for enhancing data analytics and decision-making. Notable applications include intelligent transportation, intelligent healthcare, the generation of natural language, and intrusion detection, among others. To achieve reasonable performance on a new task, these deep and high-capacity models require thousands of labeled examples, which increases the data collection effort and computation costs associated with training a model. Moreover, in many disciplines, it might be difficult or even impossible to obtain data due to concerns such as privacy and safety. This dissertation focuses on learning with limited labeled data in natural language processing and computer vision tasks. To recognize novel classes with a few examples in text classification tasks, we develop a deep learning-based model that can capture both cross- task transferable knowledge and task-specific features. We also build an uncertainty-aware self-learning framework and a semi-supervised few-shot learning method, which allow us to boost the pre-trained model with easily accessible unlabeled data. In addition, we propose a cross-domain few-shot semantic segmentation method to generalize the model to different domains with a few examples. By handling these unique challenges in learning with limited labeled data and developing suitable approaches, we hope to improve the efficiency and generalization of deep learning methods in the real world.

Andromeda is an interactive visualization tool that projects high-dimensional data into a scatterplot-like visualization using Weighted Multidimensional Scaling (WMDS). The visualization can be explored through surface-level interaction (viewing data values), parametric interaction (altering underlying parameterizations), and observation-level interaction (directly interacting with projected points). This thesis presents analyses on the collaborative utility of Andromeda in a middle school class and the insights college-level students generate when using Andromeda. The first study discusses how a middle school class collaboratively used Andromeda to explore and compare their engineering designs. The students analyzed their designs, represented as high-dimensional data, as a class. This study shows promise for introducing collaborative data analysis to middle school students in conjunction with other technical concepts such as the engineering design process. Participants in the study on college-level students were given a version of Andromeda, with access to different interactions, and were asked to generate insights on a dataset. By applying a novel visualization evaluation methodology on students' natural language insights, the results of this study indicate that students use different vocabulary supported by the interactions available to them, but not equally. The implications, as well as limitations, of these two studies are further discussed. / Master of Science / Data is often high-dimensional. A good example of this is a spreadsheet with many columns. Visualizing high-dimensional data is a difficult task because it must capture all information in 2 or 3 dimensions. Andromeda is a tool that can project high-dimensional data into a scatterplot-like visualization. Data points that are considered similar are plotted near each other and vice versa. Users can alter how important certain parts of the data are to the plotting algorithm as well as move points directly to update the display based on the user-specified layout. These interactions within Andromeda allow data analysts to explore high-dimensional data based on their personal sensemaking processes. As high dimensional thinking and exploratory data analysis are being introduced into more classrooms, it is important to understand the ways in which students analyze high-dimensional data. To address this, this thesis presents two studies. The first study discusses how a middle school class used Andromeda for their engineering design assignments. The results indicate that using Andromeda in a collaborative way enriched the students' learning experience. The second study analyzes how college-level students, when given access to different interaction types in Andromeda, generate insights into a dataset. Students use different vocabulary supported by the interactions available to them, but not equally. The implications, as well as limitations, of these two studies are further discussed.

Master of Science / Electronic theses and dissertations (ETDs) are structured documents in which chapters are major components. There is a lack of any repository that contains chapter boundary details alongside these structured documents. Revealing these details of the documents can help increase accessibility. This research explores the manipulation of ETDs marked up using LaTeX to generate chapter boundaries. We use this to create a data set of 1,459 ETDs and their chapter boundaries. Additionally, for the task of automatic segmentation of unseen documents, we prototype three deep learning models that are trained using this data set. We hope to encourage researchers to incorporate LaTeX manipulation techniques to create similar data sets.

Coherence is the quality that gives a text its conceptual unity, making a text a coordinated set of connected parts rather than a random group of sentences (turns, in the case of dialogue). Hence, coherence is an integral property of human communication, necessary for a meaningful discourse both in text and dialogue. As such, coherence can be regarded as a requirement for conversational agents, i.e. machines designed to converse with humans. Though recently there has been a proliferation in the usage and popularity of conversational agents, dialogue coherence is still a relatively neglected area of research, and coherence across multiple turns of a dialogue remains an open challenge for current conversational AI research. As conversational agents progress from being able to handle a single application domain to multiple ones through any domain (open-domain), the range of possible dialogue paths increases, and thus the problem of maintaining multi-turn coherence becomes especially critical. In this thesis, we investigate two aspects of coherence in dialogue and how they can be used to design modules for an open-domain coherent conversational agent. In particular, our approach focuses on modeling intentional and thematic information patterns of distribution as proxies for a coherent discourse in open-domain dialogue. While for modeling intentional information we employ Dialogue Acts (DA) theory (Bunt, 2009); for modeling thematic information we rely on open-domain entities (Barzilay and Lapata, 2008). We find that DAs and entities play a fundamental role in modelling dialogue coherence both independently and jointly, and that they can be used to model different components of an open-domain conversational agent architecture, such as Spoken Language Understanding, Dialogue Management, Natural Language Generation, and open-domain dialogue evaluation. The main contributions of this thesis are: (I) we present an open-domain modular conversational agent architecture based on entity and DA structures designed for coherence and engagement; (II) we propose a methodology for training an open-domain DA tagger compliant with the ISO 24617-2 standard (Bunt et al., 2012) combining multiple resources; (III) we propose different models, and a corpus, for predicting open-domain dialogue coherence using DA and entity information trained with weakly supervised techniques, first at the conversation level and then at the turn level; (IV) we present supervised approaches for automatic evaluation of open-domain conversation exploiting DA and entity information, both at the conversation level and at the turn level; (V) we present experiments with Natural Language Generation models that generate text from Meaning Representation structures composed of DAs and slots for an open-domain setting.

This study replicated a common experiment, the Desert Survival Problem, and attempted to add data to the body of knowledge for deception cues. Participants wrote truthful and deceptive essays arguing why items salvaged from the wreckage were useful for survival. Cues to deception considered here fit into four categories: those caused by a deceivers' negative emotion, verbal immediacy, those linked to a deceiver's attempt to appear truthful, and those resulting from deceivers' high cognitive load. Cues caused by a deceiver's negative emotions were mostly absent in the results, although deceivers did use fewer first-person pronouns than truth tellers. That indicated deceivers were less willing to take ownership of their statements. Cues because of deceivers' attempts to appear truthful were present. Deceivers used more words and more exact language than truth tellers. That showed an attempt to appear truthful. Deceivers' language was simpler than that of truth tellers, which indicated a higher cognitive load. Future research should include manipulation checks on motivation and emotion, which are tied to cue display. The type of cue displayed, be it emotional leakage, verbal immediacy, attempts to appear truthful or cognitive load, might be associated with particular deception tasks. Future research, including meta-analyses, should attempt to determine which deception tasks produce which cue type. Revised file, GMc 5/28/2014 per Dean DePauw / Master of Arts

There are a lot of software systems currently in use for different applications. To make sure that these systems function there is a need to properly test and maintain them.When a system grows in scope it becomes more difficult to test and maintain, and so test selection and prioritization tools that incorporate artificial intelligence, information retrieval and natural language processing are useful. In this thesis, different information retrieval models were implemented and evaluated using multiple datasets based on different filters and pre-processing methods. The data was provided by Westermo Network Technologies AB and represent one of their systems. The datasets contained data with information about the test results and what data was used for the tests.  The results showed that for models that are not trained for this data it is more beneficial to give them less data which is only related to test failures. Allowing the models to have access to more data showed that they made connections that were inaccurate as the data were unrelated. The results also showed that if a model is not adjusted to the data, a simple model could be more effective compared to a more advanced model. / Det finns många mjukvarusystem som för närvarande används för olika tjänster. För att säkerställa att dessa system fungerar korrekt är det nödvändigt att testa och underhålla dem ordentligt.När ett system växer i omfattning blir det svårare att testa och underhålla, och testverktyg för testselektion och testprioritering som integrerar artificiell intelligens, informationssökning och natural language processing är därför användbara. I denna rapport utvärderades olika informationssökningsmodeller med hjälp av olika dataset som är baserade på olika filter och förbehandlingsmetoder. Datan tillhandahölls av Westermo Network Technologies AB och representerar ett av deras system. Dataseten innehöll data med information om testresultaten och vilken data som användes för testen. Resultaten visade att för modeller som inte är tränade för denna data är det mer fördelaktigt att ge dem mindre data som endast är relaterade till testfel. Att ge modellerna tillgång till mer data visade att de gjorde felaktiga kopplingar eftersom data var orelaterad. Resultaten visade också att givet en modell inte var justerad mot data, kunde en enklare modell vara mer effektiv än en mer avancerad modell.

<p>The fact that every individual has a different sense of humor and it varies greatly from one person to another means that it is a challenge to learn any individual’s humor preferences. Humor is much more than just a source of entertainment; it is an essential tool that aids communication. Understanding humor preferences can lead to improved social interactions and bridge existing social or economic gaps.</p><p> </p><p>In this study, we propose a methodology that aims to develop a recommendation system for jokes by analyzing its text. Various researchers have proposed different theories of humor depending on their area of focus. This exploratory study focuses mainly on Attardo and Raskin’s (1991) General Theory of Verbal Humor and implements the knowledge resources defined by it to annotate the jokes. These annotations contain the characteristics of the jokes and also play an important role in determining how alike these jokes are. We use Lin’s similarity metric (Lin, 1998) to computationally capture this similarity. The jokes are clustered in a hierarchical fashion based on their similarity values used for the recommendation. We also compare our joke recommendations to those obtained by the Eigenstate algorithm (Goldberg, Roeder, Gupta, & Perkins, 2001), an existing joke recommendation system that does not consider the content of the joke in its recommendation.</p>

<p>Empowering machines to understand our physical world should go beyond models with only natural language and models with only vision. Vision and language is a growing field of study that attempts to bridge the gap between natural language processing and computer vision communities by enabling models to learn visually grounded language. However, as an increasing number of pre-trained visual linguistic models focus on the alignment between visual regions and natural language, it is difficult to claim that these models capture certain properties of objects in their latent space, such as size. Inspired by recent trends in prompt learning, this study will design a prompt learning framework for two visual linguistic models, ViLBERT and ViLT, and use different manually crafted prompt templates to evaluate the consistency of performance of these models in comparing the size of objects. The results of this study showed that ViLT is more consistent in prediction accuracy for the given task with six pairs of objects under four prompt designs. However, the overall prediction accuracy is lower than the expectation on this object size comparison task; even the better model in this study, ViLT, has only 16 out of 24 cases better than the proposed random chance baseline. As this study is a preliminary study to explore the potential of pre-trained visual linguistic models on object size comparison, there are many directions for future work, such as investigating more models, choosing more object pairs, and trying different methods for feature engineering and prompt engineering.</p>

Toxiskt språk på internet och det som ofta i vardagliga termer benämns som näthat innefattar kränkningar, hot och stötande språk. Toxiskt språk är särskilt märkbart på sociala medier. Det går att upptäcka toxiskt språk på internet med hjälp av maskininlärning som automatiskt känner igen typiska särdrag för toxiskt språk. Tidigare svensk forskning har undersökt förekomsten av toxiskt språk på sociala medier med hjälp av maskininlärning, men det saknas fortfarande forskning på den allt mer populära plattformen TikTok. Syftet med denna studie är att undersöka förekomsten och särdragen av toxiska kommentarer på TikTok med hjälp av maskininlärning och manuella metoder. Studien är menad att ge en bättre förståelse för vad unga möts av i kommentarerna på TikTok. Studien applicerar en mixad metod i en dokumentundersökning av 69 895 kommentarer. Maskininlärningsmodellen Hatescan användes för att automatiskt klassificera sannolikheten att toxiskt språk förekommer i kommentarerna. Utifrån denna sannolikhet analyserades ett urval av kommentarerna manuellt, vilket ledde till både kvantitativa och kvalitativa fynd. Resultatet av studien visade att omfattningen av toxiskt språk var relativt liten, där 0,24% av 69 895 kommentarer ansågs vara toxiska enligt en både automatiserad och manuell bedömning. Den typ av toxiskt språk som mest förekom i undersökningen visades vara obscent språk, som till majoriteten innehöll svordomar. / Toxic language on the internet and what is often referred to in everyday terms as cyberbullying includes insults, threats and offensive language. Toxic language is particularly noticeable on social media. It is possible to detect toxic language on the internet with the help of machine learning in the form of, among other things, Natural Language Processing (NLP) techniques, which automatically recognize typical characteristics of toxic language. Previous Swedish research has investigated the presence of toxic language on social media using machine learning, but there is still a lack of research on the increasingly popular platform TikTok. Through the study, the authors intend to investigate the prevalence and characteristics of toxic comments on TikTok using both a machine learning technique and manual methods. The study is meant to provide a better understanding of what young people encounter in the comments on TikTok. The study applies a mixed method in a document survey of 69 895 comments. Hatescan was used to automatically classify the likelihood of toxic language appearing in the comments. Based on this probability, a section of the comments could be sampled and manually analysed using theory, leading to both quantitative and qualitative findings. The results of the study showed that the prevalence of toxic language was relatively small, with 0.24% of 69 895 comments considered toxic based on an automatic and manual analysis. The type of toxic language that occurred the most in the study was shown to be obscene language, the majority of which contained swear words.

<p>Danas, servisi e-Uprave u različitim oblastima koriste question answer sisteme koncepta u poku&scaron;aju da se razume tekst i da pomognu građanima u dobijanju odgovora na svoje upite u bilo koje vreme i veoma brzo. Automatsko mapiranje relevantnih dokumenata se ističe kao važna aplikacija za automatsku strategiju klasifikacije: upit-dokumenta. Ova doktorska disertacija ima za cilj doprinos u identifikaciji nestruktuiranih dokumenata i predstavlja važan korak ka razja&scaron;njavanju uloge eksplicitnih koncepata u pronalaženju podataka uop&scaron;te ajče&scaron; a reprezenta vna &scaron;ema u tekstualnoj kategorizaciji je BoW pristup, kada je u pozadini veliki skup znanja. Ova disertacija uvodi novi pristup ka stvaranju koncepta zasnovanog na tekstualnoj prezantaciji i primeni kategorizacije teksta, kako bi se stvorile definisane klase u slučaju sažetih tekstualnih dokumenata Takođe, ovde je prikazan algoritam zasnovan na klasifikaciji, modelovan za upite koji odgovaraju temi. Otežavaju a okolnost u slučaju ovog koncepta, koji prezentuje termine sa visokom frekvencijom pojavljivanja u upitma, zasniva se na sličnostima u prethodno definisanim klasama dokumenata Rezultati eksperimenta iz oblasti Krivičnog zakonika Republike Srbije, u ovom slučaju i studija, pokazuju da prezentacija teksta zasnovana na konceptu ima zadovoljavaju e rezultate i u slučaju kada ne postoji rečnik za datu oblast.</p> / <p>Nowadays, the concept of Question Answering Systems (QAS) has been used by e-government services in various fields as an attempt to understand the text and help citizens in getting answers to their questions promptly and at any time. Automatic mapping of relevant documents stands out as an important application for automatic classification strategy: query-document. This doctoral thesis aims to contribute to identification of unstructured documents and represents an important step towards clarifying the role of explicit concepts within Information Retrieval in general. The most common scheme in text categorization is BoW approach, especially when, as a basis, we have a large set of knowledge. This thesis introduces a new approach to the creation of text presentation based concept and applying text categorization, with the aim to create a defined class in case of compressed text documents.Also, this paper discusses the classification based algorithm modeled for queries that suit the theme. What makes the situation more complicated is the fact that this concept is based on the similarities in previously defined classes of documents and terms with a high frequency of appearance presented in queries. The results of the experiment in the field of the Criminal Code, and this paper as well, show that the text presentation based concept has satisfactory results even in case where there is no vocabulary for certain field.</p>