Text-to-speech voices have come a long way in terms of their naturalness, and they are getting closer to human-sounding than ever. However, among the problems that still persist, the pronunciation of foreign words is still one of them. The experiments conducted in this thesis focus on using grapheme-to-phoneme (G2P) models to tackle the just-mentioned issue and, more specifically, to adjust the erroneous pronunciation of English words to an Italian English accent in Italian-speaking voices. We curated a dataset of words collected during recording sessions with an Italian voice actor reading general conversational sentences. We then manually transcribed their pronunciation in Italian English. In the second stage, we augmented the dataset by collecting the most common surnames in Great Britain and the United States, phonetically transcribed them with a rule-based phoneme mapping algorithm previously deployed by the company, and then manually adjusted the pronunciations to Italian English. Thirdly, by using the massively multilingual ByT5 model, a Transformer G2P model pre-trained on 100 languages, as well as its tokenizer-dependent versions T5_base and T5_small, and an LSTM with attention based on OpenNMT, we performed 10-fold cross-validation with the curated dataset. The results show that augmenting the data benefitted every model. In terms of PER, WER and accuracy, the transformer-based ByT5_small strongly outperformed its T5_small and T5_base counterparts even with a third or two-thirds of the training data. The second best performing model, the LSTM with attention one built with the OpenNMT framework, outperformed as well the T5 models, showed the second-best accuracy of our experiments and was the 'lightest' in terms of trainable parameters (2M) in comparison to ByT5 (299M) and the T5 ones (60 and 200M).

Passwords are common figures in everyone’s everyday life. One person has in average80 accounts for which they are supposed to use different passwords. Remembering allthese passwords is difficult and leads to people reusing, or reusing with slight modification,passwords on many accounts. Studies on memory show that information relating tosomething personal is more easily remembered. This is likely the reason as to why manypeople use passwords relating to either self, relatives, lovers, friends, or pets. Hackers will most often use either brute force or dictionary attacks to crack a password.These techniques can be quite time consuming so using machine learning could bea faster and easier approach. Segmenting someone’s previous passwords into meaningfulunits often reveals personal information about the creator and can thus be used as a basisfor password guessing. This report focuses on evaluating different sizes of the GPT-SW3model, which uses a transformer architecture, on password segmentation. The purposeis to find out if the GPT-SW3 model is suitable to use as a password segmenter and byextension if it can be used for password guessing. As training data, a list of passwords collected from a security breach on a platformcalled RockYou was used. The passwords were segmented by the author to provide themodel with a correct answer to learn from. The evaluation metric, Exact Match, checksif the model’s prediction is the same as that of the author. There were no positive resultswhen training GPT-SW3, most likely because of technical limitations. As the results arerather insufficient, future studies are required to prove or disprove the assumptions thisthesis is based on.

Gender biases are present in many NLP models. Such biased models can have large negative consequences on individuals. This work is a case study where we attempt to reduce them in a corpus consisting of Wikipedia articles about persons, in order to reduce them in models that would be trained on the corpus. For this, we apply two methods of modifying the corpus’s documents. Both methods replace gendered words (such as ‘mother’, ‘father’ and ‘parent’) with each other to change the contexts in which they each appear. The analysis and comparison of those two methods show that one of them is indeed suited to reduce gender biases. By modifying 35% of the corpus’s documents, the context of gendered words seems equal between the three considered genders (feminine, masculine and non-binary). This is confirmed through the performance of coreference resolution models trained with word embeddings fine-tuned on the corpus before and after modifying it. Evaluating these models on schemas specifically designed to point out gender biases in coreference resolution models shows that the model using the modified corpus is indeed less gender biased than the original. Our analysis further shows that the method does not compromise the corpus’s overall quality.

In this thesis we retrieve semantic information that exists implicitly in bilingual data. We gather input data by repeatedly applying the semantic mirroring procedure. The data is then represented by vectors in a large vector space. A resource of synonym clusters is then constructed by performing K-means centroid-based clustering on the vectors. We evaluate the result manually, using dictionaries, and against WordNet, and discuss prospects and applications of this method. / I det här arbetet utvinner vi semantisk information som existerar implicit i tvåspråkig data. Vi samlar indata genom att upprepa proceduren semantisk spegling. Datan representeras som vektorer i en stor vektorrymd. Vi bygger sedan en resurs med synonymkluster genom att applicera K-means-algoritmen på vektorerna. Vi granskar resultatet för hand med hjälp av ordböcker, och mot WordNet, och diskuterar möjligheter och tillämpningar för metoden.

Internet har sedan starten öppnat nya portar för kommunikation. En av de allra populäraste just nu är att blogga. Att uttrycka sig språkligt har kommit att bli så mycket mer än bara att använda sig av ord. På bloggen ges möjlighet att tillföra bild, film, färg och att använda olika typografiska medel, såsom att kursivera eller göra text fetstilt. Element som alla bidrar till hur text tolkas. Utifrån fjorton dagboksbloggar och totalt 289 blogginlägg har min uppsats syftat till att undersöka hur framställning på dessa bloggar, tillhörande gymnasieelever, skett. Mina frågeställningar jag utgått ifrån lyder: <ul type="disc">Hur använder sig gymnasieungdomar av olika uttrycksmedel för att estetiskt och kreativt skapa ett blogginlägg på så kallade dagboksbloggar? -          Hur används rubriksättning, bild, film, färg och olika stilformat på texten för att skapa kommunikation och olika uttryck på blogginläggen? <ul type="disc">Hur förhåller sig gymnasieungdomars dagboksblogg till den traditionella dagboken vad det gäller utformning och kommunikationsmöjligheter? Genom en strukturalistisk analys, med utgångspunkt hos Jurij Lotman, har jag gripit mig an blogginläggen på olika plan där jag både undersökt detaljer i texten och övergripande utformning. Jag har funnit att dagboksbloggen och dagboken skiljer sig på flera plan. Främst i fråga om kommunikationen som sker öppet på dagboksbloggen. Språkligt utmärker sig bloggen främst genom att ord och meningar betonas genom fetstilt och kursiv text, både för att göra texten mer lättövergriplig men också för att betona uttryck. Smileys och andra känslouttryck visar i sin tur hur ungdomarna undviker missförstånd på ett sätt som inte kräver bearbetning av texten. Jag vill säga att uppsatsen visar på hur en vidgad syn på språklighet och kommunikation idag är nödvändig, i och med de nya medel som tillkommit i dagens IT-samhälle. / Internet has since the beginning widened the form of communication. In recent times one of the most popular form is via blogs. To express yourself has become more than words. The blogs give you the ability to add pictures, videos, colors and more. You are also able to use typological medium like italic and bold types. All these elements contribute to how the text is read and interpreted. From 14 different diary blogs written by high school students and 289 posts in total my thesis intend to study which method of fabrications these blogs use. The question formulations I have based my thesis on are: ·         How do high school students use different ways of expressions to esthetical and creatively create posts at the so called diary blogs? -          How does headlining, pictures, film, colour and different typological medium being used to create communication and different expression on the posts? ·         How does the diary blog relate to the traditional diary regarding the formation and forms of communication? Through a structuralistic analysis method based on Jurij Lotman’s analysis I’ve approached the posts on different levels, where I examine details in the text but also the structure. I’ve found that the diary blog and the diary separate from each other on several plans, foremost the way of communication which is overt in a diary blog. Linguistically the diary blog distinguish itself from diaries by the way to be able to emphasize words or a sentence with italic and bold types. Smileys and different kinds of emotional forms of expressions are used by the blogger to avoid misconceptions. The thesis has proven that a widening way of looking at linguistic and communications are necessary due to the new medium that comes with the IT.

The demand for automation of simple tasks is constantly increasing. While some tasks are easy to automate because the logic is fixed and the process is streamlined, other tasks are harder because the performance of the task is heavily reliant on the judgment of a human expert. Matching a consultant to an offer from a client is one such task, in which case the expert is either a manager to the consultants or someone within HR at the company. One way to approach this task is to model the specific domain of interest using natural language processing. If we can capture the relationships between relevant skills and phrases within the specific domain, we could potentially use the resulting embeddings in a consultant to offer matching scheme. In this paper, we propose a key phrase-based web scraping approach to collect the data we need for a domain-specific corpus. To retrieve the key phrases needed as prompts for web scraping, we propose using the transformer-based library KeyBERT on limited domain-specific in house data belonging to the consultant firm B3 Indes, in order to retrieve the most important phrases in their respective contexts. Facebook's Word2vec based language model fasttext is then used on the processed corpus to create the fixed word embeddings. We also investigate numerous different approaches for selecting the right key phrases for web scraping in a human similarity comparison scheme, as well as comparisons to a larger pretrained general domain fasttext model. We show that utilizing key phrases for a domain-specific fasttext model could be beneficial compared to using a larger pretrained model. The results are not consistently conclusive under the current analytical framework. The results also indicate that KeyBERT is beneficial when selecting the key phrases compared to the randomized sampling of relevant phrases; however, the results are not conclusive. / Efterfrågan för automatisering av enkla uppgifter efterfrågas alltmer. Medan vissa uppgifter är lätta att automatisera eftersom logiken är fast och processen är tydlig, är andra svårare eftersom utförandet av uppgiften starkt beror på en människas expertis. Att matcha en konsult till ett erbjudande från en klient är en sådan uppgift, där experten är antingen en chef för konsulterna eller någon inom HR på företaget. En metod för att hantera denna uppgift är att modellera det specifika området av intresse med hjälp av maskininlärningsbaserad språkteknologi. Om vi kan fånga relationerna mellan relevanta färdigheter och fraser inom det specifika området, skulle vi potentiellt kunna använda de resulterande inbäddningarna i ett matchningsprocess mellan konsulter och uppdrag. I denna rapport föreslås en nyckelordsbaserad webbskrapnings-metod för att samla in data som behövs för ett domänspecifikt korpus. För att hämta de nyckelord som behövs som input för webbskrapning, föreslår vi att använda transformator-baserade biblioteket KeyBERT på begränsad domänspecifik data från konsultbolaget B3 Indes, detta för att hämta de viktigaste fraserna i deras respektive sammanhang. Sedan används Facebooks Word2vec baserade språkmodell fasttext på det bearbetade korpuset för att skapa statiska inbäddningar. Vi undersöker också olika metoder för att välja rätt nyckelord för webbskrapning i en likhets-jämnförelse mot mänskliga experter, samt jämförelser med en större förtränad fasttext-modell som inte är domänspecifik. Vi visar att användning av nyckelord för webbskrapning för träning av en domänspecifik fasttext-modell skulle kunna vara fördelaktigt jämnfört med en förtränad modell, men resutaten är inte konsekvent signifikanta enligt det begränsade analytiska ramverket. Resultaten indikerar också att KeyBERT är fördelaktigt vid valet av nyckelord jämfört med slumpmässigt urval av relevanta fraser, men dessa resultat är inte heller helt entydiga.

With the rapid growth of e-commerce, enabling effective product recommendation systems and improving product search for shoppers plays a crucial role in driving customer satisfaction. Traditional product retrieval approaches have mainly relied on unimodal models focusing on text data. However, to capture auxiliary context and improve the accuracy of similarity predictions, it is crucial to explore architectures that can leverage additional sources of information, such as images. This thesis compares the performance of multi- and unimodal methods for product similarity prediction and product retrieval. Both approaches are applied to two e-commerce datasets, one containing English and another containing Swedish product descriptions. A pre-trained multi-modal model called CLIP is used as a feature extractor. Different models are trained on CLIP embeddings using either text-only, image-only or image-text inputs. An extension of triplet loss with margins is tested, along with various training setups. Given the lack of similarity labels between products, product similarity prediction is studied by measuring the performance of a K-Nearest Neighbour classifier implemented on features extracted by the trained models. The thesis results demonstrate that multi-modal architectures outperform unimodal models in predicting product similarity. The same is true for product retrieval. Combining textual and visual information seems to lead to more accurate predictions than models relying on only one modality. The findings of this research have considerable implications for e-commerce platforms and recommendation systems, providing insights into the effectiveness of multi-modal models for product-related tasks. Overall, the study contributes to the existing body of knowledge by highlighting the advantages of leveraging multiple sources of information for deep learning. It also presents recommendations for designing and implementing effective multi-modal architectures. / I och med den snabba tillväxten av e-handel spelar att möjliggöra effektivare produktrekommendationssystem och att förbättra produktsök för konsumenter en viktig roll för att öka kundnöjdheten. Traditionella angreppsätt för produktsök har huvudsakligen tillförlitat sig på unimodala textmodeller. För att fånga ett bredare kontext och förbättra exaktheten av prediktioner av likhet mellan produkter är det viktigt att utforska arkitekturer som kan utnyttja fler informationskällor så som bilder. Den här avhandlingen jämför prestanda hos multimodala och unimodala metoder för produktlikhetsprediktioner och produktsök. Båda angreppsätten är tillämpade på två e-handelsdatamängder, en med engelska produktbeskrivningar och en med svenska. En förtränad multimodal modell kallad CLIP används för att skapa produktrepresentationer. Olika modeller har tränats på CLIPs representationer, antingen med enbart text, enbart bild eller både bild och text. En utökning av ett triplettmått med marginaler har testats som träningskriterium, i kombination med olika träningsinställningar. Givet en avsaknad av likhetsannoteringar mellan produkter så har produktlikhetsprediktion studerats genom att mäta prestandan av K-närmaste-grannar-klassificering genom att använda vektor-representationer från de tränade modellerna. Avhandlingens resultat visar att multimodala arkitekturer överträffar unimodala modeller för produktlikhetsprediktion. Att kombinera textuell och visuell information verkar leda till mer korrekta prediktioner jämfört med modeller som förlitar sig på endast en modalitet. Forskningsresultaten har markanta implikationer för e-handelsplattformar och rekommendationssystem, genom att tillhandahålla insikter i multimodala modellers effektivitet i produktrelaterade uppgifter. Överlag så bidrar studien till den existerande litteraturen genom att förtydliga fördelarna av att utnyttja flera informationskällor för djupinlärning. Den resulterar också i rekommendationer för att designa och implementera effektiva multimodala modellarkitekturer.

When clustering together synonyms, complications arise in cases of the words having multiple senses as each sense’s synonyms are erroneously clustered together. The task of automatically distinguishing word senses in cases of ambiguity, known as word sense disambiguation (WSD), has been an extensively researched problem over the years. This thesis studies the possibility of applying an unsupervised machine learning based WSD-system for analysing existing synonym clusters (N = 149) and dividing them correctly when two or more senses are present. Based on sense embeddings induced from a large corpus, cosine similarities are calculated between sense embeddings for words in the clusters, making it possible to suggest divisions in cases where different words are closer to different senses of a proposed ambiguous word. The system output is then evaluated by four participants, all experts in the area. The results show that the system does not manage to correctly divide the clusters in more than 31% of the cases according to the participants. Moreover, it is discovered that some differences exist between the participants’ ratings, although none of the participants predominantly agree with the system’s division of the clusters. Evidently, further research and improvements are needed and suggested for the future.

In this thesis, we work towards bridging the gap between two distinct areas: noisy text handling and text summarisation. The overall goal of the paper is to examine the effects of noise in the training of convolutional neural networks for text summarisation, with a view to understanding how to effectively create a noise-robust text-summarisation system. We look specifically at the problem of abstractive text summarisation of noisy data in the context of summarising error-containing documents from automatic speech recognition (ASR) output. We experiment with adding varying levels of noise (errors) to the 4 million-article Gigaword corpus and training an encoder-decoder CNN on it with the aim of producing a noise-robust text summarisation system. A total of six text summarisation models are trained, each with a different level of noise. We discover that the models with a high level of noise are indeed able to aptly summarise noisy data into clean summaries, despite a tendency for all models to overfit to the level of noise on which they were trained. Directions are given for future steps in order to create an even more noise-robust and flexible text summarisation system.

We built a system to participate in the eRisk 2019 T1 Shared Task. The aim of the task was to evaluate systems for early risk prediction on the internet, in particular to identify users suffering from eating disorders as accurately andquickly as possible given their history of Reddit posts in chronological order. In the controlled settings of this task, we also evaluated the performance of three different word representation methods: random indexing, GloVe, and ELMo.We discuss our system’s performance, also in the light of the scores obtained by other teams in the shared task. Our results show that our two-step learning approach was quite successful, and we obtained good scores on the early risk prediction metric ERDE across the board. Contrary to our expectations, we did not observe a clear-cut advantage of contextualized ELMo vectors over the commonly used and much more light-weight GloVevectors. Our best model in terms of F1 score turned out to be a model with GloVe vectors as input to the text classifier and a multi-layer perceptron as user classifier. The best ERDE scores were obtained by the model with ELMo vectors and a multi-layer perceptron. The model with random indexing vectors hit a good balance between precision and recall in the early processing stages but was eventually surpassed by the models with GloVe and ELMo vectors. We put forward some possible explanations for the observed results, as well as proposing some improvements to our system.