Att utveckla en sofistikerad chatbotlösning kräver stora mängder textdata för att kunna anpassalösningen till en specifik domän. Att manuellt skapa en komplett uppsättning textdata, specialanpassat för den givna domänen och innehållandes ett stort antal varierande meningar som en människa kan tänkas yttra, är ett enormt tidskrävande arbete. För att kringgå detta tillämpas dataaugmentering för att generera mer data utifrån en mindre uppsättning redan existerande textdata. Softronic AB vill undersöka alternativa strategier för dataaugmentering med målet att eventuellt ersätta den nuvarande lösningen med en mer vetenskapligt underbyggd sådan. I detta examensarbete har prototypmodeller utvecklats för att jämföra och utvärdera effekten av olika textaugmenteringsstrategier. Resultatet av genomförda experiment med prototypmodellerna visar att augmentering genom synonymutbyten med en domänanpassad synonymordlista, presenterade märkbart förbättrade effekter på förmågan hos en NLU-modell att korrekt klassificera data, gentemot övriga utvärderade strategier. Vidare indikerar resultatet att ett samband föreligger mellan den strukturella variationsgraden av det augmenterade datat och de tillämpade språkparens semantiska likhetsgrad under tillbakaöversättningar. / Developing a sophisticated chatbot solution requires large amounts of text data to be able to adapt the solution to a specific domain. Manually creating a complete set of text data, specially adapted for the given domain, and containing a large number of varying sentences that a human conceivably can express, is an exceptionally time-consuming task. To circumvent this, data augmentation is applied to generate more data based on a smaller set of already existing text data. Softronic AB wants to investigate alternative strategies for data augmentation with the aim of possibly replacing the current solution with a more scientifically substantiated one. In this thesis, prototype models have been developed to compare and evaluate the effect of different text augmentation strategies. The results of conducted experiments with the prototype models show that augmentation through synonym swaps with a domain-adapted thesaurus, presented noticeably improved effects on the ability of an NLU-model to correctly classify data, compared to other evaluated strategies. Furthermore, the result indicates that there is a relationship between the structural degree of variation of the augmented data and the applied language pair's semantic degree of similarity during back-translations.

Time estimation for software issues is crucial to planning projects. Developers and experts have for many decades tried to estimate time requirements for issues as accurately as possible. The methods that are used today are often time-consuming and complex. This thesis investigates if the time estimation process can be done with natural language processing and machine learning. Three different word embeddings were used to represent the free text description, bag-of-words with tf-idf weighing, word2Vec and fastText. The different word embeddings were then fed into two types of machine learning approaches, classification and regression. The classification was binary and can be formulated as will the issue take more than three hours?. The goal of the regression problem was to predict an actual value for the time that the issue would take to complete. The classification models performance were measured with an F1-score, and the regression model was measured with an R2-score. The best F1- score for classification was 0.748 and was achieved with the word2Vec word embedding and an SVM classifier. The best score for the regression analysis was achieved with the bag-of-words word embedding, which achieved an R2- score of 0.380. Further evaluation of the results and a comparison to actual estimates made by the company show that humans only performs slightly better than the models given the binary classification defined above. The F1-score of the employees was 0.792, a difference of just 0.044 from the best F1-score made by the models. This thesis concludes that the models are not good enough to use in a professional setting. An F1-score of 0.748 could be used in other settings, but the classification question in this problem is too broad to be used for a real project. The results for the regression is also too low to be of any valuable use. / Tidsuppskattning för programvaruärenden är en avgörande del för planering av projekt. Utvecklare och experter har i många årtionden försökt uppskatta tiden ett ärende kommer ta så exakt som möjligt. Metoderna som används idag är ofta tidskrävande och komplexa. Denna avhandling undersöker om tidsuppskattningsprocessen kan göras med hjälp av språkteknologi och maskininlärning. De flesta programvaruärenden har en fritextbeskrivning av vad som är fel eller behöver läggas till. Tre olika ordinbäddningar användes för att representera fritextbeskrivningen, bag-of-word med tf-idf-viktning, word2Vec och fastText. De olika ordinbäddningarna matades sedan in i två typer av maskininlärningsmetoder, klassificering och regression. Klassificeringen var binär och frågan kan formuleras som tar ärendet mer än tre timmar?. Målet med regressionsproblemet var att förutsäga ett faktiskt värde för den tid som frågan skulle ta att slutföra. Klassificeringsmodellens prestanda mättes med en F1-poäng och regressionsmodellen mättes med en R2-poäng. Den bästa F1-poängen för klassificering var 0.748 och uppnåddes med en word2Vec-ordinbäddning och en SVM-klassificeringsmodell. Den bästa poängen för regressionsanalysen uppnåddes med en bag-of-words-inbäddning, som uppnådde en R2-poäng på 0.380. Vidare undersökning av resultaten och en jämförelse av faktiskta tidsestimat som gjorts av företaget visar att människor bara är lite bättre än modellerna givet klassificeringsfrågan beskriven ovan. F1-poängen för de anställda var 0.792, bara 0.044 bättre än det bästa F1-poängen för modellerna. Slutsatsen för denna avhandling är att modellerna inte är tillräckligt bra för att användas i en professionell miljö. En F1-poäng på 0.748 kan användas i andra situationer, men klassificeringsfrågan i detta problem är för bred för att användas för ett riktigt projekt. Resultatet för regressionen är också för lågt för att vara till någon värdefull användning.

This thesis aims to evaluate different abstractive text summarization models and techniques for summarizing resumes. It has two main objectives: investigate the models’ performance on resume summarization and assess the suitability of the generated summaries as resume introductions. Although automatic abstractive text summarization has gained traction in various areas, its application in the resume domain has not yet been explored. Resumes present a unique challenge for abstractive summarization due to their diverse style, content, and length. To address these challenges, three state-of-the-art pre-trained text generation models: BART, T5, and ProphetNet, were selected. Additionally, two approaches that can handle longer resumes were investigated. The first approach, named LongBART, modified the BART architecture by incorporating the Longformer’s self-attention into the encoder. The second approach, named HybridBART, used an extractive-then-abstractive summarization strategy. The models were fine-tuned on a dataset of 653 resume-introduction pairs and were evaluated using automatic metrics as well as two types of human evaluations: a survey and expert interviews. None of the models demonstrated superiority across all criteria and evaluation metrics. However, the survey responses indicated that LongBART showed promising results, receiving the highest scores in three out of five criteria. On the other hand, ProphetNet consistently received the lowest scores across all criteria in the survey, and across all automatic metrics. Expert interviews emphasized that the generated summaries cannot be considered correct summaries due to the presence of hallucinated personal attributes. However, there is potential for using the generated texts as resume introductions, given that measures are taken to ensure the hallucinated personal attributes are sufficiently generic. / Denna avhandling utvärderar olika modeller och tekniker för automatisk textsammanfattning för sammanfattning av CV:n. Avhandlingen har två mål: att undersöka modellernas prestanda på sammanfattning av CV:n och bedöma lämpligheten att använda de genererade sammanfattningar som CV-introduktioner. Även om automatisk abstrakt textsummering har fått fotfäste inom olika sammanhang är dess tillämpning inom CV-domänen ännu outforskad. CV:n utgör en unik utmaning för abstrakt textsammanfattning på grund av deras varierande stil, innehåll och längd. För att hantera dessa utmaningar valdes tre av de främsta förtränade modellerna inom textgenerering: BART, T5 och ProphetNet. Dessutom undersöktes två extra metoder som kan hantera längre CV:n. Det första tillvägagångssättet, kallat LongBART, modifierade BART-arkitekturen genom att inkludera självuppmärksamhet från Longformer-arkitekturen i kodaren. Det andra tillvägagångssättet, kallat HybridBART, använde en extraktiv-sen-abstraktiv sammanfattningsstrategi. Modellerna finjusterades med ett dataset med 653 CV-introduktionspar och utvärderades med hjälp av automatiska mått, samt två typer av mänsklig utvärdering: en enkätundersökning och intervjuer med experter. Ingen av modellerna visade överlägsenhet på alla kriterier och utvärderingsmått. Dock indikerade enkätsvaren att LongBART visade lovande resultat, genom att få högst poäng i tre av fem utvärderingskategorier. Å andra sidan fick ProphetNet lägst poäng i samtliga utvärderingskategorier, samt lägst poäng i alla automatiska mätningar. Expertintervjuer framhävde att de genererade sammanfattningarna inte kan anses vara pålitliga som fristående sammanfattningar på grund av förekomsten av hallucinerade personliga egenskaper. Trots detta finns det potential att använda dessa sammanfattningar som introduktioner, under förutsättningen att åtgärder vidtas för att säkerställa att hallucinerade personliga attribut är tillräckligt generiska.

Plick is an online platform with the intention of being a marketplace where users may buy and sell second-hand fashion. The platform caters to younger users, and as such borrows many ideas from well-known social network platforms - such as putting more focus on user profiles and expression, rather than just the products themselves. One of these ideas is to allow users free reign over tagging their items, rather than having them select from some constrained, pre-approved, group of categories, styles, sizes - et cetera. A problem of letting users tag products however they see fit is that a subset of users will inevitably try to 'game' the system by knowingly tagging their products using incorrect labels - resulting in inaccurate search results for many of these incorrect tags.The aim of this project is to firstly develop a pre-processing algorithm to normalize the user generated tagging data - to handle situations such as a tag having multiple different (albeit possibly all correct) spellings, capitalizations, typos, languages etc. The processed data will then be used to develop two different approaches to solve the problem of incorrect tagging. The first approach involves using the normalized data to create a graph representation of the tags and their relations to each other. Each node in the graph will represent an individual tag, and each edge between nodes will explain how closely related those two tags are. An algorithm will then be developed to, utilizing the tag relation graph, describe the relatedness of an arbitrary group of tags. The algorithm should also be able to identify any tags that are outliers among the group. The second approach entails the development of a gaussian naive bayes classifier, with the goal of identifying whether an article is anomalistic or not - given the group of tags it's been assigned.

Large Language Models (LLMs) have become ever larger in recent years, typically demonstrating improved performance as the number of parameters increases. This thesis investigates how the probability distributions output by language models differ depending on the size of the model. For this purpose, three features for capturing the differences between the distributions are defined, namely the difference in entropy, the difference in probability mass in different slices of the distribution, and the difference in the number of tokens covering the top-p probability mass. The distributions are then put into different distribution classes based on how they differ from the distributions of the differently-sized model. Finally, the distributions are transformed to be more similar to the distributions of the other model. The results suggest that classifying distributions before transforming them, and adapting the transformations based on which class a distribution is in, improves the transformation results. It is also shown that letting a classifier choose the class label for each distribution yields better results than using random labels. Furthermore, the findings indicate that transforming the distributions using entropy and the number of tokens in the top-p probability mass makes the distributions more similar to the targets, while transforming them based on the probability mass of individual slices of the distributions makes the distributions more dissimilar.

I takt med den växande tekniska utvecklingen har e-handeln börjat ta en allt större plats bland konsumenter och företag. Möjligheten att göra inköp på nätet prioriteras högt, särskilt i samband med Covid-19 pandemin då e-handeln i många fall tog över helt från fysiska butiker. Samtidigt har även utvecklingen inom artificiell intelligens gått i snabb takt. Det har möjliggjort för företag att använda dessa tekniker för att göra sig konkurrenskraftiga inom e-handeln för att stödja olika aktiviteter relaterade till försäljning online. Artificiell intelligens kan definieras som den träning eller programmering av en dator som gör den kapabel att utföra uppgifter som vanligtvis kräver en människas intelligens. Vanliga delområden inom artificiell intelligens relaterade till studiens ämnesområde är maskininlärning, djupinlärning och språkteknologi. Problemet som studien grundar sig i är att e-handelsföretag använder artificiell intelligens inom marknadsföring och försäljning i en för låg utsträckning och att de därför riskerar att hamna efter, vilket ligger till grund för frågeställningen: Hur använder företag artificiell intelligens för att effektivisera försäljningen inom e-handel? Ämnet undersöks närmare genom en undersökning med kvalitativa, semistrukturerade intervjuer med lämpliga yrkeskunniga personer inom området. Insamlade data visade att e-handelsföretag främst använder artificiell intelligens i samband med riktad annonsering via tredjepartsleverantörer som Google och Facebook. Det framkom ur studien att artificiell intelligens-baserade chatbotar tillämpas för en effektivare hantering av kundtjänst. Ytterligare tredjepartsverktyg används för effektivisering av marknadsföring och försäljning vilket främst rörde text- och språkhantering men även bild- och innehållsgenerering. Detta ledde till sparad tid och arbete samt ökad effektivisering av marknadsföringen och försäljningen hos e-handelsföretag. Denna studie har endast undersökt ett urval av hur artificiell intelligens används inom e-handel, och det finns fortfarande många områden att utforska. Gällande samhälleliga konsekvenser möjliggör studiens resultat en tydligare bild av hur e-handelsföretag arbetar med artificiell intelligens inom marknadsföring och försäljning idag. Studien ger en bild av hur användning ser ut nu, med syfte att underlätta vidare forskning med vilken riktning utvecklingen kommer ta framåt samt stötta företag att börja använda artificiell intelligens. Studien har utförts med en strävan efter hög kvalitet och trovärdighet genom tillämpning av etablerade metoder och transparent dokumentation. / In line with the growing technological development, e-commerce has begun to take an increasingly large place among consumers and companies. The ability to make purchases online is of high priority, especially since the covid-19 pandemic, when e-commerce in a lot of cases completely took over from shopping in physical stores. At the same time, the development in artificial intelligence has proceeded at a rapid pace. This has enabled companies to use these technologies to stay competitive in ecommerce to support various activities related to sales online. Artificial intelligence is a machine that can perform cognitive functions typically associated with humans, such as perceiving, reasoning, learning, and interacting. Common subfields in artificial intelligence related to this study are machine learning, deep learning, and natural language processing. The problem that the study is based on is that e-commerce companies use artificial intelligence in marketing and sales to a low extent which means they risk falling behind. Based on this, the research question is: How do companies use artificial intelligence to create efficient sales in e-commerce? The subject is examined through a survey with qualitative semi-structured interviews with professionals in the field. The collected data clearly shows that e-commerce companies primarily use artificial intelligence in connection with targeted advertising through third-party providers such as Google and Facebook. It was found that artificial intelligence based chatbots are applied for more efficient management of customer service. Additional third-party tools are used for streamlining marketing and sales, which mainly involve text and language management but also image and content generation. This leads to saved time and effort as well as increased efficiency in marketing and sales for e-commerce companies. This study has only examined a selection of artificial intelligence usage within e-commerce, and there are still many areas to explore. Regarding societal consequences, the study's results help give a clearer picture of how e-commerce companies work with artificial intelligence in marketing and sales today. The study provides a snapshot of current usage to facilitate further research on the direction of future development and support companies in starting to use artificial intelligence. The study was conducted with an aim for high quality and credibility through the application of established methods and transparent documentation.

Pre-processing Arabic text for Text-to-Speech (TTS) systems poses major challenges, as Arabic omits short vowels in writing. This omission leads to a large number of homographs, and means that Arabic text needs to be diacritized to disambiguate these homographs, in order to be matched up with the intended pronunciation. Diacritizing Arabic has generally been achieved by using rule-based, statistical, or hybrid methods that combine rule-based and statistical methods. Recently, diacritization methods involving deep learning have shown promise in reducing error rates. These deep-learning methods are not yet commonly used in TTS engines, however. To examine neural diacritization methods for use in TTS engines, we normalized and pre-processed a version of the Tashkeela corpus, a large diacritized corpus containing largely Classical Arabic texts, for TTS purposes. We then trained and tested three state-of-the-art Recurrent-Neural-Network-based models on this data set. Additionally we tested these models on the Wiki News corpus, a test set that contains Modern Standard Arabic (MSA) news articles and thus more closely resembles most TTS queries. The models were evaluated by comparing the Diacritic Error Rate (DER) and Word Error Rate (WER) achieved for each data set to one another and to the DER and WER reported in the original papers. Moreover, the per-diacritic accuracy was examined, and a manual evaluation was performed. For the Tashkeela corpus, all models achieved a lower DER and WER than reported in the original papers. This was largely the result of using more training data in addition to the TTS pre-processing steps that were performed on the data. For the Wiki News corpus, the error rates were higher, largely due to the domain gap between the data sets. We found that for both data sets the models overfit on common patterns and the most common diacritic. For the Wiki News corpus the models struggled with Named Entities and loanwords. Purely neural models generally outperformed the model that combined deep learning with rule-based and statistical corrections. These findings highlight the usability of deep learning methods for Arabic diacritization in TTS engines as well as the need for diacritized corpora that are more representative of Modern Standard Arabic.

In this study, the current use of natural language processing in Swedish and international nuclear power plants has been investigated through semi-structured interviews. Furthermore, natural language processing techniques have been studied to find out how text data can be analyzed and utilized to aid operations and maintenance in the Swedish nuclear power plant industry. The state-of-the-art transformers model BERT was used to analyze text data from operations at a Swedish nuclear power plant.  This study has not managed to find any current implementations of natural language processing techniques for operations and maintenance in Swedish nuclear power plants. Natural language processing does exist in examples such as embedded search functionalities internally or chatbots on the customer side, but it does not relate to the scope of this project. Some international actors have successfully implemented natural language processing for the classification of text data such as corrective action programs. Furthermore, it was observed that the lingo and jargon in the nuclear power plant industry differ between utilities as well as from the native language. To tackle this, models further trained on domain-specific data could be beneficial to better analyze the text data and solve natural language processing tasks. As the data used in this study was unlabeled, expert input from the nuclear domain is required for a proper analysis of the results. Working for a more data-driven industry would be valuable for the implementation of natural language processing. / I denna studie har den nuvarande användningen av Natural language processing (NLP) i svenska och internationella kärnkraftverk undersökts genom semistrukturerade intervjuer. Vidare har NLP studerats för att ta reda på hur textdata kan analyseras och användas för att underlätta drift och underhåll i den svenska kärnkraftsindustrin. Transformersmodellen BERT användes för att analysera textdata från driften vid ett svenskt kärnkraftverk. Denna studie har inte lyckats hitta några aktuella implementeringar av NLP för drift och underhåll i svenska kärnkraftverk. NLP finns som inbäddade sökfunktioner internt eller chatbottar på kundsidan, men dessa omfattas inte av detta projekt. Vissa internationella aktörer har framgångsrikt implementerat NLP för klassificering av textdata som t.ex. avhjälpande underhåll (Corrective action programs). Vidare observerades att språket och jargongen inom kärnkraftsindustrin skiljer sig mellan olika kraftverk och från det vanliga språket. Genom att träna modellerna på domänspecifik data skulle modellerna kunna prestera bättre. Eftersom data som användes i denna studie var omärkt (unlabeled), krävs expertinput från kärnkraftsområdet för en korrekt analys av resultaten. Att arbeta för en mer datadriven industri skulle vara värdefullt för implementeringen av NLP / Feasibility Study on Artificial Intelligence Technologies in Nuclear Applications

This case study investigates a chatbot and its responsiveness to user expression, utility and the users' experience. A user study was performed with both a questionnaire and user interviews. A program to automatically categorize user input messages was made in order to analyze the correlations and reasons for failed predictions and misunderstandings. The current utility is not reaching the return on investment, although expansions are planned and the launch is underway. The strongest and most meaningful correlations with unsuccessful messages were with messages that were out of the chatbot's scope, inside of the scope but unsolved, messages that lacked training and test data or that contained at least one unknown word. The user rated the chatbot's ability to understand them slightly above the middle of the scale, even though the rate of unsuccessful messages was high since the chatbot is still under development. Misspellings, vernacular and organization specific words did not have a particularly strong correlation to unsuccessful messages. The opinions on humanlike interaction were split both in regard to its current likeness and what would be desirable in the future. However, most users seemed to want the conversation to feel natural as if from a human and to receive nuanced answers.

This study explores the possibilities of classifying language as governing or not. The ground premise is to examine how detecting and quantifying governing conditions from thousands of financial grants in appropriation directions can be performed automatically, as well as creating a data set to perform machine learning for this text classification task. In this study, automatic classification is performed along with an annotation process extracting and labelling data. Automatic classification can be performed by using a variety of data, methods and tasks. The classification task aims to mainly divide conditions into being governing of the conducting of the specific agency or not. The data consists of text from the specific chapter in the appropriation directions regarding financial grants. The text is split into sentences, keeping only sentences longer than 15 words. An iterative annotation process is then performed in order to receive labelled conditions, involving three expert annotators for the final data set, and laymen annotations for initial experiments. Given the data extracted from the annotation process, SVM, BiLSTM and KB-BERT classifiers are trained and evaluated. All models are evaluated using no context information, with bullet points as an exception, where a previous, generally descriptive sentence is included. Apart from this default input representation type, context regarding preceding sentence along with the target sentence, as well as adding specific agency to the target sentence are evaluated as alternative data representation types. The final inter-annotator agreement was not optimal with Cohen’s Kappa scores that can be interpreted as representing moderate agreement. By using majority vote for the test set, the non-optimal agreement was somewhat prevented for this specific set. The best performing model all input representation types considered was the KB-BERT using no context information, receiving an F1-score on 0.81 and an accuracy score on 0.89 on the test set. All models gave a better performance for sentences classed as governing, which might be partially due to the final annotated data sets being skewed. Possible future studies include further iterative annotation and working towards a clear and as objective definition of how a governing condition can be defined, as well as exploring the possibilities of using data augmentation to counteract the uneven distribution of classes in the final data sets.