• Digital Humanities and Cultural Heritage
• Publicationsclose
• Research dataclose
• Other research productsclose
• 2019-2023close
• English close
• LUTPub close

Such techniques of Natural Language Processing as information extraction and semantic text labelling had been widely utilised in recruitment sphere to decrease the labour and time resources needed to analyse CVs or labour market’s trends. However, the application of such techniques and establishing link between demand for the workforce and education providing organizations is yet to be established. In the current thesis the ideas on processing educational courses descriptions texts is provided in attempt to facilitate the information exchange between the needs of the labour market and skills supply from the educational establishments. In the literature review the analysis of the most recent methods in natural language processing methods is provided (Word2Vec, NER, Sentence Transformers) as well as commentary on their current implementations in labour market related spheres. In the empirical section state-of-the-art SBERT language model is applied to the collected open university courses’ descriptions in order to extract concrete skills from the and then the performance of the SBERT model is accessed through such metrics as precision, recall and f-score, yielding the F-score of 70.4%. As a result, an example of comparison between the skills supplies as identified by Finnish open universities educational courses and demand as identified by the job descriptions data is provided. In conclusion, research paper’s possible managerial applications and theoretical contribution are included.

In this work, the utilization of financial news alongside machine learning for predicting stock market movements is examined. The news are handled with various natural language processing methods for finding correlation between the derived attributes and stock market movements. The novelty of this work lies in the application of BNS and LDA methods as well as 2-word combinations alongside with LSTM neural network. The main point of the work is to examine the usefulness of the results achieved with the formerly mentioned methods and neural networks as well as comparing the results with market efficiency. In the research it was concluded that the models containing 2-word combinations derived with the BNS-method, produced differing results to those models, where the 2-word combinations were not used. However, the overall results followed random patterns and thus reliable results were not achieved. For achieving more reliable results, better approach could be predicting intraday stock market movements per Efficient Market hypothesis. Used datasets were possibly also too concise for the complexity of the problem. Tässä työssä tarkastellaan talousuutisten hyödyntämistä yhdessä koneoppimisen kanssa osakemarkkinoiden ennustamiseen. Uutisia käsitellään tietyillä luonnollisen kielen käsittelyn menetelmillä ja niistä pyritään löytämään korrelaatiota osakkeen liikkeiden kanssa. Työn uutuusarvona ovat BNS- ja LDA-menetelmien, sekä 2 sanan kombinaatioiden käyttö LSTM-neuroverkon yhteydessä. Työn pääasiallisena tavoitteena on tarkastella edellä mainittujen menetelmien ja neuroverkkojen yhdessä tuottamien tulosten hyödyllisyyttä ja niiden vertautumista markkinoiden tehokkuuteen. Tutkimuksessa selvisi, että mallit, jotka sisältävät BNS-menetelmällä johdettuja 2 sanan kombinaatioita, tuottavat poikkeavia tuloksia niihin malleihin verrattuna, joissa niitä ei käytetä. Tulokset kuitenkin noudattivat kaiken kaikkiaan satunnaista vaihtelua ja luotettavia tuloksia ei täten saatu. Tulosten parantamiseksi, parempi lähtökohta voisi olla päivän sisäisten vaihtelujen ennustaminen, markkinoiden tehokkuuden hypoteesin mukaisesti. Käytetyt tietoaineistot olivat myös mahdollisesti liian suppeat kompleksisuudeltaan vaativalle ongelmalle.

Fake news is not a novel concept but the scale of its spread and the damage it has and continues to cause is alarming. From the US presidential elections in 2016 to the COVID-19 pandemic and today, fake news has been circulating in news media corrupting the public opinion. Fake news alters democratic discussions polarizing people’s opinions sowing distrust in national institutions and setting different groups against each other. It is a threat to democracy and national security. It is crucial to prevent fake news from spreading and one solution is to create an automatic fake news detection system. A solution is researched using natural language processing (NLP) tasks, namely text classification. NLP is a type of artificial intelligence that is essentially taught to understand human language. Using thematic analysis, the main steps and techniques of fake news detection models are described and through a comparative analysis the state-of-the-art models are distinguished. And while there are many potential fake news detection models for English there is not much variety for other languages. So, it is additionally analysed if these benchmark models can be implemented for Finnish language as well. Valeuutiset eivät ole uusi käsite, mutta niiden leviämisen laajuus ja niiden aiheuttamat vahingot ovat huolestuttavia. Yhdysvaltain presidentinvaaleista vuonna 2016 COVID-19-pandemiaan ja nykypäivään asti, uutismediassa on kiertänyt valeuutisia, jotka muokkaavat yleisön mielipidettä. Valeuutiset muuttavat demokraattista keskustelua polarisoimalla ihmisten mielipiteitä kylväen epäluottamusta kansallisiin instituutioihin ja asettaen erilaisia ryhmiä toisiaan vastaan. Se on uhka demokratialle ja kansalliselle turvallisuudelle. On tärkeää estää valeuutisten leviäminen, ja yksi ratkaisu on luoda automaattinen valeuutisten havaitsemisjärjestelmä. Ratkaisua tutkitaan käyttämällä luonnollisen kielen käsittelyn (NLP) tehtäviä, etenkin tekstin luokittelua. NLP on tekoälyn tyyppi, missä tietokone opetetaan ymmärtämään ihmisten kieltä. Temaattisen analyysin avulla kuvataan valeuutisten havaitsemismallien päävaiheet sekä tekniikat, ja vertailevan analyysin avulla valikoidaan uusimmat ja onnistuneimmat mallit. Ja vaikka englannin kielellä on monia mahdollisia valeuutisten havaitsemismalleja, muille kielille ei ole paljon valikoimaa. Lisäksi analysoidaan, voidaanko nämä mallit toteuttaa myös suomen kielelle.

The increased amount of fake news has created a demand for different fake news detection methods. One way to detect fake news is with machine learning models. While there is a lot of research on different fake news detection models, there is not that much research done on the effects of sentiment information on the classification accuracy of the models. Sentiment information means the overall tone of each of the news articles in the dataset, whether it is positive, neutral, or negative. The goal of this thesis is to find out how sentiment information affects the performance of two of the most popular fake news detection machine learning models. These models are the Naïve Bayes and Support Vector Machine. The sentiment analysis was done with TextBlob and Vader, which are already tested and trained sentiment analysis tools available in Python. The results showed that sentiment information did not have any significant effect on the classification accuracy of the fake news detection models. In most of the cases the addition of sentiment information slightly decreased the accuracy of the models. Valeuutisten kasvanut määrä on aiheuttanut tarvetta löytää erilaisia menetelmiä valeuutisten havaitsemiseen. Yksi tapa havaita valeuutisia on koneoppimismenetelmien avulla. Vaikka erilaisia valeuutisten havaitsemiseen käytettyjä koneoppimismalleja on tutkittu paljon, ei ole vielä paljon tutkimusta siitä, miten tunnetieto vaikuttaa koneoppimismallien luokittelutarkkuuteen. Tunnetiedolla tarkoitetaan tietoaineiston kunkin uutisartikkelin yleistä sävyä, eli onko artikkeli positiivinen, neutraali vai negatiivinen. Tämän työn tavoitteena on selvittää, millainen vaikutus tunnetiedolla on kahteen suosituimpaan valeuutisten havaitsemiseen käytetyn koneoppimismallin suorituskykyyn. Nämä mallit ovat Naïve Bayes ja Support Vector Machine. Tunneanalyysi tehtiin käyttäen TextBlobia ja Vaderia, jotka ovat Pythonista löytyviä valmiiksi valmennettuja ja testattuja tunneanalyysityökaluja. Tulokset näyttivät, että tunnetiedolla ei ollut merkittävä vaikutus valeuutisia havaitsevien koneoppimismallien luokittelutarkkuuteen. Suurimmassa osassa tuloksista tunnetietojen lisääminen hiukan laski mallien tarkkuutta.

Drying, a crucial step in process engineering aimed at producing optimal product moisture content, has evolved over time from batch processing methods to continuous processing alternatives. Continuous drying methods offer uniform moisture content of the product at lower operational cost. In this study, a continuous veneer drying model was developed based on mass and energy balances. The simulated veneer dryer is a semiautomatic machine designed to maximise the drying process efficiency via control mechanisms such as the veneer transport rate, fan speed, opening angle of the inlet and outlet dampers, and radiator temperature. In the dryer, veneer plates are conveyed horizontally through the number of connected chambers where hot air is blown transversely. The optimal drying process is dynamically maintained via the Proportional–integral–derivative controllers, manipulating the rate of the damper lids opening, that are connected to the sensors monitoring the air properties in the chambers of the drying unit. The model-based sensitivity analysis ANOVA was carried out for energy optimisation purposes. The analysis outcomes indicated that radiator temperature, initial moisture content of veneer sheets and conveyor speed are the most influential parameters affecting the drying rate. Automatic control of damper lids provides optimal temperature and moisture content of drying environment at lower energy costs. Post-print / Final draft

This doctoral study was conducted as an inquiry of the international entrepreneur in order to balance off the domination of firm-level studies and limited discussion of the individual as the initial driver in the international entrepreneurship phenomenon (IE). Two research questions guide the study: ‘How do individuals make sense of themselves as becoming and being international entrepreneurs?’ and ‘How to theorise of individuals becoming and being international entrepreneurs through a narrative approach?’. With a processual view of the phenomenon, this study embraces IE as a journey and approaches histories and sense-making of individuals through narrative inquiry, paying attention to the different efforts by which entrepreneurs (and researchers) contextualize—and constitute—the personal-level IE journeys. The qualitative dataset consists of interviews and historical data. Data analysis builds on ‘hermeneutic reasoning’, suggesting that meanings and implications of journeys individuals have undertaken can be better grasped after they have unfolded in time. The findings in the four publications construct the contribution of this article-based dissertation. Publications I, II and III embrace narrative sense-making as meaning structure to past actions and lived events and illuminate how the international entrepreneurial ‘self’ as an actor and agent in retrospect manifests individuals as the ‘autobiographical authors’ in regard to developmental, transitional and generational experiences and their meanings in becoming and being an international entrepreneur. They provide evidence of how the founders’ sense-making and identity work feed into the behavioural orientations and ‘bounded and boundaryless’ career journeys of becoming and being international. Publications I, III and IV are novel attempts to address empirically the social historic process in which IE is embedded and its significance for the individual. When analysed against the (inter)generational backdrop of individuals’ actions and life-events, we may trace how international entrepreneurs are the protagonists of their own generations and leaving a legacy to the next.

The changes in the labor market accompanied by digitalization have influenced the approaches used for recruitment purposes. The development of machine learning and, specifically, natural language processing allowed new services for automated resume screening to appear. These services aim to reduce the time that HR specialists dedicate to sorting out unsuitable candidates, thus saving the costs per hire. This master thesis provides an analysis of the possible approach for matching CVs to JDs based on the taxonomy of competencies. In the literature review, there are the description and comparison of the different ontologies and taxonomies, and the analysis of the previous research performed in vacancy–CV matching provided. In the methodology part, there are the concepts of natural language processing and information extraction described, and the Sentence Transformer model and named entity recognition explained. Besides, evaluation approaches for taxonomy enrichment and vacancy–CV matching are suggested. In the solution development part, the described concepts and models are applied to the JDs and CVs data to perform taxonomy enrichment and matching tasks. The assessment of the matching algorithm performance based on the expert’s evaluation is presented. Based on the results obtained, the potential usage of the research and possible limitations are discussed.