1,721,116 research outputs found
Adequacy in Process Modeling: A Review of Measures and a Proposed Research Agenda - Position Paper -
No full textAdequacy of process design is closely connected to the notion of process model adequacy, which in turn is a surrogate for model quality. Quality of modeling is an important field of research in which, however, a comprehensive and gen-erally acknowledged understanding is still outstanding. Notions of "model", "adequacy" and "quality" often remain vague and focus on single aspects such as "syntax" or "semantics" rather than a comprehensive perspective. In this paper we review existing measures and proposals for process model quality as a surrogate for modeling adequacy. Forthcoming from this review we argue that it is foremost the question of modeling pragmatics that is of pertinence when trying to ascertain the adequacy of process modeling. We illustrate how prag-matic concerns mediate traditional conceptions of model quality. The paper ends with a proposed research agenda in the area of process model quality that stipulates an adequacy perspective with a closer focus on the socio-organizational context in which process modeling occurs
Multi-Turn Generation-Based Conversational Agents in Open Domains
Get PDFThis PhD thesis focuses on open-domain generation-based conversational agents, which are chatbots that generate responses to any input or question using natural language processing and deep learning techniques. The thesis identifies three major challenges faced by these conversational agents.
(1) Generating appropriate responses for a wide range of topics and domains. Current studies have focused on single-corpus training, which limits the model's ability to generate relevant responses for certain topics.
(2) Improving a model's performance of context attention distribution in multi-turn settings. The ability to distribute attention and assign importance to relevant information is necessary to generate appropriate responses. However, most existing works have treated multi-turn conversations as one-turn contexts, limiting the performance of the agents.
(3) Integrating knowledge under the conversational question-answering task perspective. There is a gap in research on integrating extractive question-answering techniques with instruction-based tuning and prompt-based tuning. The thesis proposes several approaches to address these challenges.
For (1), the thesis proposes Document-specific Frequency (DF) as an evaluation metric and proposes several methods for balancing multiple corpora. The best method, which integrates DF with the training, achieves an improvement by 34.1% on F1 performance and at least 20.0% on DF. A thorough human evaluation shows a highly significant (p < 0.001) improvement in all of our proposed methods.
For (2), the thesis proposes Distracting Attention Score ratio (DAS ratio) as an evaluation metric and employs self-contained negative samples and summarization techniques to improve a system's performance on context attention distribution. The proposed self-contained negative samples are applied as a training strategy, resulting in about 10% better DAS ratio. The best summarization technique setting with ORACLE gains a 23% improvement on the DAS ratio.
For (3), the thesis explores various settings of integrating extractive question answering with instruction-based tuning, prompt-based tuning, and multi-task learning. When combining prompt-based tuning with either instruction-based tuning or multi-task learning, the F1 performance is improved by about 18% over the baseline.
Together, these techniques have improved the overall performance of multi-turn conversational agents on open domains
An Ontology-Driven Recommender System for Engineering Projects
Get PDFKnowledge and information resources in enterprises are rapidly growing. The International Data Corporation (IDC) forecasts that significant yearly growth of data will result that the so-called global datasphere will have grown to 163 zettabytes (ZB) by 2025, which is 10 times of the 16.1 ZB of data generated in 2016. This happens while IT staff to manage it will grow less than 1.5 times (Reinsel, Gantz, & Rydning, 2017). A substantial number of these resources are documents that are potentially valuable for intentional reuse. Knowledge workers and engineers in particular, require specific knowledge and information embedded in different types of knowledge objects stored in internal or external resources (Hertzum & Pejtersen, 2000). However, identifying relevant knowledge from a large number of unstructured enterprise resources is challenging for users. There is a strong need for an approach that identifies users’ required information and automatically explores their preferred documents.
This PhD project focuses on improving knowledge access, sharing, and reuse challenges that people, engineers, are faced with in their daily (knowledge-based) work tasks. The proposed solution is a recommender system in professional settings to provide relevant documents for users in specific work contexts based on domain-specific ontologies. A prototype has been developed and validated on a multidisciplinary engineering use case and its performance has been evaluated. The results show that the developed system is a useful tool for improving information access in traditional engineering Projects compared to the currently applied solutions. The main contributions of this thesis are:
C1: In-depth analysis of the context of users and the document corpus in an engineering setting by applying information retrieval tools and semantic annotation.
C2. Proposing a framework for a knowledge access system combining recommendation approaches, ontologies, and information retrieval and extraction tools.
C3. Construction of a contextual ontology as knowledge domain, derived from users’ work contexts and evaluating its retrievability and coverage against existing documents as resources of knowledge and information.
C4. Validation of the concept of the recommender system for improving knowledge and information Access in engineering context by developing a system that uses the proposed ontology-based profiling approach and evaluating the performance of the developed system on a case-study
User Privacy in Recommender Systems
Get PDFMed økningen i tilgangen til nettbaserte informasjonskilder så har bruken av anbefalingssystemer tredd frem som et kraftig verktøy for å redusere mengden overført informasjon og samtidig tilby tilpasset innhold for den spesifikke målgruppen. Anbefalingssystemer er mye brukt i alle aspekter knyttet til nettet med alt fra netthandel til dynamisk tilpassede nyheter. Til tross for den økte populariteten så er ikke rekommendasjonssystemene nødvendigvis 100% troverdige på grunn av at den personlige informasjonen som disse systemene samler inn kan utgjøre en personvernrisiko. For en bruker som opplever at privat informasjon blir misbrukt av et slikt system vil naturligvis være skeptiske til slike systemer senere. Derfor tar denne avhandlingen utgangspunkt i forskningen gjort på personvern i anbefalingssystemer som viser at dette kan være et problem som det er verdt å se nærmere på. Denne avhandlingen inkluderer også personvernrisikoer og de tekniske løsningene brukt for å beskytte personlig informasjon, samt de nåværende lovene rundt personvern med tanke på bekymrede brukere.
I motsetning til tidligere forskning utført på personvern ved domeneuavhengige anbefalingssystemer så har nyhetsdomener i denne avhandlingen blitt valgt som et ekstra forskningspunkt. Mer konkret så vil denne avhandlingen identifisere de personlige opplysningene som inngår i anbefalingssystemer for nyhetsdomener. Personalisering av nyheter har blitt mer viktig da en bruker er mer interessert i å holde seg oppdatert på spesifikke nyheter innenfor en kort tidsperiode. Kvaliteten og nøyaktigheten til slike persontilpassede nyheter er avhengig av å sanke informasjon om leserne. Som et eksempel så er ønsker nyhetssamlere slik som Google News at brukere skal logge inn i systemet for å få persontilpassede nyheter. For mer generiske nyhetsforslag så samler systemet brukerens netthistorie og ser mønster i nettsidene brukeren har besøkt. Behovet for brukerprofiler øker risikoen for personvernet i nyhetsdomener, mens logging av en brukers netthistorie fører til en økt risiko for personvernet til en hvilken som helst bruker av nyhetsdomenet.
Til slutt så har det også blitt utført en brukerundersøkelse gjennom en serie med spørreskjemaer for å kartlegge brukeres meninger om personvern på nettet. Det ble konkludert med at en brukers preferanser med tanke på personvern, hva brukeren visste om innsamling av persondata, samt det eventuelle eierskapet av den innsamlede dataen hadde en stor innvirkning på en brukers mening om personvern. En analyse av resultatet fra undersøkelsen viste også at norske brukere er mindre opptatt av personvern på nettet sammenlignet med brukere fra andre nasjoner
Sentiment Analysis of Norwegian Twitter Messages
Get PDFWeb-Based Social Media (WBSM) have been on the rise for the recent several years, and have subsequently garnered interest from several groups out to proficiently utilize the vast amounts of data found on these sites. Micro-blogging site Twitter is one of the media sites that have embraced developers and interest groups, Twitter has developed accessible frameworks and allows the use of these frameworks enabling developers access to large amounts of information.
In this master thesis, a system for performing Sentiment Analysis (SA) on Norwegian Tweets is described. The system described uses a two-step binary classification process for subjectivity and polarity classification, utilizing different parameters and three different classifiers - Naive Bayes (NB), Support Vector Machines (SVM), and Maximum Entropy (MaxEnt) - for the two different classification tasks. The solution also makes an attempt at exploiting grammatical metadata given by the NTNU SmartTagger as well as cross-lingual sentiment lookup in the SentiWordNet sentiment lexicon in order to achieve improved results from the classifiers.
A system for extracting sentiment targets after classification is also described. This method is a way of utilizing the classifier in order to identify critical sentiment words in order to augment the detection of the target of a given sentiment in a tweet
Automatic Detection of Fake News in Social Media using Contextual Information
Get PDFMisinformation has become an important part of society, especially with the increase in fake news. This thesis investigates how using contextual and network data may be used as a detection system for news articles or other information pieces. Either as a standalone system or part of a bigger, hybrid solution.
A series of experiments have been conducted to explore the validity of contextual information in structured data from Facebook. Two different algorithms have been used, Logistic Regression and Harmonic Boolean Label Crowdsourcing, achieving a diverse result set shedding light on strengths and weaknesses. Using two different datasets consisting of scientific and fake news sources ranging from 4200 to 15.500 posts in size, and up to 9.5 million users, results with over 90 \% accuracy in classification in supervised training scenarios, consolidating previous results on both old and new datasets.
As a result, this thesis concludes with very promising results using contextual data only. This approach is still novel and needs more rigorous testing, but combining it with existing Natural Language Processing systems might yield better results than the current state of the art systems. A lot of work is still needed to be able to apply the methods to less structured data
Automatic Classification of Bank Transactions
Get PDFThis thesis describes implementations of several classification models to be used in a Norwegian bank transaction classification system. Our goal is to investigate which supervised machine learning methods are best suited to solve this multi-class classification problem. We are also using external semantic resources to supplement the information we have about each transaction and in turn attempt to increase the accuracy of the classification system. Determining which external semantic resources we should use and how this is to be implemented is, therefore, an important aspect of this project
Brand sentiment analysis of the Norwegian banking sector
Get PDFFor the last two decades, the world wide web has become a social arena where people can express themselves. People write opinionated texts on social media towards various targets, which is read by other people. The readers are often influenced by what is written. Companies holding brands often keep a close eye on what people write about them, but it is a burdensome task to monitor the World Wide Web. The use of sentiment analysis for automatically analysing brands has been on the rise in recent years for this reason. Norwegian banks are increasingly considering themselves as brands, and they monitor their online reputation accordingly.
This master thesis describe a system performing sentiment analysis on Norwegian reviews of the banking sector. The system use three different classifiers, Naive Bayes, Support Vector Machines and Maximum Entropy, to classify the polarity of reviews in the Norwegian banking sector. A custom made mapping between a Norwegian wordnet and the sentiment lexicon SentiWordNet aids the sentiment analysis with cross-lingual lookup, as Norwegian sentiment resources are limited. A set of unigrams and bigrams for the banking domain is also used as input to the classifiers, as well as textual features. The system also utilize knowledge of semantic structures in an attempt to extract sentiments on sub-aspects mentioned in the reviews
Sentiment Analysis of Norwegian Twitter Messages
No full textWeb-Based Social Media (WBSM) have been on the rise for the recent several years, and have subsequently garnered interest from several groups out to proficiently utilize the vast amounts of data found on these sites. Micro-blogging site Twitter is one of the media sites that have embraced developers and interest groups, Twitter has developed accessible frameworks and allows the use of these frameworks enabling developers access to large amounts of information.
In this master thesis, a system for performing Sentiment Analysis (SA) on Norwegian Tweets is described. The system described uses a two-step binary classification process for subjectivity and polarity classification, utilizing different parameters and three different classifiers - Naive Bayes (NB), Support Vector Machines (SVM), and Maximum Entropy (MaxEnt) - for the two different classification tasks. The solution also makes an attempt at exploiting grammatical metadata given by the NTNU SmartTagger as well as cross-lingual sentiment lookup in the SentiWordNet sentiment lexicon in order to achieve improved results from the classifiers.
A system for extracting sentiment targets after classification is also described. This method is a way of utilizing the classifier in order to identify critical sentiment words in order to augment the detection of the target of a given sentiment in a tweet
Understanding Organizations’ Adoption of AI Technologies: Challenges, Opportunities and Impact
Get PDFDer hvor mange publiserte verk og forskningsbidrag har lett for å bli påvirket av diverse eksterne krefter, vil en masteroppgave fra NTNU ha fordelen av å representere et akademisk forankret produkt. Denne oppgaven utforsker potensialet ved å sammenligne og karakterisere organisasjoner på tvers av industrier, og se hva slags verdi som kan bli hentet ut i kontekst av en organisasjons forhold til kunstig intelligens. Informasjonsinnhenting har blitt gjort gjennom kvalitative and kvantitative forskningsmetoder hvor intervjuer og en undersøkelse står i sentrum, og de ulike organisasjonene utgjør offisielle samarbeidspartnere til oppgaven.
Totalt 8 ulike temaer relatert til hvordan den enkelte organisasjon forholder seg til og definerer adopsjon og bruk av AI, har blitt trukket ut fra intervjuene hvor ulike sammenligninger er understøttet og validert gjennom direkte sitater fra de ulike representantene. En undersøkelse har også bidratt til å ytterligere karakterisere forskjeller og prioriteter mellom de ulike partnerne. Kunnskapen har bidratt til å stadfeste at verdi kan hentes ut i form av den lærdommen og erfaringen de ulike partnerne "tar med til bordet", både likheter og ulikheter, og gjennom å vise at det er hensiktsmessig å kunne utføre en studie av denne typen.
En diskusjon av forskningen og state-of-the-art i respekt av AI-modenhet har resultert i et grovkornet sluttprodukt, en tilnærming mot en AI maturity model. Tilnærmingen er uttrykt gjennom fem definerte dimensjoner i forhold til organisatorisk AI: technical, data, people, societal, og responsible, og fem nivåer av AI-modenhet: initial, believer, adopter, managed, og optimized. Dette fundamentet utgjør et viktig resultat med et potensial for å kunne fortsette forskningen der denne oppgaven slapp, eller ta det i en helt ny retning.Where many published works and research contributions can easily be influenced by various external forces, a master thesis from NTNU has the benefit of representing an academically-rooted product. This thesis explores the potential of comparing and characterizing organizations across industries, and see what kind of value can be extracted in the context of an organization's relationship with artificial intelligence. The information retrieval has been made possible through qualitative and quantitative research methodologies were interviews and a survey takes center stage, and the different organizations make up the official collaborating partners of the thesis.
In total, 8 different themes related to how an organization relates to and defines the adoption and use of AI have been extracted from the interviews where different comparisons are supported and validated through direct quotes from the different representatives. A survey has also contributed to further characterize differences and priorities between the different partners. The knowledge has contributed to confirming that value can be extracted in the form of the learning experience and experience which the different partners bring to the table, both similarities and differences, and by showing that it's relevant to conduct a study of this kind.
A discussion of the research and state-of-the-art in respect of AI-maturity has resulted in a coarse end-product, an approach towards an AI maturity model. The approach is expressed through five defined dimensions in relation to organizational AI: technical, data, people, societal, and responsible, and five levels of AI-maturity: initial, believer, adopter, managed, and optimized. This foundation constitutes an important result with the potential to continue the research of this thesis, or to guide it in a new direction
- …
