From: 2024-08-29 10:15 to 16:00
Place: See information for each presentation
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se

Six MSc theses to be presented on Thursday August 29, 2024

Thursday August 29 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Six MSc theses will be presented.

You will find information about how to follow along under each presentation. There will be presentations in two different rooms: E:2116 and E:4130 (Lucas). See room for each presentation. A preliminary schedule follows.

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.

10:15-11:00 only via Zoom (see link below)

Presenter: Desnoyer Jonathan
Title: Leveraging Large Language Models for Event Extraction
Examiner: Jacek Malec
Supervisors: Nugues Pierre (LTH), Martin Carla (QuantCube Technology)

In light of generative AI's recent explosion in popularity, a lot of research is being made to leverage large language models that have demonstrated impressive capabilities across numerous fields. These developments could be advantageous for event extraction, as previous methods depended heavily on large annotated datasets and lacked generalizable methodology. Using three recent large language models (Phi-2, Phi-3, and Mistral 7B) that are categorized as small language models, we investigate the process of fine-tuning for the event extraction task. We show that fine-tuning these models enhances the results on both document and sentence level event extraction, with Mistral 7B outperforming Phi-2 and Phi-3 on both levels. Document level extraction being a challenging task, reaching 0.329 and 0.089 F1-score against 0.576 and 0.316 for sentence-level extraction on trigger and argument extraction, respectively.

Link to popular science summary: To be uploaded

Link to Zoom presentation: https://lu-se.zoom.us/j/61652376857?pwd=nhuniOzQPuQFHY67CsPZpdxaTvkHnd.1

11:15-12:00 in E:4130 (Lucas)

Presenter: Tom Richter
Title: Enhancing Sentiment Analysis of Mobile App Reviews: Investigating Novel Techniques in Aspect-Based Opinion Mining
Examiner: Jacek Malek
Supervisor: Markus Borg (LTH)

This thesis investigates the application of Aspect-Based Sentiment Analysis (ABSA) on mobile app reviews using large language models (LLMs). While general-purpose LLMs like GPT-3.5 and GPT-4 offer powerful capabilities, their proprietary nature and high computational requirements pose challenges for deployment in resource-constrained environments. This research explores the Distilling Step-by-Step methodology to create smaller, specialized models that retain the reasoning capabilities of larger LLMs.. Using a large dataset of customer reviews from Apples's App Store, this study aims to develop a high-performing model for ABSA by leveraging advanced distillation techniques. The findings demonstrate the feasibility and effectiveness of deploying smaller models for sentiment analysis, offering practical solutions for businesses with limited resources.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240829_11Richter.pdf

13:15-14:00 in E:2116

Presenter: Marie Ask Uggla
Title: Hybrid work within agile software development teams
Examiner: Per Runeson
Supervisor: Elizabeth Bjarnason (LTH)

Hybrid work has become increasingly common in recent years. Integrating hybrid work with agile practices presents challenges, as agile principles rely heavily on close collaboration and co-location of team members. This thesis investigates how hybrid work is currently utilized in agile teams, the challenges it creates, and the practices applied to address these challenges. A case study approach was utilized, consisting of five semi-structured interviews and a survey distributed to different case companies, amassing a total of 125 responses. Most employees choose to work remotely to some extent and many companies establish guidelines, particularly on remote work limits. The biggest challenge in incorporating a hybrid work environment in agile teams was creating a successful meeting strategy and getting hybrid meetings to work. Most participants felt that hybrid work had a positive impact on their well-being and the added flexibility was positive for many individuals' work-life balance.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240829_13AskUggla.pdf

14:15-15:00 in E:2116

Presenter: Max Fogwall
Title: Synchronizing CI Workflows Across Repositories Using GitHub Actions
Examiner: Elizabeth Bjarnason
Supervisors: Ulf Asklund (LTH), Lars Andersson & José Díaz López (QlikTech International AB)

GitHub Actions is a CI/CD tool integrated into repositories of GitHub. In this paper, I explore the reusability challenges of having a set of cross-repository workflows in GitHub Actions by analyzing the industry organization, Qlik, and 6 large open source GitHub organizations. I find that GitHub Actions requires each included repository to have copies of workflows, and that its public reusability features are insufficient to properly manage these copies all at once, such as adding new workflows or changing an existing trigger condition. Instead, a tool is necessary to synchronize such changes across repositories, which has its own challenges. The improvement I propose to Qlik includes stripping the copied workflows down to references of central code, along with proper versioning. A small prototype of this suggested that most problems that Qlik had, including instability, high maintenance cost, and long integration times, could be mitigated by this, assuming it scales well.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240829_14Fogwall.pdf

14:15-15:00 in E:4130 (Lucas)

Presenter: Erik Dahlberg
Title: Examining the Impact of Generating Synthetic Images for Polyp Segmentation in Colonoscopy
Examiner: Jacek Malec
Supervisor: Marcus Klang (LTH)

Colon cancer, responsible for over 800,000 and 1.3\% of deaths annually worldwide, remains one of the five most common and deadly cancer. Colonoscopy is the primary screening method used to inspect the colon and identify polyps and carcinomas. If left undetected, pre-cancerous polyps and carcinomas can develop into fatal cancer. Moreover, polyps can be easily missed during screening, due to, for instance, difficult navigation of the colon or because certain polyps, like the sessile serrated polyp, closely resemble colon tissue. To assist colonoscopists in identifying polyps and carcinomas, Computer Aided Detection (CADe) systems have been developed for polyp detection and segmentation. However, the high cost and need of obtaining medical image data for CADe systems, have prompted research into using generative models to enhance dataset diversity. This dissertation investigates the potential of generated colon data by conducting a feasibility study on available Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs). Two DDPM architectures were trained and evaluated for generating colon polyp segmentation data. The inclusion of generated images improved the IoU performance of polyp segmentation models by a maximum of 3.3%.

Link to popular science summary: To be added

15:15-16:00 in E:4130 (Lucas)

Presenter: Maxime Pakula
Title: Resume Quality Assurance and Enhancement Using Small LLMs
Examiner: Jacek Malec
Supervisors: Pierre Nugues (LTH), François-Pierre Chalopin (Takima)

Resumes are essential for standing out in a competitive job market, but the process requires careful attention to both content and form. Experienced professionals often lack the time to help first-time writers. Recent advancements in Natural Language Processing (NLP) and Large Language Models (LLM) can help automate quality assurance and enhancement of resumes, enabling applicants to focus more on content rather than formatting. This research focuses on using small LLMs and prompt engineering to evaluate the quality and enhance French-language resumes. I focused on a specific section of the resume and stuck to particular guidelines. I tested a wide range of small models and used various prompts, experimenting with different instruction languages, guideline types, and prompt lengths. This approach allowed me to draw conclusions on how to effectively prompt small LLMs for optimal performance in the desired tasks.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240829_15Pakula.pdf