The thesis examines the use of large language processing methods for automated classification and analysis of citizen proposals and comments on Slovenian e-participation portal predlagam.vladi.si. Since the site’s launch in 2009 the portal has accumulated 19 thousand proposals and 135 thousand comments, which are not being automatically categorized or analyzed, making their effective processing challenging. In the empirical part of the study there were several approaches and models evaluated across three tasks: thematic classification of proposals (9 classes), emotional tone of proposals (4 classes) and comment sentiment (3 classes). A manually labeled sample of 500 proposals and 300 comments was created and split into training and test set in a 70:30 ratio. Six models were evaluated on this data: classical TF-IDF + SVM and fastText approaches, local models Llama 3.2-3B and GaMS3-12B in zero-shot setting as well as fine-tuned with QLoRA and commercial OpenAI GPT-5-mini model. The latter achieved the best result across all three tasks, correctly classifying almost three-quarters of proposals in thematic classification task. The closest performance was achieved by Slovenian local model GaMS3-12B fine-tuned with QLoRA with only a few percentage points difference. Emotional tone classification proved challenging for all models with the best result reaching an F1 score of 0.49. This work provides a systematic overview of the applicability of different NLP methods to Slovenian public administration texts. It establishes a manually labeled dataset and defines three categorization schemes based on the collected portal data.
|
