| Form of studies |
Bachelor |
| Title of the study programm |
Computer Systems |
| Title in original language |
Nestrukturētu sludinājumu datu apstrāde un atslēgas informācijas ieguve ar lielo valodas modeļu palīdzību |
| Title in English |
Processing of Unstructured Advertisement Data and Extraction of Key Information with the Help of Large Language Models |
| Department |
Faculty Of Computer Science Information Tehnology And Energy |
| Scientific advisor |
Arnis Staško |
| Reviewer |
Balasuriyage Aritha Dewnith Kumarasinghe |
| Abstract |
The rapid growth of classified digital marketplaces has led to an abundance of unstructured, free-form advertisement texts that contain messy formatting, multilingual content, abbreviations, and slang. Traditional rule-based and supervised machine learning methods are often unsuitable for processing this highly variable data without extensive manual effort. Therefore, this thesis explores the feasibility of using modern commercial Large Language Models (LLMs) to automatically extract structured key information from these chaotic listings.
The research conducts a systematic empirical evaluation comparing two state-of-the-art models: Google's Gemini 3.1 Pro and Anthropic's Claude 4.6 Sonnet. The models are tested using two prompting strategies: Standard Zero-Shot and a more sophisticated Chain-of-Thought approach with negative constraints. The experiments utilize a manually curated ground-truth dataset of 102 multilingual used car listings from the Latvian portal ss.com.
The results demonstrate that both models achieve near-perfect accuracy for easy-to-extract attributes like Make and Year. While Chain-of-Thought prompting successfully mitigates the over-extraction of complex attributes for both models, it causes a drastic, model-specific degradation in recall for Claude concerning numerical attributes like mileage and price. Gemini, however, does not experience a significant decrease in these same metrics. The findings conclude that seemingly universal prompt engineering techniques produce model-specific effects, meaning optimal strategies depend heavily on the specific architecture used. Guided by these comparisons, the thesis provides practical guidelines for developing effective LLM-based information extraction pipelines and highlights directions for future research.
Volume Information: The graduation thesis contains 63 pages , 3 figures , 1 appendix , and 12 sources of reference. |
| Keywords |
Lielie valodu modeļi, informācijas iegūšana, nestrukturēti dati, uzvedņu inženierija. |
| Keywords in English |
Large Language Models, Information Extraction, Unstructured Data, Prompt Engineering. |
| Language |
eng |
| Year |
2026 |
| Date and time of uploading |
26.05.2026 23:56:43 |
