20.500.12556/RUL-176566 Veliki jezikovni modeli in njihova uporaba pri analizi sentimenta finančnih novic magistrsko delo Large language models and their use for sentiment analysis of financial news Magistrsko delo obravnava uporabo velikih jezikovnih modelov pri analizi sentimenta finančnih novic. Delo izpostavlja prehod od tradicionalnih metod, ki temeljijo na slovarjih in preprostih pravilih, k naprednim modelom nevronskih mrež, sposobnim zajemanja konteksta in pomena besedila. Jedro teoretičnega dela predstavlja opis arhitekture transformerjev, ki omogoča učinkovito obdelavo zaporednih podatkov z mehanizmom samopozornosti. Predstavljeni so ključni modeli, kot sta BERT in GPT, ter njune različice, prilagojene finančni domeni, med katerimi izstopa FinBERT. V empiričnem delu je izvedena analiza sentimenta finančnih novic z uporabo klasičnega modela in treh velikih jezikovnih modelov. Pridobljeni rezultati napovedanega sentimenta so primerjani z dejanskimi spremembami finančnega indeksa z namenom ocene povezanosti med razpoloženjem v novicah in gibanjem trga. Rezultati potrjujejo, da veliki jezikovni modeli v nekaterih primerih presegajo tradicionalne pristope pri prepoznavanju konteksta in zaznavanju čustvenih odtenkov besedil. To odpira možnosti za njihovo uporabo pri samodejni interpretaciji tržnih informacij in podpori investicijskim odločitvam. The master’s thesis explores the application of large language models in the sentiment analysis of financial news. It emphasizes the shift from traditional lexicon-based and rule-driven approaches to advanced neural architectures capable of contextual understanding and semantic interpretation. The theoretical part presents the transformer architecture, which enables efficient processing of sequential data through the self-attention mechanism. Key models such as BERT and GPT, along with domain-specific adaptations like FinBERT, are discussed. The empirical part involves sentiment analysis of financial news using one classical model and three large language models. The predicted sentiment is compared with movements in financial indices to examine the relationship between news sentiment and market dynamics. The results show that large language models sometimes outperform traditional methods in contextual comprehension and sentiment detection, demonstrating their potential for automated interpretation of market information and enhanced decision-making in finance. veliki jezikovni modeli strojno učenje nevronske mreže analiza sentimenta obdelava naravnega jezika transformer finančne novice samopozornost delniški trgi large language models machine learning neural networks sentiment analysis natural language processing transformer financial news self-attention stock markets true false false Slovenski jezik Angleški jezik Magistrsko delo/naloga 2025-12-04 08:15:14 2025-12-04 08:15:25 2025-12-22 04:05:02 0000-00-00 00:00:00 2025 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.42 156748 259411459 21083.pdf 21083.pdf 1 79FF587E7E8FC6B81130A2B69B4CA920 2975a5fe5ede1ebd9c1326561a779b855e35a721aa5bca6b8b5448b00e0eca74 a0ff6344-d0e0-11f0-9328-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=222974 Fakulteta za matematiko in fiziko 0 0 0