Prilagodba modela transformatora za strojno prevođenje u južnoslavenskim dijalektima

undergraduate thesis

• Title (english): Adapting transformer models for machine translation in South-Slavic dialects
• Author: Gašpar Haramija
• Mentor: Jan Šnajder (mentor)
• Committee member: Jan Šnajder (predsjednik povjerenstva)
• Committee member: David Dukić (član povjerenstva)
• Committee member: Ana Barić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-03, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovom radu istražuju se metode za poboljšanje strojnog prevođenja za južnoslavenske dijalekte i jezike s malim resursima, fokusirajući se na čakavski, cerljanski i torlački dijalekt. Koristeći model No Language Left Behind (NLLB), ovaj rad istražuje tri pristupa koji su se temeljili na skupu podataka COPA: osnovni, fino ugađani i model obrnutog prijevoda. Fino ugađani modeli nadmašili su osnovne za 17.70 BLEU i 17.99 chrF2 bodova, dok je model obrnutog prijevoda, koji se temeljio na tehnici obrnutog prijevoda za proširenje izvornog skupa podataka, nadmašio fino ugađane modele za 6-16 BLEU i 11-28 chrF2 bodova. Međutim, rezultati na dodatnom paralelnom skupu podataka pokazali su ograničeni potencijal generalizacije.

Abstract (english)

This work investigates methods to enhance machine translation for low-resourced South Slavic dialects, focusing on Chakavian, Cerkno, and Torlak dialects. Using the No Language Left Behind (NLLB) model, this study explores three approaches that were based on the COPA dataset: baseline, fine-tuned, and back-translation model. Fine-tuned models surpassed baseline ones by 17.70 BLEU and 17.99 chrF2 points, while the back-translation model, which was based on the technique of back-translation for augmenting the original dataset, outperformed fine-tuned models by 6-16 BLEU and 11-28 chrF2 points. However, results on an additional parallel dataset showed limited generalization potential.

• Keywords: obrada prirodnog jezika; strojno prevođenje; jezici s malim resursima; dijalekt
• Keywords (english): natural language processing; machine translation; low-resource languages; dialect
• Language: english
• URN:NBN: urn:nbn:hr:168:713476
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 20:46:30