Analiza strojnega prevajanja terminologije v angleških in slovenskih besedilih

Simonišek, Živa (author), Vintar, Špela (mentor)

Nevronski strojni prevajalniki se kljub visoki kakovosti še vedno soočajo z določenimi težavami, med drugim s prevajanjem terminologije in s terminološko doslednostjo. V magistrskem delu analiziramo strojne prevode terminov v jezikovnem paru angleščina-slovenščina. Angleške in slovenske članke s področja krasoslovja smo prevedli z nevronskima strojnima prevajalnikoma DeepL in Google Translate, nato pa smo preučili vrste napak in jih razdelili v sedem kategorij. Strojne prevode terminov smo primerjali s prevodnimi ustreznicami v terminološki bazi in z referenčnim prevodom ter izračunali odstotke ujemanj. Največji odstotek smo zabeležili pri ujemanju prevodov Googlovega prevajalnika s termini v terminološki bazi (83 %). Nadaljnja analiza je pokazala, da sta strojna prevajalnika uspešnejša pri prevajanju terminov v slovenščino kot pri prevajanju v angleščino. Razlog za to je, da je mednarodna literatura določene krasoslovne termine prevzela iz slovanskih jezikov, npr. polje, doline, ponor. Ugotavljamo tudi, da Googlov prevajalnik dosega boljšo kakovost prevodov v obeh jezikovnih smereh, prav tako je pri prevajanju doslednejši.

• Language: Slovenian
• Keywords: strojno prevajanje, nevronski prevajalniki, prevajanje terminologije, terminološka doslednost, krasoslovje
• Work type: Master's thesis/paper
• Organization: FF - Faculty of Arts
• Year: 2023
• PID: 20.500.12556/RUL-144398
• COBISS.SI-ID: 144643587
• Publication date in RUL: 19.02.2023

Secondary language

• Language: English
• Title: Analysis of machine translation of terminology in English and Slovenian texts

Abstract

Despite their high quality, neural machine translation systems still face certain problems, including terminology translation and terminology consistency. This thesis analyzes machine translation of terms for the language pair English-Slovene. First, English and Slovene texts in karstology were translated using two NMT systems, DeepL and Google Translate. Incorrect translations of terms were then divided into seven categories. Machine translations of terms were compared with the translation equivalents in the term base and with the reference translation, and the match percentages were calculated. The highest match percentage was observed in matching the Google Translate translations with terms in the term base (83%). Further analysis revealed that the NMT systems achieve higher quality when translating terms into Slovene; the reason for this is that international literature has adopted certain karst terms from Slavic languages; for example, polje, doline, and ponor. More correct term translations were created by Google Translate, which also achieves higher terminology consistency.

• Keywords: machine translation, neural machine translation systems, terminology translation, terminology consistency, karstology