Title Dirbtinio intelekto priemonių taikymas verčiant žodžiu /
Translation of Title Application of artificial intelligence tools in interpreting.
Authors Martin Marin, Lauryna
Full Text Download
Pages 112
Abstract [eng] This paper seeks to evaluate how artificial intelligence-based machine interpreting systems translate European Parliament speeches on various topics. The aim is to determine which delivers higher quality interpretations from English into Lithuanian: the artificial intelligence system ChatGPT or professional interpreters. The study analyses original speeches in English, their official simultaneous interpretations into Lithuanian performed by professional interpreters, and interpretations of the English speeches into Lithuanian generated by the artificial intelligence system ChatGPT, simulating simultaneous interpretation. The dataset consists of five European Parliament speeches, ranging from 5 minutes 9 seconds to 8 minutes 51 seconds in length, with speech rates between 112 and 168 words per minute. Interpretation quality was assessed using a mixed human evaluation framework, incorporating the EU institutions’ interpreter accreditation criteria and the MQM error typology. The results indicate that professional interpreters produced higher quality interpretations. A total of 95 errors were identified in the professionals’ interpretations, compared to 106 in those produced by the artificial intelligence system ChatGPT. Further analysis of the errors based on their impact on meaning revealed that approximately 24% of the errors in professionals’ interpretations affected the conveyed meaning, whereas this figure rose to approximately 47% in ChatGPT’s interpretations. These findings suggest that professional interpreters more effectively convey the intended meaning of the original speeches. The results also showed that from all the errors in ChatGPT’s interpretations the most were omissions of information (the whole segment) and literal translation errors, while professional interpreters’ most errors were of omitted elements and self-corrections. Both types of interpretation had a similar number of mistranslation errors. Terminology errors were assessed using IATE database, the Lithuanian Encyclopaedia website and the EUR-Lex portal. The analysis revealed that ChatGPT’s interpretations contained no terminology errors, while the professional interpretations contained just one. Additionally, ChatGPT’ interpretations had fewer addition errors than those of professional interpreters. However, professional interpreters demonstrated better results in avoiding literal translations, resulting in more fluent and natural-sounding interpretations. Despite the challenges highlighted by the analysis of the interpretations, the findings suggest that artificial intelligence-based machine interpreting systems show promise for achieving adequate interpretation quality in the future. Overall, the study highlights the importance of continued research and technological development in the field of machine interpretating.
Dissertation Institution Vilniaus universitetas.
Type Master thesis
Language Lithuanian
Publication date 2025