Exploring challenges in audiovisual translation: A comparative analysis of human- and AI-generated Arabic subtitles in Birdman

Abstract

Movies often use allusions to add depth, create connections, and enrich the storytelling. However, translators may face challenges when subtitling movie allusions, as they must render both meaning and culture accurately despite existing language and cultural barriers. These challenges could be further complicated by the use of available AI tools attempting to subtitle movie allusions, while probably unaware of existing cultural complexities. This research investigates these challenges using qualitative and descriptive quantitative approaches by analyzing the movie Birdman or (The Unexpected Virtue of Ignorance), comprising13.014 words, to identify the types of allusions used and compare the human- vs. AI (ChatGPT)-generated Arabic subtitles in terms of the subtitling strategies, their frequency, and quality. The results revealed that the movie used 52 Noun Phrase (NP) allusions, where the writer intertextually employed a proper name to convey meaning, and 8 Key-Phrase (KP) allusions, where the writer used phrases that convey implicit meaning easily perceived by members of the source culture (by referring to religious, literary, or entertainment texts). For NP allusions, both the human translator and AI opted for retentive strategies; however, the human translator’s preference to add guidance/parentheses to mark NP allusions was distinct. Additionally, it was observed that AI used neologism to render technology-related allusions, which could be a suggested strategy for NP subtitling into Arabic. For KP allusions, while the human translator seemed to be cognizant of the idea that KP allusions typically require a change in wording, AI fell short. Specifically, the human translator employed reduction in 5 out of 8 KPs, opting for minimum change/literal translation only three times. Conversely, AI utilized literal translation in all 8 examples, despite its awareness of the allusion and its intricate meaning/reference. As for the FAR assessment, for NP allusions, it revealed minor semantic errors in AI’s subtitles that did not affect the plot. Regarding KP allusions, AI’s subtitles were penalized in 5 out of its 8 Arabic renditions, in contrast to the human translator. Most of the errors were serious semantic errors that likely disrupted the flow of reading the subtitles due to conveying irrelevant meanings in the movie’s/scene’s context. Despite its functionality, this study suggests adding an extra parameter to the FAR model: consistency, as it plays a role in enhancing audience involvement and understanding. Its absence, as observed in some AI instances, can be misleading.