Recycling texts: human evaluation of example-based machine translation subtitles for DVD

Abstract

This project focuses on translation reusability in audiovisual contexts. Specifically, the project seeks to establish (1) whether target language subtitles produced by an EBMT system are considered intelligible and acceptable by viewers of movies on DVD, and (2)whether a relationship exists between the ‘profiles’ of corpora used to train an EBMT system, on the one hand, and viewers’ judgements of the intelligibility and acceptability of the subtitles produced by the system, on the other. The impact of other factors, namely: whether movie-viewing subjects have knowledge of the soundtrack language; subjects’ linguistic background; and subjects’ prior knowledge of the (Harry Potter) movie clips viewed; is also investigated. Corpus profiling is based on measurements (partly using corpus-analysis tools) of three characteristics of the corpora used to train the EBMT system: the number of source language repetitions they contain; the size of the corpus; and the homogeneity of the corpus (independent variables). As a quality control measure in this prospective profiling phase, we also elicit human judgements (through a combined questionnaire and interview) on the quality of the corpus data and on the reusability in new contexts of the TL subtitles. The intelligibility and acceptability of EBMT-produced subtitles (dependent variables) are, in turn, established through end-user evaluation sessions. In these sessions 44 native German-speaking subjects view short movie clips containing EBMT-generated German subtitles, and following each clip answer questions (again, through a combined questionnaire and interview) relating to the quality characteristics mentioned above. The findings of the study suggest that an increase in corpus size along with a concomitant increase in the number of source language repetitions and a decrease in corpus homogeneity, improves the readability of the EBMT-generated subtitles. It does not, however, have a significant effect on the comprehensibility, style or wellformedness of the EBMT-generated subtitles. Increasing corpus size and SL repetitions also results in a higher number of alternative TL translations in the corpus that are deemed acceptable by evaluators in the corpus profiling phase. The research also finds that subjects are more critical of subtitles when they do not understand the soundtrack language, while subjects’ linguistic background does not have a significant effect on their judgements of the quality of EBMT-generated subtitles. Prior knowledge of the Harry Potter genre, on the other hand, appears to have an effect on how viewing subjects rate the severity of observed errors in the subtitles, and on how they rate the style of subtitles, although this effect is training corpus-dependent. The introduction of repeated subtitles did not reduce the intelligibility or acceptability of the subtitles. Overall, the findings indicate that the subtitles deemed the most acceptable when evaluated in a non-AVT environment (albeit one in which rich contextual information was available) were the same as the subtitles deemed the most acceptable in an AVT environment, although richer data were gathered from the AVT environment.