Impact of Automatic Segmentation on the Quality, Productivity and Self-reported Post-editing Effort of Intralingual Subtitles

Impact of automatic segmentation on the quality, productivity and self-reported post-editing effort of intralingual subtitles Aitor ´Alvarez1, Marina Balenciaga1, Arantza del Pozo1, Haritz Arzelus1, Anna Matamala2, Carlos-D. Mart´ınez-Hinarejos3 1 Human Speech and Language Technology Group, Vicomtech-IK4, San Sebastian, Spain 2 Department of Translation, Interpreting and East Asian Studies, UAB, Barcelona, Spain 3 Pattern Recognition and Human Language Technologies Research Center, Universitat Polit`ecnica de Val`encia, Spain 1{aalvarez, mbalenciaga, adelpozo, harzelus}@vicomtech.org, 2anna.matamala@uab.cat, 3cmartine@dsic.upv.es Abstract This paper describes the evaluation methodology followed to measure the impact of using a machine learning algorithm to automatically segment intralingual subtitles. The segmentation quality, productivity and self-reported post-editing effort achieved with such approach are shown to improve those obtained by the technique based in counting characters, mainly employed for automatic subtitle segmen- tation currently. The corpus used to train and test the proposed automated segmentation method is also described and shared with the community, in order to foster further research in this area. Keywords: automatic subtitling, subtitle segmentation, machine learning 1. Introduction Society and the governments are increasingly requesting larger amounts of subtitled TV content (Neto et al., 2008), since subtitles are the most practical technique to guarantee the accessibility of audiovisual material to those who can- not access the audio (AENOR, 2012). The current demand and promising future of intralingual subtitling has accelerated research into more productive methods that help to cover challenging subtitling situations such as, for example, live broadcasts. In recent years, tech- nological advances in speech recognition have enabled au- tomatic intralingual subtitling to be a reality ( ´Alvarez et al., 2015). However, automatic subtitling technology has lim- itations, a major one being its inability to segment subtitle text in a logical way. Within the subtitling field, segmenta- tion refers to the division of the original text into sections that viewers can understand immediately (D´ıaz-Cintas and Remael, 2007), playing a fundamental role in the creation of quality subtitles. This work analyses the application of a machine learning algorithm to automatically segment the text contained in intralingual subtitles and compares its performance against that of the main technique based in counting characters em- ployed in most automatic subtitling systems currently. Its impact is measured in terms of subtitle quality and regard- ing the productivity achieved and self-reported effort when integrated in the subtitling process through post-editing. Also, given the little work carried out so far in automatic subtitle segmentation, the corpus employed to train and test the presented machine learning algorithm is described and shared with the aim of fostering further research and tech- nology development. 2. Background 2.1. From traditional to automatic subtitling Traditional subtitling is carried out by professionals who aim to reproduce in text on screen the original dialogues, the discursive elements in the image and, when addressed to those who cannot hear the original audio, the informa- tion contained in the soundtrack of audiovisual contents (D´ıaz-Cintas and Remael, 2007). From a linguistic perspective, subtitles can be classified as intralingual, interlingual or bilingual. Depending on the time available for their preparation, …

Content truncated to 500 words