Error correction in lightly supervised alignment of broadcast subtitles

This paper presents a range of error correction techniques aimed at improving the accuracy of a lightly supervised alignment task for broadcast subtitles. Lightly supervised approaches are frequently used in the multimedia domain, either for subtitling purposes or for providing a more reliable source for training speech–based systems. The proposed methods focus on directly correcting of the alignment output using different techniques to infer word insertions and words with inaccurate time boundaries. The features used by the classification models are the outputs from the alignment system, such as confidence measures, and word or segment duration. Experiments in this paper are based on broadcast material provided by the BBC to the Multi–Genre Broadcast (MGB) challenge participants. Results, show that the order alignment F–measure improves up to 2.6% absolute (15.8% relative) when combining insertion and word– boundary correction