Movie and TV subtitles (such as OpenSubtitles in OPUS) are a highly valuable resource for the compilation of parallel corpora thanks to their availability in large numbers and across many languages.
However, the quality of the resulting sentence alignments is often much lower than for other parallel corpora. Low-quality alignments may be due to several factors, such as:
- - The input or target sentence may be poorly written or formatted
- - The alignment may be incorrect due to small timing differences in the display times of the subtitles.
It would be very useful to automatically assess the quality of those subtitles, both by filtering out low-quality subtitles (monolingual filtering) and by prune sentence pairs that are likely to be misaligned. The student will develop such filtering methods and evaluate their performance on downstream tasks such as machine translation.
Prerequisites: Good programming skills and interest in multilingual NLP. We are going to work with very large datasets, so some basic knowledge of high-performance computing is an advantage.