Identification of the parallel documents from multilingual news websites

摘要

We present the initial results of our experiments on document alignment for the online news domain. Specifically, as apposed to cross-site comparable news alignment, we focus on the identification of parallel documents from within the same multilingual websites. In such a setting parallel news stories oftentimes turn out to be direct translations of each other with a tendency of sharing common media and displaying proximity in publication date. We leverage this domain-specific property of the data and propose a straightforward yet competitive heuristic that performs on par with a machine learning-based method in terms of precision, and outperforms a widely used bitext extraction system on a range of metrics. Moreover, this heuristic has allowed us to identify comparable documents overlooked by a human annotator. Although both rule- and learning-based methods that we present are language independent, we specifically focus on the Russian-Kazakh language pair as the present study is one of the initial steps towards a greater objective of building a corresponding parallel corpus and a machine translation system.