Trillions of Comparable Documents

We propose a novel multilingual Web crawler and sentence mining system to continuously mine and extract parallel sentences from trillions of websites, unconstrained by domain or url structures, or publication dates. The system is divided into three main modules, namely Web crawler, comparable and parallel website matching and parallel sentence extraction. Previous methods in mining parallel sentences from the Web focus on specific websites, such as newspaper agencies, or sites sharing the same URL parents. The output of these previous systems are limited in scope and static in nature. As the Web is boundless and growing, we propose to continuously crawl the Web and update the pool of parallel sentences extracted. One main objective of our work is to improve statistical machine translation systems. Another objective is to take advantage of the heterogeneous website documents to discover parallel sentences in henceforth undiscovered domains and genres, such as user generated content. We investigate a host of recall-oriented vs precisionoriented algorithms for comparable and parallel document matching, as well as parallel sentence extraction. In the future, this system can be extended to mine other monolingual or bilingual linguistic resources from the Web.