Sentence Alignment in Parallel, Comparable, and Quasi-comparable Corpora

We explore the usability of different bilingual corpora for the purpose of multilingual and cross-lingual natural language processing. The usability of bilingual corpus is evaluated by the lexical alignment score calculated for the bi-lexicon pair distributed in the aligned bilingual sentence pairs. We compare and contrast a number of bilingual corpora, ranging from parallel, to comparable, and to nonparallel corpora. We compare different methods of mining parallel sentences and bilingual lexicon from bilingual corpora. These methods make several sentence-level assumptions on the bilingual corpora. We have found that some of them are applicable to bilingual parallel documents but non-applicable to non-parallel, comparable documents. None of the sentence-level assumptions can be made about nonparallel and quasi-comparable corpora. The latter contain bilingual documents that may or may not be on the same topic. By postulating additional assumptions on comparable documents, we propose a completely unsupervised method to extract useful material, such as parallel sentences and bilexicons, from quasi-comparable corpora. The lexical alignment score for the comparable sentences extracted with our unsupervised method is found to be very close to that of the parallel corpus. This shows that our extraction method is effective.