Corpus-Driven Splitting of Compound Words

A method is presented for splitting compound words into their constituents based on cognate words in the other language of a parallel corpus. A minor extension to the method using a bilingual lexicon (which may be statistically derived from the corpus) allows the decompounding of words that do not have cognates in the other language. Further, the algorithm can produce, as a by-product, a mapping from compound words in one language to phrases in the other language. The method described in this paper is applied to an example-based machine translation (EBMT) by decompounding the training corpus, and training both the EBMT system and the bilingual lexicon it uses for subsentential alignment from the decompounded corpous. Compared with the original corpus, the decompounded corpus substantially reduces the incidence of word-alignment failure, resulting in a modest overall improvement in performance.