Bilingual Chunk Alignment In Statistical Machine Translation*

Aiming to overcome the shortcomings of word-based Abstract In [his Paper new algorithm called Mulfi- methods, alignment based on phrases, Layer Filtering (MZF) is proposed for extracting bilingual chunks or structures have been proposed [ S, 6, 7, 81. These translations since chunk-based translations caphxe local parallel corpus. Multiple layers are used to extract bilingual chunks according '0 di@erentfeatures of chunks phrase Or in the bilingual corpus. And the alignment chunks are one- chunk alignment algorithms have been based on complex to-one corresponding with each other. The chunking and syntax information e.g. by incorporating parsing focused on certain special kinds of phrases. These methods tagging, parsing, syntax analyzing or segmenting for have proven to yield poor performance when dealing with Chinese corpus as most conventional atgorifhms do. Preliminary experimental results show that the algorithm long sentences. Further, the methods heavily depend on the achieves a good performance in chunking and alignment. performance of associated tools, such as parsers, part-of- Moreover, the translations generated by this algorithm are speech taggers, word segmenters for Chinese and so much better than the results penerated bv the baseline automaticalb from a Chinese modifications are advantageous and lead to more fluent phenomena, H ~ ~ ~ ~ ~ ~ , until now, algorithm doesn't On lhe infomation from technology with crossing cons&aints or have heen narrowly I (word-based stafisfical machine translalion). Kelwords: Alignment, chunking, multi-layer filtering, statistical machine translation