Abstract
In this paper, we introduce a method to align an English multiword expression (MWE) with Chinese translation equivalent (TE) in a given bilingual parallel corpus. In our approach, we make use of an existing word alignment tool which provides information about word alignment results to estimate context-independent lexical translation probability (LTP). However, such estimates for lexical translation probability often have unsatisfactory precision and recall rates due to the inherent limitation of word alignment. Consider context and morphological information may ease the problem. More specifically, words with related meanings usually have some characters in common. Therefore, we build on word-alignment results to estimate context-sensitive LTP at the morphological level for further alignment. At runtime, we align each word in an English MWE individually with Chinese words in the target sentence according to this new version of LTP, and combine the alignments into the final TE. We implement the method on verbal MWEs randomly selected from WordNet and several MWEs manually examined from corpus. The evaluation of the experimental results shows our procedure outperforms the underlying word alignment tool. Our methodology helps to establish more precise translation memory, which may improve the performance of machine translation systems.