Targeted paraphrasing is a new approach to the problem of obtaining cost-effective, reasonable quality translation that makes use of simple and inexpensive human computations by monolingual speakers in combination with machine translation. The key insight behind the process is that it is possible to spot likely translation errors with only monolingual knowledge of the target language, and it is possible to generate alternative ways to say the same thing (i.e. paraphrases) with only monolingual knowledge of the source language. Evaluations demonstrate that this approach can yield substantial improvements in translation quality. 1

To increase the model coverage, sourcelanguage paraphrases have been utilized to boost SMT system performance. Previous work showed that word lattices constructed from paraphrases are able to reduce out-ofvocabulary words and to express inputs in different ways for better translation quality. However, such a word-lattice-based method suffers from two problems: 1) path duplications in word lattices decrease the capacities for potential paraphrases; 2) lattice decoding in SMT dramatically increases the search space and results in poor time efficiency. Therefore, in this paper, we adopt word confusion networks as the input structure to carry source-language paraphrase information. Similar to previous work, we use word lattices to build word confusion networks for merging of duplicated paths and faster decoding. Experiments are carried out on small-, medium- and large-scale English– Chinese translation tasks, and we show that compared with the word-lattice-based method, the decoding time on three tasks is reduced significantly (up to 79%) while comparable translation quality is obtained on the largescale task. 1