Abstract: The performance of statistical machine translation (SMT) suffers from the insufficiency of parallel corpus. To solve the problem, the authors propose a paraphrase based SMT framework with three solutions: 1) acquiring paraphrase knowledge based on a third language; 2) expressing multiple paraphrases of input sentence in a lattice and modifying decoder to be able to process it; 3) integrating paraphrase knowledge as features into log-linear model. In this way, not only more expressions in source language can be covered, but also more expressions in target language can be generated as candidate translations. To verify proposed method, experiments are conducted on three training data sets with different sizes, and evaluate the improvement of the performance of SMT system contributed by paraphrasing. Experimental results show that the translation performance is improved significantly (BLEU+1.4%) when the parallel corpus is small (10 K), and a good performance (BLEU+0.32%) is also achieved when parallel corpus is large enough (1 M).

Key words: paraphrase, phrase translation table, features, decoder

摘要： 为了缓解双语语料不足导致的翻译知识欠缺问题, 提出基于复述技术的翻译框架。此框架利用第三种语言获取带有概率的复述知识表, 以Lattice表示输入句子的多种复述形式, 扩展解码器使之可以对Lattice形式的输入进行解码, 将复述知识作为特征加入到对数线性模型的目标函数中。在保持原始翻译知识表不变的情况下, 此框架不仅可以增大短语翻译表对源语言现象的覆盖率, 也能够增加候选译文表现形式的多样性。在3个不同规模训练集上的对比实验结果表明, 在训练语料规模最小的情况下(10 K句对), 系统性能有明显提升(BLEU+1.4%); 在训练语料规模最大的情况下(1 M句对), 系统性能也取得一定提升(BLEU+ 0.32%)。

关键词: 复述知识, 短语翻译表, 特征, 解码器