基于立方剪枝的短语机器解码算法改进  

Improvement to the Decoding Algorithm of Phrase-Based Statistical Machine Translation Based on Cube Pruning

作　　者：朱海[1,2] 李淼[1] 乌达巴拉[1] 张建[1] 

机构地区：[1]中国科学院合肥智能机械研究所,安徽合肥230031 [2]中国科学技术大学信息科学技术学院,安徽合肥230027

出　　处：《微电子学与计算机》2011年第2期153-156,共4页Microelectronics & Computer

基　　金：中国科学院知识创新工程重要方向项目(KGCX2-SW-511)

摘　　要：柱搜索算法是短语统计机器翻译广泛使用的解码算法.文中在立方剪枝算法的基础上提出了另一种对柱搜索算法的改进算法-基于立方剪枝的逆向递归算法.柱搜索算法对栈中所有的假进行扩展,立方剪枝算法有选择地扩展栈中的前k个最好假设,而基于立方剪枝的逆向递归算法只扩展栈中评分高的假设.实验结果表明获得相同的翻译质量,立方剪枝算法比柱搜索算法快10倍,立方剪枝的改进算法在大规模语料中比立方剪枝算法更节省时间.Beam search is the popular decoding algorithm for phrase-based statistical machine translation systems.This paper describes another improved algorithm to beam search algorithm based on cube pruning algorithm-reverse recursive algorithm based on cube pruning.Beam search expands all the possible hypothesis on the stack.Cube pruning only expands the top k best hypothesis and reverse recursive algorithm based on cube pruning just expands the hypothesis scoring high.The results show that cube pruning algorithm use less time to achieve the same translation quality and reverse recursive algorithm based on cube pruning is quicker than cube pruning without affecting the translation quality much.

关 键 词：基于短语统计机器翻译系统 柱搜索算法 立方剪枝 逆向递归 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]