翻译规则剪枝与基于半强制解码和变分贝叶斯推理的模型训练  

Translation Rule Pruning and Model Training with Semi-Forced Decoding and Variational Bayesian Inference

作　　者：高恩婷[1] 段湘煜[2] 巢佳媛[2] 张民[2] 

机构地区：[1]苏州科技学院电子与信息工程学院,江苏苏州215011 [2]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《中文信息学报》2014年第5期141-147,共7页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61373095)

摘　　要：统计机器翻译一般采用启发式方法训练翻译模型。但启发式方法的理论基础不够完善,因此,会导致翻译模型规模庞大以及模型参数精确率不高。针对以上两个问题,该文提出一种基于变分贝叶斯推理的模型训练方法,形成更精确的精简翻译模型。该方法首先通过强制解码对齐语料,然后利用变分贝叶斯EM算法获得模型参数。该文的实验语料为NIST汉英翻译任务数据,实验结果显示,基于句法(基于短语)的统计机器翻译中,超过95%(76%)的规则被剪枝,且BLEU值显著提高。SMT usually learns translation models with heuristics, which leads to large models and potentially less accurate model parameters due to the poor theoretical justification of heuristics. This paper presents a variational Bayesian inference-based training method to address these two issues, targeting to learn a compact translation model with more accurate translation probabilities. It is achieved by translation model parameter estimation using variational Bayesian EM over alignments obtained by forced decoding. Experimental results on the Chinese-English NIST translation data shows that our proposed method is very effective, resulting in more than 95% （76%） rule pruned out with significant performance improvement in Bleu score for syntax-based SMT and phrase-based SMT. Key words., machine translation; rule pruning~ semi-forced decoding; variational bayesian

关 键 词：机器翻译 规则剪枝 半强制解码 变分贝叶斯 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]