语言建模中最小化样本风险算法的研究和改进

A Study and Improvement of Minimum Sample Risk Methods for Language Modeling

目前,一些主流的判别学习算法只能优化光滑可导的损失函数,但在自然语言处理(natural language processing,简称NLP)中,很多应用的直接评价标准(如字符转换错误数(character error rate,简称CER))都是不可导的阶梯形函数.为解决此问题,研究了一种新提出的判别学习算法——最小化样本风险(minimum sample risk,简称MSR)算法.与其他判别训练算法不同,MSR算法直接使用阶梯形函数作为其损失函数.首先,对MSR算法的时空复杂性作了分析和提高;同时,提出了改进的算法MSR-II,使得特征之间相关性的计算更加稳定.此外,还通过大量领域适应性建模实验来考察MSR-II的鲁棒性.日文汉字输入实验的评测结果表明:(1)MSR/MSR-II显著优于传统三元模型,使错误率下降了20.9%;(2)MSR/MSR-II与另两类主流判别学习算法Boosting和Perceptron表现相当;(3)MSR-II不仅在时空复杂度上优于MSR,特征选择的稳定性也更高;(4)领域适应性建模的结果证明了MSR-II的良好鲁棒性.总之,MSR/MSR-II是一种非常有效的算法.由于其使用的是阶梯形的损失函数,因此可以广泛应用于自然语言处理的各个领域,如拼写校正和机器翻译.

Most existing discriminative training methods adopt smooth loss functions that could be optimized directly. In natural language processing （NLP）, however, many applications adopt evaluation metrics taking a form as a step function, such as character error rate （CER）. To address the problem, a newly-proposed discriminative training method is analyzed, which is called minimum sample risk （MSR）. Unlike other discriminative methods, MSR directly takes a step function as its loss function. MSR is firstly analyzed and improved in time/space complexity. Then an improved version MSR-Ⅱ is proposed, which makes the computation of interference in the step of feature selection more stable. In addition, experiments on domain adaptation are conducted to investigate the robustness of MSR-Ⅱ. Evaluations on the task of Japanese text input show that： （1） MSR/MSR-Ⅱ significantly outperforms a traditional trigram model, reducing CER by 20.9%; （2） MSR/MSR-Ⅱ is comparable to the other two state-of-the-art discriminative algorithms, Boosting and Perceptron; （3） MSR-Ⅱ outperforms MSR not only in time/space complexity but also in the stability of feature selection; （4） Experimental results of domain adaptation show the robustness of MSR-Ⅱ. In all, MSR/MSR-Ⅱ is a quite effective algorithm. Given its step loss function, MSR/MSR-Ⅱ could be widely applied to many fields of NLE such as spelling check and machine translation.