针对大边距估计(Large Margin Estimation,LME)准则仅选取支持集内的最小边距进行调整导致边距利用不合理的问题,本文提出一种大边距准则目标函数的改进形式,通过增强竞争假设中与正确标注竞争关系较强的路径的似然得分,使训练数据的分类边距在一定程度上变小,从而进一步提高大边距估计的训练效果。并在此基础上,提出一种新的逼近优化方法,即当某点目标函数与辅助函数梯度方向相同时,在该点邻近的一定范围内,优化辅助函数即可带来目标函数相应的优化。在微软语料库上的实验成功证明了本文算法的有效性。
Aiming at the problem that unreasonable use of margin caused by Large Margin Estimation (LME) only chooses the minimum margin in the support set to adjust, we present a modified form of the LME objective function which gives improved results for discriminative training. The modification consists of boosting the likelihoods of paths in the competing hypothesis that have a higher phone error relative to the correct transcript. So that, the minimum margin of all training data becomes smaller. What's more, we present a new approximate optimization method. When auxiliary function's and objective function's gradient direction phase at the same time, in the close neighborhood, the optimal solution found for the approximate auxiliary function is expected to improve the original objective function as well. The recognition results on the Microsoft corpora can prove our idea.