.png){kind=icon}
返回
Large-margin minimum classification error training: A theoretical risk minimization perspective
.png){kind=icon}
PRE
AI
DOI：10.1016/j.csl.2008.03.002.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Large-margin discriminative training of hidden Markov models has received significant attention recently, A natural and interesting question is whether the existing discriminative training algorithms can be extended directly to embed the concept of margin. Ill this paper, we give this question all affirmative answer by showing that the sigmoid bias ill the conventional minimum classification error (MCE) training call be interpreted as a soft margin. We justify this claim from a theoretical classification risk minimization perspective where the loss function associated with a non-zero sigmoid bias is shown to include not only empirical error rates but also it margin-bound risk. Based oil this perspective, we propose a practical optimization strategy that adjusts the margin (sigmoid bias) incrementally in the NICE training process so that a desirable balance between the empirical error rates oil the training set and the margin call be achieved. We call this modified NICE training process large-margin minimum classification error (LM-MCE) training to differentiate it from the conventional MCE. Speech recognition experiments have been carried out oil two tasks. First, ill the TIDIGITS recognition task, LM-MCE outperforms the state-of-the-art NICE method with 17% relative digit-error reduction and 19%, relative string-error reduction. Second, oil the Microsoft internal large vocabulary telephony speech recognition task (with 2000 h of training data and 120 K words ill the vocabulary), significant recognition accuracy improvement is achieved, demonstrating that our formulation of LM-MCE call be successfully scaled up and applied to large-scale speech recognition tasks. (c) 2008 Elsevier Ltd. All rights reserved.
Keyword:
Minimum classification error training
Discriminative training
Large-margin training
Speech recognition
Large-scale speech recognition
Theoretical classification risk minimization
AI总结
对已上传原文的论文进行重点信息的提取，主要内容包括：简要概述、研究摘要、背景介绍、关键亮点、图文解析、展望与总结。