论文部分内容阅读
针对基于汉语词的 Ｎｇｒａｍ 模型统计数据稀疏问题和应用域变化造成原统计模型识别性能降低，提出具有应用域适应能力的 Ｎｇｒａｍ 模型平滑算法。对两种应用域的语料进行了前、后向 ０ 到３ 元文法统计，采用隐马尔可夫模型（ Ｈ Ｍ Ｍ）在语音识别中的成功经验，由 Ｂａｕｍ ｗ ｅｌｃｈ 算法来获得优化权值，每个权值代表相关模型的统计可靠性。由前后向的３ｇｒａｍ 模型可得到５ｇｒａｍ 文法约束的平滑算法，以弥补统计矩阵数据的稀疏现象。将《人民日报》语料的统计结果作为先验统计结果，和《计算机世界》作为转换域的专业语料进行后继训练，得到一种适应应用域的３ｇｒａｍ 模型。实验结果表明，前后向约束的３ｇｒａｍ 文法得到的５ｇｒａｍ 平滑可以较小的存储代价得到较高的文法约束，大大降低了统计模型的困惑度
Aiming at the sparseness of N-gram model based on Chinese words and the change of application domain, the recognition performance of the original statistical model is reduced. An N-gram model smoothing algorithm with adaptability in application domain is proposed. The corpus of the two kinds of application domains were analyzed before and after the 0 to 3 grammes of grammar, the successful experience of HMM in speech recognition was adopted, and the Baum elch algorithm was used to obtain the optimal weights , Each weight represents the statistical reliability of the relevant model. The smoothing algorithm of 5-gram grammar constraints can be obtained from the forward-backward 3-gram model to make up for the sparseness of statistical matrix data. The statistical results of the “People’s Daily” corpus as a priori statistical results, and the “Computer World” as a conversion domain of professional corpus for subsequent training, to get a 3ggram model adapted to the application domain. The experimental results show that the 5-gram smoothing obtained by the 3-gram grammar with the backward and backward constraint can obtain a higher grammar constraint with less storage cost and greatly reduce the confusion of the statistical model