用决策树指导TBL进行多音字消歧

Polyphone disambiguation based on tree-guided TBL

多音字消歧是普通话语音合成系统中字音转换模块的核心问题。选择了常见易错的33个多音字和24个多音词作为研究对象,构建了一个平均每个多音字(词)5000句的语料库,并且提出了一种结合决策树和基于转换的错误驱动的学习(Transformation-Basederror-driven Learning,TBL)的混合算法。该方法根据决策树的指导,自动生成TBL算法的模板,避免了手工总结模板这一费时费力的过程。实验结果表明,该方法生成的模板与手工模板性能相当,其平均准确率达90.36%,明显优于决策树。

Polyphone disambiguation is the core issue of the grapheme-to-phoneme conversion in Mandarin Text-To-Speech （＇ITS） system.This paper selects 33 key polyphones and 24 key polyphonic words which are most ambiguous and frequently used as study objects,and builds a polyphone corpus of 5 000 sentences per polyphone on average.Furthermore,a hybrid algorithm called Tree-Guided Transformation-Based Leaming（TGTBL）,which combines decision tree with Transformation-Based error-driven Leaming（TBL）,is proposed to resolve the polyphonic ambiguity.It automatically generates TBL templates,thereby avoiding manually summarizing templates, which is time-consuming and laborious in conventional TBL.Results of comparative experiments show that, for the task of polyphone disambiguation, templates automatically generated by decision tree achieve comparable performance to manually summarized templates,and the average precision of TGTBL reaches 90.36%,siguificantly higher than that of decision tree.