英语口语重复修正检错中语法网络和搜索过滤算法  

Repair filtering and grammar network for automatic detection of lack of fluency

作　　者：黄申[1] 李宏言[1] 王士进[1] 徐波[1] 

机构地区：[1]中国科学院自动化所研究所,数字内容技术研究中心,北京100190

出　　处：《清华大学学报（自然科学版）》2011年第9期1282-1287,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金重点项目(90820303)

摘　　要：针对口语重复修正检错这一计算机辅助语言教学中的难点,提出一种基于音节单元WFST(weighted finite-state transducer)网络的容错对齐和搜索过滤算法。该算法将对齐后识别结果中邻近匹配词所对应的脚本建立上述语法网络进行二次识别下的容错对齐,得到的候选被修改部分和替换部分作为搜索过滤的查询和模板。最终,重复修正检错结果由搜索过滤算法的置信度决定。为此,提出了基于顺序假设的k-difference算法和基于随机假设的n-gram算法。实验表明:不使用二次容错对齐时,以音节为建模单元的多n-gram混合搜索过滤取得了相对最优的结果;使用二次容错对齐时,能够使F-measure获得3～4个百分比的进一步提升。The detection of the lack of speech fluency is a difficult task in computer aided language learning.This paper presents a second-pass recognition grammar based on weighted finite-state transducer（WFST） and repair filtering.The system first performs a fine-grained re-match on successive matched words in the recognized speech,generating candidates for repair.Then k-difference and n-gram repair filtering is used to filter out false positive errors.The final decision is controlled by the confidence score for the repair filtering.The results show that with single-pass recognition,multi-gram based syllable repair filtering gives the best performance and introducing second pass re-match via the syllable WFST brings about another 3%~4% increase in the F-measure.

关 键 词：计算机辅助语言教学 语法网络 流利度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]