摘要
词性标注主要面临兼类词消歧以及未知词标注的难题,传统隐马尔科夫方法不易融合新特征,而最大熵马尔科夫模型存在标注偏置等问题。论文引入条件随机域建立词性标注模型,易于融合新的特征,并能解决标注偏置的问题。此外,又引入长距离特征有效地标注复杂兼类词,以及应用后缀词与命名实体识别等方法提高未知词的标注精度。在条件随机域模型框架下,进一步探讨了融合模型的方法及性能。词性标注开放实验表明,条件随机域模型获得了96.10%的标注精度。
The main difficulties in POS tagging are multi-class word disambiguation and unknown word tagging. However,more features cannot be added into Hidden Markov Model,and there is label bias problem in Maximum Entropy Markov Model.So Conditional Random Field(CRF) is introduced to build POS tagging model in this paper,in order to overcome above problems.In addition,long distance features are extracted and utilized to label complicated multi-class word.As for the unknown word tagging,named entities recognition and suffix-based method etc. are adopted to improve the POS tagging performance.Moreover,we explore the mixing models' performance based on CRF.The experiment indicates our model can achieve a good performance with 96.10% tagging precision.
出处
《计算机工程与应用》
CSCD
北大核心
2006年第21期13-16,42,共5页
Computer Engineering and Applications
基金
国家自然科学基金重点资助项目(编号:60435020)
国家自然科学基金资助项目(编号:60504021)
关键词
词性标注
条件随机域
触发对
POS tagging, Conditional Random Fields,trigger