基于半监督CRF的缩略词扩展解释识别  

Abbreviation Expansion Interpretation Recognition Based on Semi-supervised CRF

作　　者：陈季梦[1] 刘杰[1] 黄亚楼[1,2] 刘天笔[2] 刘才华[1] 

机构地区：[1]南开大学信息技术科学学院,天津300071 [2]南开大学软件学院,天津300071

出　　处：《计算机工程》2013年第4期203-209,共7页Computer Engineering

基　　金：国家自然科学基金资助项目(61105049);高等学校博士学科点专项科研基金资助项目(20100031110096);中央高校基本科研业务费专项基金资助项目(65010571)

摘　　要：缩略词拓展解释识别任务中标注样本较少,无法从中总结出全面的规则或采用有监督的学习方法来学习。为此,提出一种基于半监督条件随机场(CRF)的缩略词扩展解释识别模型,利用广泛的未标注样本和较少的标注样本寻找序列文本中恰当的语句,以解释给定的缩略词。使用较少的标注序列样本训练一个全监督CRF模型,针对未标注序列样本,采用最小序列熵学习样本之间的联系,结合标注样本和未标注样本,利用半监督自学习方法学习两者的关系。实验结果表明,该模型的序列F1值达到84.73%,高于支持向量机和全监督CRF基准算法。Because lack of labeled samples in acronym expansion recognition task, rule-based or supervised methods are unsuitable to find a proper phrase in sequential texts to explain a given acronym. A semi-supervised Conditional Random Field（CRF） is designed to make use of sufficient unlabeled samples and a few labeled samples. A supervised CRF model by a few labeled sequential samples is trained. Minimum sequential entropy is employed to learn relationship between samples without labels. A self training model is applied to analyzes the relationship between labeled and unlabeled samples. Experimental results show this model achieves the best performance of sequential F1 （84.73%） against the state-of-the-art baselines including Support Vector Machine（SVM） and supervised CRF.

关 键 词：扩展解释 半监督 条件随机场 序列熵 序列标注 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]