中文名实体识别中的特征组合与特征融合的比较

Comparing features combination with features fusion in Chinese named entity recognition

先分析了最大熵模型常用的特征线性组合方法中的权值偏置问题,然后提出了在线性组合之前,对特征进行融合,并根据融合特征和目标类别之间的互信息选择有效复合特征的方法。通过在包含2000个人名的语料库上的测试,表明特征融合能有效地提高名实体识别的精度和召回率。

Maximum entropy model is usually used for named entity recognition, in which the features related to a random event are linearly combined. The problem of the weight bias in the features combination was pointed out, and a strategy of performing features fusion before linearly combining was proposed. The result of experiment on corpus containing 2000 human names shows that features fusion can improve the precision and recall of named entity recognition effectively.