【作者】 刘克彬； 李芳； 刘磊； 韩颖；

【Author】 Kebin Liu, Fang Li, Lei Liu, Ying Han (Dept. of Computer Science & Engineering, Shanghai Jiaotong University, Shanghai 200240)

【机构】 上海交通大学计算机科学与工程系；

【摘要】 词序列核函数是Convolution核的一种,它处理的对象是离散的词序列。词序列核函数的计算过程不需要显式地构造特征向量,而且具有良好的复合特性,在自然语言处理领域有广泛的应用。传统的词序列核函数没有考虑到语义信息,本文对传统的词序列核函数加入了特征选择和语义扩展,得到一种新的核函数。本文使用该新函数进行实体关系自动抽取的实验, 并与基于特征向量的方法和传统的词序列核函数作了比较。结果证明新函数的效果好于传统的方法,特别是在小规模训练集合上面具有较大的优势。更多还原

【Abstract】 Word-Sequence Kernel is one kind of the Convolution kernels; it works on the discrete word sequences. Word-Sequence kernel has many fine properties such as the combination property and has been widely used in natural language processing field. But traditional Word-Sequence kernel failed to consider the semantic information. This paper proposed a new kernel function which embeds the semantic information and feature selection into kernel function calculation. The experiments of relation extraction were carried out to compare the new kernel function with the feature-based approach and traditional Word-Sequence kernel. The conclusion is that, the new kernel function has better generalization ability and gains a better performance than other two traditional methods especially on a small training set.更多还原