基于支持向量机的英语名词短语指代消解

Anaphora Resolution of Noun Phrase Based on SVM

提出一种基于支持向量机(SVM)的英语名词短语的指代消解方法,并给出具体实现系统。实验采用了几个常用的基本特征,在MUC-6公开语料上测试得到的F值为68.6,优于同类型的其他原型系统。分析SVM中不同核函数对分类结果的影响以及不同的特征对指代消解的作用。实验结果表明,同位语、别名和字符串匹配3个特征对指代消解非常重要,距离作为特征使用时对指代消解没有帮助,但可在训练样例生成时作为限制条件来使用。

This paper proposes an anaphora resolution of noun phrases based on Support Vector Machine（SVM）. Evaluation on the MUC-6 corpus using several widely used features shows that the system achieves the F-measure of 68.6% and outperforms other similar systems. Further analysis shows that appositive, name alias and full string matching contributes most for anaphora resolution. It also shows that the distance between the antecedent candidate and the anaphor is very useful in constraining the instance generation, although including it as a feature does not help for anaphora resolution.