基于近似随机测试的语义关系抽取比较

Comparison on Semantic Relation Extraction Based on Approximation Random Test

为比较结构化信息和句法分析器对树核函数的关系抽取的作用,提出一种基于近似随机测试语义关系比较方法。对于2种不同配置关系的抽取结果,采用随机标号互换的方法重复产生样本,通过计算这些样本的性能差异进行显著性分析。实验结果表明,动态关系树是最佳的结构化信息,句法分析器Charniak和Berkeley性能均优于Stanford。

To scientifically compare the effect of structured information and parsers on kernel-based relation extraction, a comparison method based on random approximate test is proposed. It gives two relation extraction results for different settings, samples are produced repeatedly using random label exchange from re-sampling techniques, and significant tests are conducted by calculating the performance differences between these samples. Experimental results show that dynamic relation tree is the best structured information, and the performance of Charniak and Berkeley are better than Stanford.