基于Deep Belief Nets的中文名实体关系抽取

摘要

Relation extraction js a fundamental task in information extraction, which is to identify the semantic relationships between two entities in the text. In this paper, deep belief nets (DBN), which is a classifier of a combination of several unsupervised learning networks, named RBM (restricted Boltzmann machine) and a supervised learning network named BP (back-propagation), is presented to detect and classify the relationships among Chinese name entities. The RBM layers maintain as much information as possible when feature vectors are transferred to next layer. The BP layer is trained to classify the features generated by the last RBM layer. The experiments are conducted on the Automatic Content Extraction 2004 dataset. This paper proves that a character-based feature is more suitable for Chinese relation extraction than a word-based feature. In addition, the paper also performs a set of experiments to assess the Chinese relation extraction on different assumptions of an entity categorization feature. These experiments showed the comparison among models with correct entity types and imperfect entity type classified by DBN and without entity type. The results show that DBN is a successful approach in the high-dimensional-feature-space information extraction task. It outperforms state-of-the-art learning models such as SVM and back-propagation networks.%关系抽取是信息抽取的一项子任务,用以识别文本中实体之间的语义关系.提出一种利用DBN(deep belief nets)模型进行基于特征的实体关系抽取方法,该模型是由多层无监督的RBM(restricted Boltzmann machine)网络和一层有监督的BP(back-propagation)网络组成的神经网络分类器.RBM网络以确保特征向量映射达到最优,最后一层BP网络分类RBM网络的输出特征向量,从而训练实体关系分类器.在ACE04语料上进行的相关测试,一方面证明了字特征比词特征更适用于中文关系抽取任务 ;另一方面设计了3组不同的实验,分别使用正确的实体类别信息、通过实体类型分类器得到实体类型信息和不使用实体类型信息,用以比较实体类型信息对关系抽取效果的影响.实验结果表明,DBN非常适用于基于高维空间特征的信息抽取任务,获得的效果比SVM和反向传播网络更好.