融合词嵌入表示特征的实体关系抽取方法研究

发布时间：2018-05-31 16:48

本文选题：关系抽取 + 词嵌入表示　；参考：《数据分析与知识发现》2017年09期

【摘要】：【目的】为解决已有方法中单词特征表示不具有语义信息这一问题,对词嵌入表示特征在关系抽取中的作用进行探讨。【方法】考虑词嵌入表示级别、词汇级别和语法级别三种类型特征,利用朴素贝叶斯模型、决策树模型和随机森林模型进行对比实验,并选出代表全部特征的有效特征子集。【结果】使用全部特征时,决策树算法的准确率达到0.48,关系抽取效果最佳,Member-Collection(E_2,E_1)类型关系的F_1值达到0.70,特征排序结果表明依存关系有助于关系抽取。【局限】对小样本量和情况复杂的关系类型识别效果有待提高,以及词向量训练及方法的相关参数需要进一步优化。【结论】实验证明选取的三种类型特征的有效性,词嵌入表示级别特征在实体关系抽取问题中可以发挥重要作用。
[Abstract]:[objective] in order to solve the problem that word feature representation does not have semantic information in existing methods, the function of word embedding representation feature in relation extraction is discussed. [methods] the level of word embedding representation is considered. Three types of features, lexical level and grammatical level, are compared by using naive Bayes model, decision tree model and random forest model, and the effective feature subsets representing all the features are selected. [results] when using all the features, The accuracy of decision tree algorithm is 0.48, and the F-1 value of relationship extraction is 0.70. The result of feature ranking shows that dependency relationship is helpful for relation extraction. [limitation] for small sample size and complicated relation type, the FSP _ 1 value of the relation is 0.70, and the result of feature ranking shows that the dependency relation is helpful to the relation extraction. The recognition effect needs to be improved. The training of word vector and the related parameters of the method need to be further optimized. [conclusion] experiments show that the selected three types of features are effective, and the level feature of word embedding representation can play an important role in the problem of entity relation extraction.
【作者单位】：中国科学院文献情报中心;中国科学院大学;中国科学院武汉文献情报中心;
【基金】：ISTIC-EBSCO文献大数据发现服务联合实验室基金项目“基于clique子团聚类的文本主题识别方法研究”的研究成果之一
【分类号】：TP391.1

【相似文献】