本体驱动的文本虚拟样本构造方法研究

Research on Ontology-driven Text Virtual Sample Constructing

构造虚拟样本能够为机器学习中的训练集融入先验知识,从而改善标注瓶颈问题。提出了一种本体驱动的文本虚拟样本构造方法。在确保类别不变性的前提下,该方法依据领域相关本体所明晰表达的领域知识,基于本体树的点、边、子树,从同义、父子、语义同构的多个词义关系角度实现了文本虚拟样本的构造。初步实验表明,该方法与原分类及类似方法相比具有更好的分类精度和推广能力。

Constructing virtual examples can incorporate prior knowledge into training set in machine learning, so as to alleviate the labeling bottleneck. An Ontology-driven scheme to construct text virtual sample is proposed. Under the precondition of label invariability, the proposal constructs virtual samples according to the domain knowledge explicitly formalized by domain-specific Ontology. Based on the different Ontology tree structures, namely nodes, edges, and sub-trees, various lexical-semantic relations, including synonymy, paternity, and semantic isomorphs, are applied into text virtual example constructing. The primary experimental results show the scheme outperforms original text catego- rizations and other similar ones in precision and generalization ability.