Distant Supervision Relation Extraction via adaptive dependency-path and additional knowledge graph supervision. (February 2021)

• Record Type: Journal Article
• Authors: Shi, Yong; Xiao, Yang; Quan, Pei; Lei, MingLong; Niu, Lingfeng
• Is Part Of: Neural networks. Volume 134(2021)
• Page Start: 42
• Page End: 53
• Publication Date: 2021-02
• Subjects: Distant Supervision Relation Extraction -- Dependency-path -- Adaptive method -- Additional knowledge graph supervision; Neural computers -- Periodicals; Neural networks (Computer science) -- Periodicals; Neural networks (Neurobiology) -- Periodicals; Nervous System -- Periodicals; Ordinateurs neuronaux -- Périodiques; Réseaux neuronaux (Informatique) -- Périodiques; Réseaux neuronaux (Neurobiologie) -- Périodiques; Neural computers; Neural networks (Computer science); Neural networks (Neurobiology); Periodicals; 006.32
• DOI: 10.1016/j.neunet.2020.10.012
• Languages: English
• ISSNs: 0893-6080

Abstract

Abstract: Relation Extraction systems train an extractor by aligning relation instances in Knowledge Base with a large amount of labeled corpora. Since the labeled datasets are very expensive, Distant Supervision Relation Extraction (DSRE) utilizes rough corpus annotated with Knowledge Graph to reduce the cost of acquisition. Nevertheless, the data noise problem limits the performance of the DSRE. Dependency trees can be used to filter the wrong-labeled instances in the distant supervision bag. However, existing dependency tree relation extraction strategies are all based on manually-set paths between the subject and object entities, and suffer from the problem of pruning the trees too aggressively or too insufficiently. To circumvent the shortcomings, in this paper, we propose a novel DSRE framework A 2 DSRE, based on the A daptive dependency-path and A dditional KG supervision. To obtain the dependency paths related to entity relations adaptively, we introduce an advanced graph neural network—GeniePath into DSRE, which assigns higher weights to those direct neighbor words that contribute more to relation prediction through breadth exploration, and conducts depth exploration to determine the correlation between relations and high-order neighbors. In this way, the irrelevant nodes are pruned while the relevant nodes are kept, our method can obtain more appropriate paths associated with relations. At the same time, to further reduce the noises in the data, we incorporate additional supervision information from the knowledge graph by retracting the margin between the representation of the bag and the pre-training knowledge graph embedding. Extensive numerical experiments validate the effectiveness of our new method.