Generic and Trend-aware Curriculum Learning for Relation Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2202 - 2213 July 10-15, 2022 ©2022 Association for Computational Linguistics Generic and Trend-aware Curriculum Learning for Relation Extraction in Graph Neural Networks Nidhi Vakil Department of Computer Science University of Massachusetts Lowell nvakil@cs.uml.edu Hadi Amiri Department of Computer Science University of Massachusetts Lowell hadi@cs.uml.edu Abstract We present a generic and trend-aware curricu- lum learning approach for graph neural net- works. It extends existing approaches by in- corporating sample-level loss trends to better discriminate easier from harder samples and schedule them for training. The model effec- tively integrates textual and structural infor- mation for relation extraction in text graphs. Experimental results show that the model pro- vides robust estimations of sample difficulty and shows sizable improvement over the state- of-the-art approaches across several datasets. 1 Introduction Relation extraction is the task of detecting (of- ten pre-defined) relations between entity pairs. It has been investigated in both natural language pro- cessing (Mintz et al., 2009; Lin et al., 2016; Peng et al., 2017; Zhang et al., 2018) and network sci- ence (Zhang and Chen, 2018; Fout et al., 2017). Relation extraction is a challenging task, especially when data is scarce. Nonetheless, the ability to automatically link entity pairs is a crucial task as it can reveal relations that have not been previously identified, e.g., informing clinicians about a causal relation between a gene and a phenotype or dis- ease. Figure 1 shows an example sentence from a PubMed article in the Gene Phenotype Relation (PGR) dataset (Sousa et al., 2019), which describes the application domain of the present work as well. Previous research has extensively investigated relation extraction at both sentence (Zeng et al., 2015; dos Santos et al., 2015; Sousa et al., 2019) and document (Yao et al., 2019b; Quirk and Poon, 2017) levels. Furthermore, effective graph-based neural network approaches have been developed for various prediction tasks on graphs, including link prediction between given node pairs (Kipf and Welling, 2017; Hamilton et al., 2017; Xu et al., 2018; Veliˇckovi´c et al., 2018). Several recent ap- proaches (Li et al., 2020; Zhang and Chen, 2018; Our study further emphasizes that NDUFS6 sequence should be analyzed in patients presenting with lethal neonatal lactic acidemia because of isolated complex I deficiency Phenotype Gene Causal Figure 1: An example showing the report of a causal relation between a gene and a phenotype (symptom) from the PGR dataset (Sousa et al., 2019). Alsentzer et al., 2020) illustrated the importance of enhancing graph neural networks using structurally- informed features such as shortest paths, random walks and node position features. In this work, we develop a graph neural network titled Graph Text Neural Network (GTNN) that employs structurally-informed node embeddings as well as textual descriptions of nodes at predic- tion layer to avoid information loss for relation extraction. GTNN can be trained using a standard approach where data samples are fed to the network in a random …

Content truncated to 500 words