Keyphrase Generation via Soft and Hard Semantic Corrections

Procedingsofthe202ConferenceonEmpiricalM ethodsinNaturalLanguageProcesing ,pages825-838 December7-1,202©202AsociationforComputationalLinguistics LightEA: A Scalable, Robust, and Interpretable Entity Alignment Framework via Three-view Label Propagation Xin Mao1,2, Wenting Wang3, Yuanbin Wu1,2, Man Lan1,2 1School of Computer Science and Technology, East China Normal University 2Shanghai Institute of AI for Education, East China Normal University 3TikTok Group, Singapore xmao@stu.ecnu.edu.cn,{ybwu,mlan}@cs.ecnu.edu.cn wenting.wang@bytedance.com Abstract Entity Alignment (EA) aims to find equivalent entity pairs between KGs, which is the core step of bridging and integrating multi-source KGs. In this paper, we argue that existing GNN- based EA methods inherit the inborn defects from their neural network lineage: weak scal- ability and poor interpretability. Inspired by recent studies, we reinvent the Label Propa- gation algorithm to effectively run on KGs and propose a non-neural EA framework — LightEA, consisting of three efficient compo- nents: (i) Random Orthogonal Label Genera- tion, (ii) Three-view Label Propagation, and (iii) Sparse Sinkhorn Iteration. According to the extensive experiments on public datasets, LightEA has impressive scalability, robustness, and interpretability. With a mere tenth of time consumption, LightEA achieves comparable results to state-of-the-art methods across all datasets and even surpasses them on many. 1 Introduction Knowledge Graph (KG) describes the real-world entities and their internal relations by triples (head, rel, tail), expressing the information on the internet in a form closer to human cognition. To this day, KGs have facilitated a mount of down- stream internet applications (e.g., search engines (Yang et al., 2019b) and dialogue systems (Yang et al., 2020)) and become one of the core driving forces in the development of artificial intelligence. In practice, KGs are usually constructed by vari- ous departments with multi-source data. Therefore, they typically contain complementary knowledge while having overlapping parts. Integrating these independent KGs could significantly improve the coverage rate, which is especially beneficial to low- resource language users. Entity Alignment (EA) aims to find equivalent entity pairs between KGs (as shown in Figure 1), which is the core step of bridging and integrating ? Figure 1: A toy example of entity alignment. multi-source KGs. Therefore, EA attracts enor- mous attention and progresses rapidly. Most exist- ing methods regard EA as a graph representation learning task and share the same two-stage archi- tecture: (i) encoding the KGs into low-dimensional spaces via graph encoders (e.g., TransE (Bordes et al., 2013) and GCN (Kipf and Welling, 2016)) and (ii) mapping the embeddings of equivalent en- tity pairs into a unified vector space through con- trastive losses (Hadsell et al., 2006). Recently, Graph Neural Network (GNN) has achieved impressive success in many sorts of graph applications. Following Wang et al. (2018b), who first introduced Graph Convolutional Net- work (GCN) into EA, numerous new fancy mecha- nisms are proposed and stacked over vanilla GCN for better performance, such as Graph Matching (Fey et al., 2020), Relational Attention (Mao et al., 2020a), and Hyperbolic embedding (Sun et al., 2020a). According to an EA paper list 1 on Github, over 90% of EA methods adopted GNNs as their graph encoders in recent three years. However, there is no such …

Content truncated to 500 words