Learning from Sibling Mentions with Scalable Graph Inference in Fine-Grained Entity Typing

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2076 - 2087 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learning from Sibling Mentions with Scalable Graph Inference in Fine-Grained Entity Typing Yi Chen1,3∗, Jiayang Cheng2∗, Haiyun Jiang†, Lemao Liu Haisong Zhang, Shuming Shi, Ruifeng Xu1,4† 1School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China 2Department of Computer Science and Engineering, HKUST 3Joint Lab of HITSZ and China Merchants Securities, Shenzhen, China 4Peng Cheng Laboratory, Shenzhen, China yichennlp@gmail.com,jchengaj@cse.ust.hk xuruifeng@hit.edu.cn Abstract In this paper, we firstly empirically find that existing models struggle to handle hard men- tions due to their insufficient contexts, which consequently limits their overall typing perfor- mance. To this end, we propose to exploit sib- ling mentions for enhancing the mention rep- resentations. Specifically, we present two dif- ferent metrics for sibling selection and employ an attentive graph neural network to aggre- gate information from sibling mentions. The proposed graph model is scalable in that un- seen test mentions are allowed to be added as new nodes for inference. Exhaustive experi- ments demonstrate the effectiveness of our sib- ling learning strategy, where our model outper- forms ten strong baselines. Moreover, our ex- periments indeed prove the superiority of sib- ling mentions in helping clarify the types for hard mentions. 1 Introduction Fine-Grained Entity Typing (FGET) aims to as- sign one or more fine-grained types to an entity mention given its context. For instance, the men- tion Steve Jobs should be classified as Person and Entrepreneur under the context “Steve Jobs co- founded Apple ...”. Many tasks have witnessed the importance of FGET, such as relation extrac- tion (Jiang et al., 2020b; Chu et al., 2020; Jiang et al., 2020a; Cheng et al., 2021), entity linking (Onoe and Durrett, 2020), and other tasks (Jiang et al., 2020c; Zhang et al., 2020b; Liu et al., 2021b). It is challenging to learn effective representa- tions for contextualized mentions1 in many infor- mation extraction tasks (Gao et al., 2022), espe- ∗Equal Contribution †Corresponding Authors 1To simplify the statement, in the rest of this paper, the term “mention” is referred to as the contextualized mention, i.e., a mention accompanied with its context. cially in FGET, since the representations are re- quired to well distinguish fine-grained types with similar but different semantics. Noticeable efforts have been made to learn type-aware representa- tions for mentions (Ren et al., 2016; Xin et al., 2018; Choi et al., 2018; Zhang et al., 2018; Lin and Ji, 2019; Abhishek et al., 2017; Xu and Barbosa, 2018; Ali et al., 2020; Chen et al., 2021) and sig- nificant progress has been achieved. However, as supported by our empirical experiments, existing SOTA models perform poorly on a certain num- ber of “hard” mentions, leading to limited overall performance. The main reasons are the following challenges. First, the structure of some contexts surrounding the hard mentions are inherently too complex to extract informative features for identi- fying entity types. Second, …

Content truncated to 500 words