Type-enriched Hierarchical Contrastive Strategy for Fine-Grained Entity Typing

Proceedings of the 29th International Conference on Computational Linguistics, pages 2405–2417 October 12–17, 2022. 2405 Type-enriched Hierarchical Contrastive Strategy for Fine-Grained Entity Typing Xinyu Zuo, Haijin Liang, Ning Jing, Shuang Zeng, Zhou Fang and Yu Luo Tencent Inc. {xylonzuo,hodgeliang,shuangzeng,akirafang,yamiluo}@tencent.com ning.jing.ustc@gmail.com Abstract Fine-grained entity typing (FET) aims to de- duce specific semantic types of the entity men- tions in text. Modern methods for FET mainly focus on learning what a certain type looks like. And few works directly model the type differences, that is, let models know the extent that one type is different from others. To alle- viate this problem, we propose a type-enriched hierarchical contrastive strategy for FET. Our method can directly model the differences be- tween hierarchical types and improve the abil- ity to distinguish multi-grained similar types. On the one hand, we embed type into entity contexts to make type information directly per- ceptible. On the other hand, we design a con- strained contrastive strategy on the hierarchi- cal structure to directly model the type dif- ferences, which can simultaneously perceive the distinguishability between types at differ- ent granularity. Experimental results on three benchmarks, BBN, OntoNotes, and FIGER show that our method achieves significant per- formance on FET by effectively modeling type differences. 1 Introduction Entity typing is a fundamental research problem in natural language processing (NLP), which aims to deduce the semantic types of the entity mentions in text. With the deepening of text understanding, the type sets of entities become more refined and rang- ing in from dozens (Hovy et al., 2006) to hundreds (Weischedel and Brunstein, 2005; Ling and Weld, 2012) or thousands (Choi et al., 2018). Therefore, fine-grained entity typing (FET) has gained more attention, which focuses on assigning more spe- cific types to entities. For sentence in Figure 1, a FET system needs to assign a coarse-grained type "/person" and a fine-grained type "/person/actor" to the entity "Vivien Leigh". The inferred fine- grained types could provide more specific prior knowledge for downstream NLP tasks, such as He described his portrait of actress Vivien Leigh as lit by a top spotlight diffused by tracing paper. root person location building actor coach county city airport hospital ... ... ... ... FET Model athlete country library Figure 1: Example of fine-grained entity typing based on FIGER ontology. Green Box: the scope of visi- ble types about "person" of the fine-grained contrastive strategy. Red Box: the scope of visible types of the coarse-grained contrastive strategy. question answering (Lee et al., 2006) and entity linking (Leszczynski et al., 2022). Considering the partial ontology of the FIGER dataset (Ling and Weld, 2012) in Figure 1, fine- grained entity types are often linked together in a hierarchical taxonomy, which makes the type boundaries increasingly blurred, especially for sub- types under the same coarse type. As shown in Fig- ure 1, the fine-grained types "coach", "athlete" and "actor", all of which fall into the coarse-grained type "person", are less differentiated. In order to identify fine-grained types, prior work has concentrated …

Content truncated to 500 words