Leveraging Type Descriptions for Zero-shot Named Entity Recognition and Classification

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1516–1528 August 1–6, 2021. ©2021 Association for Computational Linguistics 1516 Leveraging Type Descriptions for Zero-shot Named Entity Recognition and Classification Rami Aly1, Andreas Vlachos1, Ryan McDonald2∗ 1Computer Laboratory, University of Cambridge, U.K. 2ASAPP {rami.aly|andreas.vlachos}@cl.cam.ac.uk, ryanmcd@asapp.com Abstract A common issue in real-world applications of named entity recognition and classification (NERC) is the absence of annotated data for target entity classes during training. Zero- shot learning approaches address this issue by learning models that can transfer information from observed classes in the training data to unseen classes. This paper presents the first approach for zero-shot NERC, introducing a novel architecture that leverage the fact that textual descriptions for many entity classes oc- cur naturally. Our architecture addresses the zero-shot NERC specific challenge that the not-an-entity class is not well defined, since different entity classes are considered in train- ing and testing. For evaluation, we adapt two datasets, OntoNotes and MedMentions, em- ulating the difficulty of real-world zero-shot learning by testing models on the rarest en- tity classes. Our proposed approach outper- forms baselines adapted from machine read- ing comprehension and zero-shot text classifi- cation. Furthermore, we assess the effect of different class descriptions for this task. 1 Introduction Named entity recognition and classification (NERC) is the task of identifying spans of text corresponding to named entities and classifying these spans from a set of pre-defined entity classes. A prevalent issue for many real-world applications is that annotated data does not readily exist. This motivates the focus on the zero-shot setting (Xian et al., 2018; Wang et al., 2019), where annotated data is not available for the classes of interest. Instead, information available from observed classes must be transferred to unseen target classes. Recently zero-shot approaches making use of textual representations to represent entity classes ∗Work done when author was working at Google. were explored for entity linking (EL) (Logeswaran et al., 2019; Wu et al., 2020) and named entity typ- ing (NET) (Obeidat et al., 2019), which are similar to the NERC subtask of named entity classification (NEC). However, no previous work has addressed the task of zero-shot NERC, which additionally re- quires the detection of which tokens make up an entity in addition to its type, i.e. Named Entity Recognition (NER). This paper is the first to study zero-shot NERC, by leveraging entity type descriptions. The task is illustrated in Figure 1. During testing, the in- put is a sentence and a set of target entity classes. each accompanied by its description, and the goal is to recognize and classify entities in these tar- get classes. Descriptions contain crucial informa- tion for the task. Given as input “Shantou Har- bour, a natural river seaport, opens to the South China Sea.” and a class Facility in Figure 1, us- ing a description “Names of human-made struc- tures: infrastructure (streets, bridges), [...]” connection between Facility and Shantou Harboura can be …

Content truncated to 500 words