Description-Based Zero-shot Fine-Grained Entity Typing

Proceedings of NAACL-HLT 2019, pages 807–814 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 807 Description-Based Zero-shot Fine-Grained Entity Typing Rasha Obeidat, Xiaoli Fern, Hamed Shahbazi and Prasad Tadepalli School of Electrical Engineering and Computer Science, Oregon State University {obeidatr, Xfern, shahbazh, tadepall}@eecs.oregonstate.edu Abstract Fine-grained Entity typing (FGET) is the task of assigning a fine-grained type from a hierar- chy to entity mentions in the text. As the tax- onomy of types evolves continuously, it is de- sirable for an entity typing system to be able to recognize novel types without additional train- ing. This work proposes a zero-shot entity typ- ing approach that utilizes the type description available from Wikipedia to build a distributed semantic representation of the types. During training, our system learns to align the entity mentions and their corresponding type repre- sentations on the known types. At test time, any new type can be incorporated into the sys- tem given its Wikipedia descriptions. We eval- uate our approach on FIGER, a public bench- mark entity tying dataset. Because the exist- ing test set of FIGER covers only a small por- tion of the fine-grained types, we create a new test set by manually annotating a portion of the noisy training data. Our experiments demon- strate the effectiveness of the proposed method in recognizing novel types that are not present in the training data. 1 Introduction Entity Typing assigns a semantic type (e.g., per- son, location, organization) to an entity mention in text based on the local context. It is useful for enhancing a variety of Natural Language Process- ing(NLP) tasks such as question answering (Han et al., 2017; Das et al., 2017), relation extrac- tion (Liu et al., 2014; Yaghoobzadeh et al., 2016), and entity linking (Stern et al., 2012). Traditional Named Entity Typing systems consider a small set of coarse types (e.g., person, location, organiza- tion) (Tjong Kim Sang and De Meulder, 2003; Kr- ishnan and Manning, 2006; Chieu and Ng, 2002). Recent studies address larger sets of fine-grained types organized in type hierarchies (e.g., per- son/artist, person/author) (Ling and Weld, 2012; Corro et al., 2015; Xu and Barbosa, 2018; Murty et al., 2018). Fine-Grained Entity Typing (FGET) is usually approached as a multi-label classifica- tion task where an entity mention can be assigned multiple types that usually constitute a path in the hierarchy (Ren et al., 2016). In real-world scenarios, there is a need to deal with ever-growing type taxonomies. New types emerge, and existing types are refined into finer sub-categories. Traditional methods for entity typ- ing assume that the training data contains all pos- sible types, thus require new annotation effort for each new type that emerges. Zero-shot learning (ZSL), a special kind of transfer learning, allows for new types to be incorporated at the predic- tion stage without the need for additional annota- tion and retraining. The main idea behind ZSL is to learn a shared semantic space for representing both the seen and unseen types, which …

Content truncated to 500 words