Label Embedding for Zero-shot Fine-grained Named Entity Typing

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 171–180, Osaka, Japan, December 11-17 2016. Label Embedding for Zero-shot Fine-grained Named Entity Typing Yukun Ma1,2, Erik Cambria1,2, Sa Gao1,2 1School of Computer Science and Engineering, Nanyang Technological University, Singapore 2Rolls-Royce@NTU Corporate Lab, Nanyang Technological University, Singapore mayu0010@e.ntu.edu.sg, cambria@ntu.edu.sg, gaos0011@e.ntu.edu.sg Abstract Named entity typing is the task of detecting the types of a named entity in context. For instance, given “Eric is giving a presentation”, our goal is to infer that ‘Eric’ is a speaker or a presenter and a person. Existing approaches to named entity typing cannot work with a growing type set and fails to recognize entity mentions of unseen types. In this paper, we present a label embedding method that incorporates prototypical and hierarchical information to learn pre-trained label em- beddings. In addition, we adapt a zero-shot framework that can predict both seen and previously unseen entity types. We perform evaluation on three benchmark datasets with two settings: 1) few-shots recognition where all types are covered by the training set; and 2) zero-shot recog- nition where fine-grained types are assumed absent from training set. Results show that prior knowledge encoded using our label embedding methods can significantly boost the performance of classification for both cases. 1 Introduction Named entity typing (NET) is the task of inferring types of named entity mentions in text. NET is a useful pre-processing step for many natural language processing (NLP) tasks, e.g., auto-categorization and sentiment analysis. Named entity linking, for instance, can use NET to refine entity candidates of a given mention (Ling and Weld, 2012). Besides, NET is capable of supporting applications based on a deeper understanding of natural language, e.g., knowledge completion (Dong et al., 2014) and question answering (Lin et al., 2012; Fader et al., 2014). Standard NET approaches or sometime known as named entity recognition (Chinchor and Robinson, 1997; Tjong Kim Sang and De Meulder, 2003; Doddington et al., 2004) are concerned with coarse-grained types (e.g, person, location, organization) that are flat in structure. In comparison, fine-grained named entity typing (FNET) (Ling and Weld, 2012), which has been studied as an extension of standard NET task, uses a tree-structured taxonomy including not only coarse-grained types but also fine-grained types of named entities. For instance, given “[Intel] said that over the past decade”, standard NET only classifies Intel as organization, whereas FNET further classifies it as organization/corporation. FNET is faced with two major challenges: growing type set and label noises. Since the type hierarchy of entities is typically built from knowledge bases such as DBpedia, which is regularly updated with new types (especially fine-grained types) and entities, it is natural to assume that the type hierarchy is growing rather than fixed over time. However, current FNET systems are impeded from handling a growing type set for that information learned from training set cannot be transferred to unseen types. Another problem with FNET is that the weakly supervised tagging process used for automatically generating labeled data …

Content truncated to 500 words