Empower Entity Set Expansion via Language Model Probing

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8151–8160 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8151 Empower Entity Set Expansion via Language Model Probing Yunyi Zhang1, Jiaming Shen1, Jingbo Shang2, Jiawei Han1 1 University of Illinois at Urbana-Champaign, IL, USA 2 University of California San Diego, CA, USA 1{yzhan238, js2, hanj}@illinois.edu 2 jshang@ucsd.edu Abstract Entity set expansion, aiming at expanding a small seed entity set with new entities belong- ing to the same semantic class, is a critical task that benefits many downstream NLP and IR applications, such as question answering, query understanding, and taxonomy construc- tion. Existing set expansion methods boot- strap the seed entity set by adaptively select- ing context features and extracting new enti- ties. A key challenge for entity set expansion is to avoid selecting ambiguous context fea- tures which will shift the class semantics and lead to accumulative errors in later iterations. In this study, we propose a novel iterative set expansion framework that leverages automati- cally generated class names to address the se- mantic drift issue. In each iteration, we select one positive and several negative class names by probing a pre-trained language model, and further score each candidate entity based on selected class names. Experiments on two datasets show that our framework generates high-quality class names and outperforms pre- vious state-of-the-art methods significantly. 1 Introduction Entity set expansion aims to expand a small set of seed entities (e.g., {“United States”, “China”, “Canada”}) with new entities (e.g., “United King- dom”, “Australia”) belonging to the same semantic class (i.e., Country). The entities so discovered may benefit a variety of NLP and IR applications, such as question answering (Wang et al., 2008), query understanding (Hua et al., 2017), taxonomy construction (Shen et al., 2018a), and semantic search (Xiong et al., 2017; Shen et al., 2018b). Most existing entity set expansion methods boot- strap the initial seed set by iteratively selecting context features (e.g., co-occurrence words (Pantel et al., 2009), unary patterns (Rong et al., 2016), and coordinational patterns (Mamou et al., 2018)), Entities Hearst Pattern [NP0] such as [NP1],[NP2], and [NP3] {USA, China, Canada} [MASK] such as USA, China, and Canada Class-probing Query Class Name Entity Hearst Pattern [NP0], [NP1],or other [NP2] countries Canada Entity-probing Query Canada, [MASK], or other countries Retrieved Class Names countries states large countries … cities … Retrieved Entities Japan United Kingdom Mexico … Toronto … LanguageModel (e.g. BERT/XLNet) Figure 1: Examples of class-probing and entity- probing queries generated based on Hearst patterns. while extracting and ranking new entities. A key challenge to set expansion is to avoid selecting am- biguous patterns that may introduce erroneous en- tities from other non-target semantic classes. Take the above class Country as an example, we may find some ambiguous patterns like “* located at” (which will match more general Location enti- ties) and “match against *” (which may be asso- ciated with entities in the Sports Club class). Furthermore, as bootstrapping is an iterative pro- cess, …

Content truncated to 500 words