Learning Structured Representations of Entity Names using Active Learning and Weak Supervision

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6376–6383, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6376 Learning Structured Representations of Entity Names using Active Learning and Weak Supervision Kun Qian, Poornima Chozhiyath Raman, Lucian Popa, Yunyao Li IBM Research qian.kun@ibm.com, {pchozhi, lpopa, yunyaoli}@us.ibm.com Abstract Structured representations of entity names are useful for many entity-related tasks such as entity normalization and variant generation. representationsLearning the implicit structured of entity names without context and external knowledge is particularly challenging. In this paper, we present a novel learning framework that combines active learning and weak supervision to solve this problem. Our experimental evaluation show that this framework enables the learning of high-quality models from merely a dozen or so labeled examples. 1 Introduction Entity normalization and variant generation are fundamental for a variety of other tasks such as semantic search and relation extraction (Bhutani et al., 2018; Arasu and Kaushik, 2009). Given an entity name E, the goal of entity normalization is to convert E to a canonical form (e.g., “Jordan, Michael” → “Michael Jordan”), while the goal of entity variant generation is to convert E to a set of different textual representations that refer to the same entity as E (e.g., “Michael Jordan” → {“Jordan, Michael”, “MJ”, “M. Jordan”, . . .}). Typically, entity normalization and variant generation are done by first performing entity linking (Moro et al., 2014; Zhao et al., 2019; Li et al., 2017), i.e., matching entity names appearing in some context (e.g., free text) to named entities in curated knowledge bases (KBs), then use the canon- ical form or variations (of the linked entities) resid- ing in the KBs to complete the tasks. Unfortunately, in some scenarios, such as search (Thompson and Dozier, 1997), entity names are not surrounded by context. Furthermore, for specialized domain- specific applications, there may not be a knowledge base to govern the names of the relevant entities. Thus, entity linking is not always applicable. In this paper, we take the view that entity normaliza- tion and variant generation can be done without contextual information or external KBs if we understand the internal structures of entity names. As observed in (Campos et al., 2015; Bhutani et al., 2018; Arasu and Kaushik, 2009; Katiyar and Cardie, 2018; Finkel and Manning, 2009), entity names often have implicit structures that can be exploited to solve entity normalization and variant generation. Table 1 shows how we can manipu- late such structured representations of entity names to generate different variations without help from context or external knowledge. Declarative frameworks are proposed in (Arasu and Kaushik, 2009; Campos et al., 2015) to al- low developers to manually specify rules that parse entity names into a structured representation. To avoid such low-level manual effort, (Katiyar and Cardie, 2018; Finkel and Manning, 2009) used fully supervised methods for identifying nested entities embedded in flat named entities. Unfor- tunately, labeled data are rarely available to lever- age these methods in the real-world. To mitigate the need …

Content truncated to 500 words