Learning to Bootstrap for Entity Set Expansion

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 292–301, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 292 Learning to Bootstrap for Entity Set Expansion Lingyong Yan1,3, Xianpei Han1,2,∗, Le Sun1,2, Ben He3,1,∗ 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China {lingyong2014, xianpei, sunle}@iscas.ac.cn, benhe@ucas.ac.cn Abstract Bootstrapping for Entity Set Expansion (ESE) aims at iteratively acquiring new instances of a specific target category. Traditional bootstrap- ping methods often suffer from two problems: 1) delayed feedback, i.e., the pattern evalua- tion relies on both its direct extraction qual- ity and the extraction quality in later iterations. 2) sparse supervision, i.e., only few seed enti- ties are used as the supervision. To address the above two problems, we propose a novel boot- strapping method combining the Monte Carlo Tree Search (MCTS) algorithm with a deep similarity network, which can efficiently esti- mate delayed feedback for pattern evaluation and adaptively score entities given sparse su- pervision signals. Experimental results con- firm the effectiveness of the proposed method. 1 Introduction Bootstrapping is widely used for Entity Set Ex- pansion (ESE), which acquires new instances of a specific category by iteratively evaluating and se- lecting patterns, while extracting and scoring en- tities. For example, given seeds {London, Paris, Beijing} for capital entity expansion, a bootstrap- ping system for ESE iteratively selects effective patterns, e.g., “the US Embassy in *”, and extracts new capital entities, e.g., Moscow, using the se- lected patterns. The main challenges of effective bootstrapping for ESE owe to the delayed feedback and the sparse supervision. Firstly, bootstrapping is an iterative process, where noises brought by cur- rently selected patterns can affect successive itera- tions (Movshovitz-Attias and Cohen, 2012; Qadir et al., 2015). Indeed, the pattern evaluation relies on not only its direct extraction quality but also the extraction quality in later iterations, which are correspondingly denoted as instant feedback and ∗Corresponding authors. ���������� ���������� * is the capital of the meeting held in * will be held in * ���������� ���������� ������� �������� ����� ������� ������ ������� ����� ������ ��������� ������ ������ �������� ������ ������� �������� ����� ���� ����� ������ ������ ������ ���� ����� ����� ���������� ���������� the Japanese Embassy in * the US Embassy in * * is a big city ��� ��� Figure 1: Example of the delayed feedback problem, which expands the capital seeds {London, Paris, Bei- jing}. We demonstrate the top entities extracted by a pattern (correct ones are shown in blue), where P is the precision of extracted entities. delayed feedback in this paper. For instance, as shown in Figure 1, although “* is a big city” and “the US Embassy in *” have equal direct extraction quality, the former is considered less useful since its later extracted entities are mostly unrelated. As a result, selecting patterns with high instant …

Content truncated to 500 words