Fine-grained Entity Typing without Knowledge Base

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5309–5319 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5309 Fine-grained Entity Typing without Knowledge Base∗ Jing Qian1, Yibin Liu2, Lemao Liu3, Yangming Li3, Haiyun Jiang3, Haisong Zhang3, Shuming Shi3 1University of California, Santa Barbara 2Peking University 3Tencent AI Lab jing_qian@cs.ucsb.edu liuybinpku@pku.edu.cn {redmondliu,newmanli,haiyunjiang}@tencent.com {hansonzhang,shumingshi}@tencent.com Abstract Existing work on Fine-grained Entity Typing (FET) typically trains automatic models on the datasets obtained by using Knowledge Bases (KB) as distant supervision. However, the re- liance on KB means this training setting can be hampered by the lack of or the incom- pleteness of the KB. To alleviate this limi- tation, we propose a novel setting for train- ing FET models: FET without accessing any knowledge base. Under this setting, we pro- pose a two-step framework to train FET mod- els. In the first step, we automatically cre- ate pseudo data with fine-grained labels from a large unlabeled dataset. Then a neural net- work model is trained based on the pseudo data, either in an unsupervised way or using self-training under the weak guidance from a coarse-grained Named Entity Recognition (NER) model. Experimental results show that our method achieves competitive performance with respect to the models trained on the origi- nal KB-supervised datasets. 1 Introduction Entity Typing is a fundamental task in Natural Language Processing. Traditional entity typing re- search focuses on a limited number of entity types while recent studies strive for finer granularity. A major challenge in fine-grained entity typing (FET) is the absence of human-annotated data. To ad- dress this problem, a common practice is to seek distant supervision from knowledge bases. Typi- cally, the training data is obtained by linking entity mentions and drawing their types from knowledge bases. For example, Ling and Weld (2012) uti- lize the information encoded in anchor links from Wikipedia text along with the type information in Freebase (Bollacker et al., 2008). After that, classi- fication models are trained using the resultant FET dataset, as illustrated in the left part of Figure 1. ∗The first two authors (Jing and Yibin) contributed equally to this work during the internships at Tencent AI Lab. Figure 1: A comparison between the previous FET framework (left) and our proposed FET framework (right). In the previous one, the type ontology is cou- pled with KB whereas it is not in ours. Despite its success, this framework has one limi- tation: The type ontology for training and testing is strongly restricted by the underlying KB. As a result, the application of the trained models under this framework can be hampered by the lack or the incompleteness of the KB (Choi et al., 2018). In real-world applications of FET, the ontology of the fine-grained types depends on the applications instead of the KB. Typically, there may not exist a one-to-one mapping from the ontology of appli- cations to that of the existing KB. For example, a game-related application may need to identify the named entities of some specific video …

Content truncated to 500 words