Improving Fine-grained Entity Typing with Entity Linking

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6210–6215, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6210 Improving Fine-grained Entity Typing with Entity Linking Hongliang Dai1, Donghong Du1,3, Xin Li2, and Yangqiu Song1 1Department of CSE, HKUST 2Tencent Technology (SZ) Co., Ltd. 1{hdai,yqsong}@cse.ust.hk 2alonsoli@tencent.com 3dduaa@connect.ust.hk Abstract Fine-grained entity typing is a challenging problem since it usually involves a relatively large tag set and may require to understand the context of the entity mention. In this pa- per, we use entity linking to help with the fine- grained entity type classification process. We propose a deep neural model that makes pre- dictions based on both the context and the in- formation obtained from entity linking results. Experimental results on two commonly used datasets demonstrates the effectiveness of our approach. On both datasets, it achieves more than 5% absolute strict accuracy improvement over the state of the art. 1 Introduction Given a piece of text and the span of an entity men- tion in this text, fine-grained entity typing (FET) is the task of assigning fine-grained type labels to the mention (Ling and Weld, 2012). The as- signed labels should be context dependent (Gillick et al., 2014). For example, in the sentence “Trump threatens to pull US out of World Trade Organi- zation,” the mention “Trump” should be labeled as /person and /person/politician, although Donald Trump also had other occupations such as busi- nessman, TV personality, etc. This task is challenging because it usually uses a relatively large tag set, and some mentions may require the understanding of the context to be cor- rectly labeled. Moreover, since manual annotation is very labor-intensive, existing approaches have to rely on distant supervision to train models (Ling and Weld, 2012; Ghaddar and Langlais, 2018). Thus, the use of extra information to help with the classification process becomes very important. In this paper, we improve FET with entity linking (EL). EL is helpful for a model to make typing de- cisions because if a mention is correctly linked to its target entity, we can directly obtain the type in- formation about this entity in the knowledge base (KB). For example, in the sentence “There were some great discussions on a variety of issues fac- ing Federal Way,” the mention “Federal Way” may be incorrectly labeled as a company by some FET models. Such a mistake can be avoided after link- ing it to the city Federal Way, Washington. For cases that require the understanding of the context, using entity linking results is also beneficial. In the aforementioned example where “Trump” is the mention, obtaining all the types of Donald Trump in the knowledge base (e.g., politician, business- man, TV personality, etc.) is still informative for inferring the correct type (i.e., politician) that fits the context, since they narrows the possible labels down. However, the information obtained through EL should not be fully trusted since it is not …

Content truncated to 500 words