How Can Cross-lingual Knowledge Contribute Better to Fine-Grained Entity Typing?

Findings of the Association for Computational Linguistics: ACL 2022, pages 3071 - 3081 May 22-27, 2022 c⃝2022 Association for Computational Linguistics How Can Cross-lingual Knowledge Contribute Better to Fine-Grained Entity Typing? Hailong Jin1,2, Tiansi Dong3, Lei Hou1,2∗, Juanzi Li1,2 Hui Chen4, Zelin Dai4, Qu Yincen4 1Department of Computer Science and Technology, BNRist 2KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China 3B-IT, University of Bonn, Germany 4Alibaba Group, Hangzhou, China {jinhl, houlei}@mail.tsinghua.edu.cn Abstract Cross-lingual Entity Typing (CLET) aims at im- proving the quality of entity type prediction by transferring semantic knowledge learned from rich-resourced languages to low-resourced lan- guages. In this paper, by utilizing multilingual transfer learning via the mixture-of-experts ap- proach, our model dynamically capture the re- lationship between target language and each source language, and effectively generalize to predict types of unseen entities in new languages. Extensive experiments on multi- lingual datasets show that our method signifi- cantly outperforms multiple baselines and can robustly handle negative transfer. We ques- tioned the relationship between language sim- ilarity and the performance of CLET. With a series of experiments, we refute the common- sense that the more source the better, and pro- pose the Similarity Hypothesis for CLET. 1 Introduction Fine-grained Entity Typing (FET) aims at labeling entity mentions in a particular context with one or more specific types organized in a type hierarchy. For example, Donald Trump is classified as hav- ing the path of following types: President, Politi- cian, Person. President is a subclass of Politician that in turn is a subclass of Person. FET provides accurate type information, and is therefore quite useful for various downstream NLP tasks, such as entity linking (Onoe and Durrett, 2020; Chen et al., 2020a; Zhu et al., 2020), relation extrac- tion (Vashishth et al., 2018; Kuang et al., 2020), text generation (Dong et al., 2021; Elsahar et al., 2018), and so on. Supervised learning approaches to FET need huge amount of labeled training data (Ren et al., 2016; Shi et al., 2020; Chen et al., 2020b), and can be applied for a few rich-resourced languages, e.g., English, which have enough qualified labeled data. ∗ Corresponding author Trump was inaugurated as the 45th president of the United States on January 20, 2017. During his 1月23日，特朗普签署了上任后第一份行政 命令，正式宣布美国退出跨太平洋战略经济 伙伴关系协议 Source: English Target: Chinese Root Loc. Person Org. City Politician University Artist Type Hierarchy instanceOf instanceOf ??? Figure 1: Example of Cross-lingual Entity Typing. We use knowledge from source language (English) to help with entity typing task in target language (Chinese). For the vast majority of low-resourced languages, we have insufficient training data, or even do not have labeled data at all. However, languages are not independent, instead, some are more similar than others, and form a family tree. For example, Portuguese is similar to Spanish; Dutch can even be thought of as half way between German and English. This motivates us to utilize the knowl- edge from rich-resourced languages (source) to help to predict missing types in a low-resourced language (target), which is called …

Content truncated to 500 words