DAMO-NLP at SemEval-2022 Task 11: A Knowledge-based System for Multilingual Named Entity Recognition

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1457 - 1468 July 14-15, 2022 ©2022 Association for Computational Linguistics DAMO-NLP at SemEval-2022 Task 11: A Knowledge-based System for Multilingual Named Entity Recognition Xinyu Wang⋄⋆, Yongliang Shen♠⋆, Jiong Cai⋄⋆, Tao Wang, Xiaobin Wang†, Pengjun Xie† Fei Huang†, Weiming Lu♠, Yueting Zhuang♠, Kewei Tu⋄, Wei Lu‡, Yong Jiang†∗ †DAMO Academy, Alibaba Group ⋄School of Information Science and Technology, ShanghaiTech University ♠College of Computer Science and Technology, Zhejiang University ‡StatNLP Research Group, Singapore University of Technology and Design {wangxy1,caijiong,tukw}@shanghaitech.edu.cn {syl,luwm}@zju.edu.cn, luwei@sutd.edu.sg yongjiang.jy@alibaba-inc.com Abstract The MultiCoNER shared task aims at detecting semantically ambiguous and complex named entities in short and low-context settings for multiple languages. The lack of contexts makes the recognition of ambiguous named entities challenging. To alleviate this issue, our team DAMO-NLP proposes a knowledge-based sys- tem, where we build a multilingual knowledge base based on Wikipedia to provide related con- text information to the named entity recogni- tion (NER) model. Given an input sentence, our system effectively retrieves related contexts from the knowledge base. The original input sentences are then augmented with such con- text information, allowing significantly better contextualized token representations to be cap- tured. Our system wins 10 out of 13 tracks in the MultiCoNER shared task.1 1 Introduction The MultiCoNER shared task (Malmasi et al., 2022b) aims at building Named Entity Recognition (NER) systems for 11 languages, including English, Spanish, Dutch, Russian, Turkish, Korean, Farsi, German, Chinese, Hindi, and Bangla. The task has three kinds of tracks including one multilingual track, 11 monolingual tracks and one code-mixed track. The multilingual track requires training mul- tilingual NER models that are able to handle all languages. The monolingual tracks require training individual monolingual models where each model works for only one language. The code-mixed track requires handling code-mixed samples (sen- tences that may involve multiple languages). The datasets mainly contain sentences from three do- mains: Wikipedia, web questions and user queries, ∗: project lead. ⋆: equal contributions. 1Our code is publicly available at https://github. com/Alibaba-NLP/KB-NER. köpings is rate Input Sentence: Retrieve in KB Predict with knowledge Predict without knowledge köpings | LOC is rate köpings is | GRP rate Retrieval Results: 1. Kis: Köpings IS, a Swedish sports club 2. Köping, Sweden: Köpings IS, association club, bandy and handball 3. Kalle Samuelsson: Kalle Samuelsson (born February 15, 1986) is a Swedish Bandy player who plays for Västerås SK as a goalkeeper. Kalle was a youth product of Köpings IS. 4. … Figure 1: A motivating example from the English test set. In the retrieval results, the bold phrases are the title of the retrieved page and the underlined phrases contain the hyperlinks to other pages. LOC and GRP are entity labels representing location and group respectively. which are usually short and low-context sentences. Moreover, these short sentences usually contain se- mantically ambiguous and complex entities, which makes the problem more difficult. In practice, professional annotators usually use their domain knowledge to disambiguate such kinds of entities. They …

Content truncated to 500 words