PAI at SemEval-2023 Task 2: A Universal System for Named Entity Recognition with External Entity Information

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 744–750 July 13-14, 2023 ©2023 Association for Computational Linguistics PAI at SemEval-2023 Task 2: A Universal System for Named Entity Recognition with External Entity Information Long Ma∗, Kai Lu∗, Tianbo Che, Hailong Huang, Weiguo Gao, Xuan Li {malong633,lukai113,chetianbo263,huanghailong590}@pingan.com.cn {gaoweiguo801,lixuan208}@pingan.com.cn Ping An Life Insurance Company of China, Ltd. Abstract The MultiCoNER II task aims to detect com- plex, ambiguous, and fine-grained named en- tities in low-context situations and noisy sce- narios like the presence of spelling mistakes and typos for multiple languages. The task poses significant challenges due to the scarcity of contextual information, the high granularity of the entities(up to 33 classes), and the inter- ference of noisy data. To address these issues, our team PAI proposes a universal Named En- tity Recognition (NER) system that integrates external entity information to improve perfor- mance. Specifically, our system retrieves enti- ties with properties from the knowledge base (i.e. Wikipedia) for a given text, then concate- nates entity information with the input sentence and feeds it into Transformer-based models. Fi- nally, our system wins 2 first places, 4 second places, and 1 third place out of 13 tracks. The code is publicly available at https://github. com/diqiuzhuanzhuan/semeval-2023. 1 Introduction The objective of the MultiCoNER II shared task (Fetahu et al., 2023b) is to develop a robust named entity recognition (NER) system for multiple lan- guages (including English, Spanish, Hindi, Bangla, Chinese, Swedish, Farsi, French, Italian, Por- tuguese, Ukranian, German and aforementioned mixed languages) that can perform well even when the input data has spelling errors, lacks contextual information, or contains out of knowledge base en- tities. This task entails the recognition of a total of 33 fine-grained entity types within six coarse- grained entity categories. The entity types within each category are highly prone to confusion and ambiguity, and the testing data includes noise such as spelling errors that must be considered. Similar to the previous competition (Malmasi et al., 2022b), the data in this competition contains a lot of short *These authors contributed equally to this work. and low-context sentences, which means the ac- curate prediction of entity types relies heavily on external knowledge. Therefore, the key objective of our current work is to achieve effective integration of the model and external knowledge. In this paper, extending our previous work (Ma et al., 2022), which has demon- strated the efficacy of the Dictionary-fused model, we propose a NER system based on Entity Prop- erty Knowledge Base. We constructed our knowl- edge base using all entities and their associated properties (including entity name, alias, sitelink title, description, instanceof, subclassof and occu- pation) from WikiData. In the retrieval module, we use string matching to retrieve entities with prop- erties for a given sentence. Given the sequence length limitations of our model, we keep longer entities in sentences. Subsequently, we put the en- tity and property values into the external context in a specific format and feed the external context and original sentence …

Content truncated to 500 words