OPDAI at SemEval-2022 Task 11: A hybrid approach for Chinese NER using outside Wikipedia knowledge

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1488 - 1493 July 14-15, 2022 ©2022 Association for Computational Linguistics OPDAI at SemEval-2022 Task 11: A hybrid approach for Chinese NER using outside Wikipedia knowledge Ze Chen, Kangxu Wang, Jiewen Zheng Zijian Cai, Jiarong He and Jin Gao Interactive Entertainment Group of Netease Inc., Guangzhou, China {jackchen, wangkangxu, zhengjiewen}@corp.netease.com {caizijian01, gzhejiarong, jgao}@corp.netease.com Abstract This article describes the OPDAI submission to SemEval-2022 Task 11 on Chinese complex NER. First, we explore the performance of model-based approaches and their ensemble, finding that fine-tuning the pre-trained Chinese RoBERTa-wwm model with word semantic representation and contextual gazetteer repre- sentation performs best among single models. However, the model-based approach performs poorly on test data because of low-context and unseen-entity cases. Then, we extend our sys- tem into two stages: (1) generating entity can- didates by using neural model, soft-templates and Wikipedia lexicon. (2) predicting the fi- nal entity results within a feature-based rank model. For the evaluation, our best submission achieves an F1 score of 0.7954 and attains the third-best score in the Chinese sub-track. 1 Introduction Named Entity Recognition (NER)(Yadav and Bethard, 2019) aims to detecting the boundaries of named entities and recognizing their cate- gories(e.g., person or location). It plays an im- portant role in many downstream tasks, such as information extraction and question answering. SemEval-2022 Task 11: Multilingual Complex Named Entity Recognition(MultiCoNER) (Mal- masi et al., 2022b) is a shared task which en- courages participants to develop NER system to detect semantically ambiguous and complex enti- ties in short and low-context settings for 11 lan- guages. Participants can build a NER model that works only for one language or for all the lan- guages. And an additional track with code-mixed data are offered in this task. Different from ordi- nary NER, this task focuses on complex and unseen entities. Complex entities, like the titles of creative works(movie/book/software names) are harder to recognize. Additional test sets on questions and short search queries are offered in the test phase, which contains large proportion of unseen entities. Our main interest is to build a NER system which can process complex entities and adapt to other do- mains in practical scenarios in Chinese language. This paper describes our two-stage hybrid approach for Chinese NER. A description of datasets pro- vided in this shared task and additional datasets adopted in our system is given in Section 3. The implementation details of our system are listed in Section 4. We first experiment with model-based methods and integrate word semantic feature and gazetteer feature with neural model to improve in- ference performance. Further, we extend our model system to a two-stage prediction system: entity candidates generation and entity confidence rank- ing. Confidence ranking is used to pick out high- confidence entities from candidates. With the help of Wikipedia lexicon and soft templates for gener- ating entity candidates, our hybrid system shows good performance and great domain adaption capa- bility in the final evaluation phase. 2 Related Work …

Content truncated to 500 words