NCUEE-NLP at SemEval-2022 Task 11: Chinese Named Entity Recognition Using the BERT-BiLSTM-CRF Model

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1597 - 1602 July 14-15, 2022 ©2022 Association for Computational Linguistics NCUEE-NLP at SemEval-2022 Task 11: Chinese Named Entity Recognition Using the BERT-BiLSTM-CRF Model Lung-Hao Lee, Chien-Huan Lu, and Tzu-Mi Lin Department of Electrical Engineering National Central University No. 300, Zongda Rd., Zhongli Dist., Taoyan City 32001, Taiwan lhlee@ee.ncu.edu.tw, 10952107@ncu.edu.tw,110521087@ncu.edu.tw Abstract This study describes the model design of the NCUEE-NLP system for the Chinese track of the SemEval-2022 MultiCoNER task. We use the BERT embedding for character represen- tation and train the BiLSTM-CRF model to recognize complex named entities. A total of 21 teams participated in this track, with each team allowed a maximum of six submissions. Our best submission, with a macro-averaging F1-score of 0.7418, ranked the seventh position out of 21 teams. 1 Introduction Named Entity Recognition (NER) is a fundamental task in information extraction that locates the men- tions of named entities and classifies them (e.g., person, organization and location) in unstructured texts. The NER is a traditional NLP task that has been solved as a sequence labeling problem, where entity boundaries and category labels are jointly predicted. It is difficult to recognize com- plex named entities like the titles of creative works (e.g., books, songs, movies) that can take the form of any linguistic constituent (Ashwini and Choi, 2014). Syntactic and semantic ambiguity makes it challenging to recognize such complex named entities based on their context. The SemEval-2022 Task 11 (MultiCoNER) orga- nized a challenge to develop multilingual complex NER system for 11 human languages (Malmasi et al., 2022b). This task focuses on detecting se- mantically ambiguous and complex entities in short and low-context settings. The languages include: English, Spanish, Dutch, Russian, Turkish, Ko- rean, Farsi, German, Chinese, Hindi, and Bangla. The named entity categories are Person (labeled as PER), Location (LOC), Group (GRP), Corpora- tion (CORP), Product (PROD), and Creative Work (CW). The task evaluation framework is divided in three broad tracks. 1) Multi-lingual (Track 1): participants train a single multi-lingual NER model for all the languages; 2) Mono-lingual (Track 2-12): participants train a model that works for only one language; 3) Code-mixed (Track 13): testing sam- ples include tokens from any of the 11 mentioned languages in the shared task. We only participated the Track 9 for Chinese language. Chinese NER is more difficult to process than English NER. Chinese language is logographic and provides no conventional features like capitaliza- tion. In addition, due to a lack of delimiters be- tween characters, Chinese NER is correlated with word segmentation tasks, and named entity bound- aries are also word boundaries. However, incor- rectly segmented entity boundaries will cause error propagation in NER. For example, in a short con- text “這首歌出現在華特迪士尼動畫動物方城 市中” (This song appeared in the Walt Disney ani- mation Zootopia), a creative work “動物方城市” (Zootopia) may be incorrectly segmented into three words: “動物” (animal), “方”(square), and “城市” (city). Hence, it has been shown that character- based approaches outperform word-based methods for Chinese NER …

Content truncated to 500 words