CMB AI Lab at SemEval-2022 Task 11: A Two-Stage Approach for Complex Named Entity Recognition via Span Boundary Detection and Span Classification

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1603 - 1607 July 14-15, 2022 ©2022 Association for Computational Linguistics CMB AI Lab at SemEval-2022 Task 11: A Two-Stage Approach for Complex Named Entity Recognition via Span Boundary Detection and Span Classification Keyu PU, Hongyi LIU, Yixiao YANG, Jiangzhou JI, Wenyi LV and Yaohan HE CMB AI Lab {pukeyu, lhy24, yangyixiao}@cmbchina.com {jesse, lvwymail, heyh18}@cmbchina.com Abstract This paper presents a solution for the SemEval- 2022 Task 11 Multilingual Complex Named Entity Recognition. What is challenging in this task is detecting semantically ambiguous and complex entities in short and low-context settings. Our team (CMB AI Lab) propose a two-stage method to recognize the named entities: first, a model based on biaffine layer is built to predict span boundaries, and then a span classification model based on pooling layer is built to predict semantic tags of the spans. The basic pre-trained models we choose are XLM-RoBERTa and mT5. The evaluation result of our approach achieves an F1 score of 84.62 on sub-task 13, which ranks the third on the learder board. 1 Introduction Named entity recognition (NER)(Tjong Kim Sang and De Meulder, 2003) is a fundamental task in natural language processing, aiming at identifying the spans of texts that refer to entities. NER is widely applied to information extraction and data mining(Lin et al., 2019)(Cao et al., 2019), which is greatly challenging in practical and open domain settings. However, the previous research has not paid much attention on processing complex and ambiguous named entities. SemEval 2022 task 11 (Malmasi et al., 2022b) containing a total of 13 sub-tasks is a complex NER task which focuses on detecting semantically ambiguous and complex entities in short and low- context settings (Meng et al., 2021). For the pur- pose of testing the domain adaption capability of the participating models, the task not only set 11 base sub-tasks: English, Spanish, Dutch, Russian, Turkish, Korean, Farsi, German, Chinese, Hindi and Bangla, but also set two additional testing sets on questions and short queries: Multilingual, and code-mixed (Fetahu et al., 2021). We conduct a two-stage method to deal with the code-mixed sub- task, which achieves an F1 score of 84.62. This paper is structured as follows. The related work of NER is briefly introduced in Section 2. The data for training and testing the model is presented in Section 3. The details of the two-stage method is described in Section 4. The experimental results of our method are exhibited in Section 5. Section 6 summarizes this paper. 2 Related Work In the NLP field, the NER task is usually consid- ered as a sequence labeling problem (Liu et al., 2018) (Lin et al., 2019) (Cao et al., 2019). With well-designed features, CRF-based models have achieved the leading performance (Lafferty et al., 2001) (Finkel et al., 2005) (Liu et al., 2011). Re- cently, neural network models have been exploited for feature representations (Chen and Manning, 2014). Moreover, contextualized word represen- tations such as ELMo (Peters et al., 2018) …

Content truncated to 500 words