Cross-domain Named Entity Recognition via Graph Matching

Findings of the Association for Computational Linguistics: ACL 2022, pages 2670 - 2680 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Cross-domain Named Entity Recognition via Graph Matching Junhao Zheng, Haibin Chen, Qianli Ma* School of Computer Science and Engineering, South China University of Technology, Guangzhou, China junhaozheng47@outlook.com, haibin_chen@foxmail.com, qianlima@scut.edu.cn∗ Abstract Cross-domain NER is a practical yet challeng- ing problem since the data scarcity in the real- world scenario. A common practice is first to learn a NER model in a rich-resource general domain and then adapt the model to specific do- mains. Due to the mismatch problem between entity types across domains, the wide knowl- edge in the general domain can not effectively transfer to the target domain NER model. To this end, we model the label relationship as a probability distribution and construct label graphs in both source and target label spaces. To enhance the contextual representation with label structures, we fuse the label graph into the word embedding output by BERT. By rep- resenting label relationships as graphs, we for- mulate cross-domain NER as a graph matching problem. Furthermore, the proposed method has good applicability with pre-training meth- ods and is potentially capable of other cross- domain prediction tasks. Empirical results on four datasets show that our method outperforms a series of transfer learning, multi-task learning, and few-shot learning methods. 1 Introduction Named entity recognition (NER) is a crucial com- ponent in many language understanding tasks (Shaalan, 2014; Nadeau and Sekine, 2007) and is often applied in various domains. Due to the data scarcity in the real-world scenario, obtaining ad- equate domain-specific data is usually expensive and time-consuming. Hence, cross-domain NER, which is capable of adapting NER models to spe- cific domains with limited data, has been drawing increasing attention in recent years. However, one of the primary challenges of cross- domain NER is the mismatch between source and target domain labels (Yang and Katiyar, 2020). For example, the label sets between ATIS (Hakkani-Tür et al., 2016) and CoNLL 2003 are non-overlapping. ∗*Corresponding author Figure 1: A demonstration of graph matching. In both two cases, our model learns graph structures from the source label space and makes correct predictions. In two label spaces, each node is a target label and the matching nodes and edges are opaque. To address this issue, some approaches utilize multi-task learning (Jia and Zhang, 2020; Wang et al., 2020) for transferring knowledge across do- mains. However, these methods require full train- ing on both source and target domain data when adapting to each new domain. Since the source do- main dataset is usually much larger than the target domain dataset, the multi-task learning methods are inefficient when adapting to low-resource domains. Recently, as Pre-trained Language Models (PLMs) such as BERT (Devlin et al., 2019) have shown remarkable success in NER, transfer- learning-based methods also show effectiveness for cross-domain NER. A typical approach is to first train a NER model initialized with PLM on rich-resource domain (e.g., CoNLL 2003 (Sang and Meulder, 2003)), …

Content truncated to 500 words