Fusing Heterogeneous Factors with Triaffine Mechanism for Nested Named Entity Recognition

Findings of the Association for Computational Linguistics: ACL 2022, pages 3174 - 3186 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Fusing Heterogeneous Factors with Triaffine Mechanism for Nested Named Entity Recognition Zheng Yuan12∗ Chuanqi Tan2 Songfang Huang2 Fei Huang2 1Tsinghua University 2Alibaba Group yuanz17@mails.tsinghua.edu.cn {chuanqi.tcq,songfang.hsf,f.huang}@alibaba-inc.com Abstract Nested entities are observed in many domains due to their compositionality, which cannot be easily recognized by the widely-used se- quence labeling framework. A natural solution is to treat the task as a span classification prob- lem. To learn better span representation and increase classification performance, it is cru- cial to effectively integrate heterogeneous fac- tors including inside tokens, boundaries, labels, and related spans which could be contributing to nested entities recognition. To fuse these heterogeneous factors, we propose a novel tri- affine mechanism including triaffine attention and scoring. Triaffine attention uses boundaries and labels as queries and uses inside tokens and related spans as keys and values for span rep- resentations. Triaffine scoring interacts with boundaries and span representations for classi- fication. Experiments show that our proposed method outperforms previous span-based meth- ods, achieves the state-of-the-art F1 scores on nested NER datasets GENIA and KBP2017, and shows comparable results on ACE2004 and ACE2005. 1 Introduction Named entity recognition (NER) is a fundamental natural language processing task that extracts enti- ties from texts. Flat NER has been well studied and is usually viewed as a sequence labeling problem (Lample et al., 2016). However, nested entities also widely exist in real-world applications due to their multi-granularity semantic meaning (Alex et al., 2007; Yuan et al., 2020), which cannot be solved by the sequence labeling framework since tokens have multiple labels (Finkel and Manning, 2009). Various paradigms for nested NER have been proposed in recent years. A representative direc- tion is the span-based approach that learns deep representation for every possible span and then classifies it to the corresponding type (Zheng et al., ∗ Work done at Alibaba DAMO Academy. a defective NF - chi B site was completely inactive in EBV - transformed B cells , … protein DNA cell type cell line Figure 1: An example sentence with nested entities from the GENIA dataset. 2019; Xia et al., 2019; Wadden et al., 2019; Tan et al., 2020; Wang et al., 2020; Yu et al., 2020). By leveraging the large-scale pretrained language model, several works show that the simple model structure for span representation and classification can achieve satisfactory results (Luan et al., 2019; Zhong and Chen, 2021). However, we still believe that explicit modeling of some relevant features will further benefit the span representation and clas- sification under the complex nested setting. Taking Figure 1 as an example, we claim that the following factors are critical for recognizing whether a span is an entity. (1) Tokens: It is obvious that tokens of the given span contribute to the recognition. (2) Boundaries: We emphasize boundaries (or bound- ary tokens) because they are special tokens with rich semantics. Works with simple structure may just produce …

Content truncated to 500 words