FLAT: Chinese NER Using Flat-Lattice Transformer

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6836–6842 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6836 FLAT: Chinese NER Using Flat-Lattice Transformer Xiaonan Li, Hang Yan, Xipeng Qiu∗, Xuanjing Huang Shanghai Key Laboratory of Intelligent Information Processing, Fudan University School of Computer Science, Fudan University lixiaonan xdu@outlook.com, {hyan19, xpqiu, xjhuang}@fudan.edu.cn Abstract Recently, the character-word lattice structure has been proved to be effective for Chinese named entity recognition (NER) by incorpo- rating the word information. However, since the lattice structure is complex and dynamic, most existing lattice-based models are hard to fully utilize the parallel computation of GPUs and usually have a low inference-speed. In this paper, we propose FLAT: Flat-LAttice Transformer for Chinese NER, which converts the lattice structure into a flat structure con- sisting of spans. Each span corresponds to a character or latent word and its position in the original lattice. With the power of Trans- former and well-designed position encoding, FLAT can fully leverage the lattice informa- tion and has an excellent parallelization ability. Experiments on four datasets show FLAT out- performs other lexicon-based models in perfor- mance and efficiency. 1 Introduction Named entity recognition (NER) plays an indis- pensable role in many downstream natural lan- guage processing (NLP) tasks (Chen et al., 2015; Diefenbach et al., 2018). Compared with English NER (Lample et al., 2016; Yang et al., 2017; Liu et al., 2017; Sun et al., 2020), Chinese NER is more difficult since it usually involves word segmenta- tion. Recently, the lattice structure has been proved to have a great benefit to utilize the word infor- mation and avoid the error propagation of word segmentation (Zhang and Yang, 2018). We can match a sentence with a lexicon to obtain the latent words in it, and then we get a lattice like in Figure 1(a). The lattice is a directed acyclic graph, where each node is a character or a latent word. The lat- tice includes a sequence of characters and potential ∗Corresponding author. 和 And 庆 Qing 人 People 重 Chong 店 Shop 药 Drug 人和药店 Renhe Pharmacy 重庆 Chongqing 药店 Pharmacy (a) Lattice. 和 And 庆 Qing 人 People 重 Heavy 店 Shop 药 Drug 人和药店 Renhe Pharmacy 重庆 Chongqing 药店 Pharmacy B-LOC E-LOC E-LOC B-LOC I-LOC I-LOC (b) Lattice LSTM. Renhe Pharmacy 和 And 庆 Qing 人 People 重 Chong 店 Shop 药 Drug 人和药店 重庆 Chongqing 药店 Pharmacy 1 1 2 2 3 3 4 4 5 5 6 6 1 2 3 6 5 6 Transformer Encoder B-LOC E-LOC E-LOC B-LOC I-LOC I-LOC (c) Flat-Lattice Transformer. Figure 1: While lattice LSTM indicates lattice struc- ture by dynamically adjusting its structure, FLAT only needs to leverage the span position encoding. In 1(c), , , denotes tokens, heads and tails, respectively. words in the sentence. They are not ordered se- quentially, and the word’s first character and last character determine its position. Some words in lattice may be important for NER. For example, in Figure …

Content truncated to 500 words