Neural Chinese Address Parsing

Proceedings of NAACL-HLT 2019, pages 3421–3431 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3421 Neural Chinese Address Parsing Hao Li and Wei Lu StatNLP Research Group Singapore University of Technology and Design hao li@mymail.sutd.edu.sg luwei@sutd.edu.sg Pengjun Xie and Linlin Li DAMO Academy Alibaba Group chengchen.xpj@alibaba-inc.com linyan.lll@alibaba-inc.com Abstract This paper introduces a new task – Chinese address parsing – the task of mapping Chi- nese addresses into semantically meaningful chunks. While it is possible to model this problem using a conventional sequence la- belling approach, our observation is that there exist complex dependencies between labels that cannot be readily captured by a simple linear-chain structure. We investigate neural structured prediction models with latent vari- ables to capture such rich structural informa- tion within Chinese addresses. We create and publicly release a new dataset consisting of 15,000 Chinese addresses, and conduct ex- tensive experiments on the dataset to investi- gate the model effectiveness and robustness. We release our code and data at http:// statnlp.org/research/sp. 1 Introduction Addresses play an important role in modern soci- ety. They are typically used as identifiers to lo- cations and entities in the world that can be used to facilitate various social activities, such as busi- ness correspondences, meetings and events. Re- cent research efforts show that systems that per- form address parsing, the task of automatically parsing addresses into semantically meaningful structures, can be useful for tasks such as build- ing e-commerce or product recommendation sys- tems (Jia et al., 2017; Avvenuti et al., 2018). Due to historical reasons, the English addresses come with a standardized format, mostly written in order from most specific to most general. Meaningful chunks in an English address are also separated by punctuation or the new-line symbols. Such char- acteristics make parsing English addresses a rela- tively easy task. However, addresses written in eastern Asian languages such as Chinese present several unique 浙江省杭州市拱墅区登云路639号1号楼电子市场230飞阳电子 浙江省PROVINCE 杭州市CITY 拱墅区DISTRICT (Zhejiang Province) (Hangzhou City) (Gongshu District) 登云路ROAD 639号ROADNO 1号楼HOUSENO (Dengyun Road) (No. 639) (Unit #1) 电子市场POI 230ROOMNO 飞阳电子SUBPOI (Electronic Market) (Feiyang Dianzi LLC.) 观沙街道观沙小区观沙嘉园安置小区9栋5单元9栋5单元1705 观沙街道TOWN 观沙小区POI 观沙嘉园SUBPOI (Guansha Town) (Guansha Residence) (Guansha Sub-residence) 安置小区PERSON 9栋HOUSENO 5单元CELLNO (Anzhi Sector) (Block 9) (Unit #5) 9栋REDUNDANT 5单元REDUNDANT 1705ROOMNO (Block 9) (Unit #5) Figure 1: Two example Chinese addresses and the ex- pected structures after parsing. Each chunk is under- lined with its corresponding label in blue. challenges. Unlike English addresses, Chinese addresses are typically written in the form of a consecutive sequence of Chinese characters (pos- sibly intermixed with digits and English letters). Figure 1 presents two example Chinese addresses and their desired output structures after parsing – chunks annotated with their labels indicating se- mantics (such as province, road, etc). The Chi- nese addressing system is also different from that of English. Though it is generally believed that the system uses the opposite ordering – starting from most general (e.g., province) and ending with most specific (e.g., room no.), in practice it can be ob- served that the format is far …

Content truncated to 500 words