Domain-Specific NER via Retrieving Correlated Samples

Proceedings of the 29th International Conference on Computational Linguistics, pages 2398–2404 October 12–17, 2022. 2398 Domain-Specific NER via Retrieving Correlated Samples Xin Zhang1, Yong Jiang, Xiaobin Wang, Xuming Hu2, Yueheng Sun3, Pengjun Xie, Meishan Zhang4∗ 1School of New Media and Communication, Tianjin University, China 2School of Software, Tsinghua University, China 3College of Intelligence and Computing, Tianjin University, China 4Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China hsinz@tju.edu.cn, hxm19@mails.tsinghua.edu.cn jiangyong.ml@gmail.com, czwangxiaobin@foxmail.com yhs@tju.edu.cn, xpjandy@gmail.com, zhangmeishan@hit.edu.cn Abstract Successful Machine Learning based Named En- tity Recognition models could fail on texts from some special domains, for instance, Chinese ad- dresses and e-commerce titles, where requires adequate background knowledge. Such texts are also difficult for human annotators. In fact, we can obtain some potentially helpful infor- mation from correlated texts, which have some common entities, to help the text understand- ing. Then, one can easily reason out the correct answer by referencing correlated samples. In this paper, we suggest enhancing NER models with correlated samples. We draw correlated samples by the sparse BM25 retriever from large-scale in-domain unlabeled data. To ex- plicitly simulate the human reasoning process, we perform a training-free entity type calibrat- ing by majority voting. To capture correlation features in the training stage, we suggest to model correlated samples by the transformer- based multi-instance cross-encoder. Empirical results on datasets of the above two domains show the efficacy of our methods. 1 Introduction Named Entity Recognition (NER), which first lo- cates entity positions and then labels their types se- quentially, is a fundamental topic in both academia and industry (Li et al., 2022). Normal NER models consider the input samples to be independent of each other, learning the common intra-instance pat- terns and making predictions in a sequential way. This paradigm has shown surprising successes in decades, especially with the help of emerging deep learning (Shang et al., 2018; Zhang et al., 2018b; Liu et al., 2019; Luo et al., 2020; Lison et al., 2020; Fang et al., 2021; Meng et al., 2021). However, learned models will fail at some hard cases, which would be inevitably encountered in ∗Corresponding author. 吉林⽩城镇赉县⽕车站 (Train Station, Zhen-Lai County, Bai-Cheng, Ji-Lin) 吉林⽩城市镇赉县站前街⽕车站 (Train Station, Zhan-Qian Street, Zhen-Lai County, Bai-Cheng City, Ji-Lin) 吉林省⽩城市镇赉县⼴通东路⽕车头商店 (Trainhead Shop, Guang-Tong East Road, Zhen-Lai County, Bai-Cheng City, Ji-Lin Province) 吉林⽩城镇赉站前街春丽旅店 (Chun-Li Hotel, Zhan-Qian Street, Zhen-Lai, Bai-Cheng, Ji-Lin) 吉林省⽩城市镇赉县站前街富达⼩区0栋 (Building 0, Fu-Da Community, Zhan-Qian Street, Zhen-Lai County, Bai-Cheng City, Ji-Lin Province) 赉县⽕车站 (Lai County Train Station) POI 吉林 (Ji-Lin) City ⽩城镇 (Bai-Cheng Town) Town ⽕车站 (Train Station) POI 吉林 (Ji-Lin) Prov ⽩城 (Bai-Cheng) City 镇赉县 (Zhen-Lai County) Dist Ground- Truth Prediction Input Retrieved Samples 镇 (Zhen) = 镇 (Town) Figure 1: An address example with retrieved texts. The model incorrectly predicted “白城镇 (Baicheng Town)” and “赉县火车站 (Lai County Train Station)” because they match common patterns, i.e. “XX镇” (Xx Town) and “X县火车站” (X County Train Station). “吉林” is ambiguous, which is both a province and a city. real scenarios (Li et al., 2019; Ding et al., 2019). Figure 1 shows an example …

Content truncated to 500 words