Distantly Supervised NER with Partial Annotation Learning and Reinforcement Learning

Proceedings of the 27th International Conference on Computational Linguistics, pages 2159–2169 Santa Fe, New Mexico, USA, August 20-26, 2018. 2159 Distantly Supervised NER with Partial Annotation Learning and Reinforcement Learning Yaosheng Yang, Wenliang Chen, Zhenghua Li, Zhengqiu He, Min Zhang Institute of Artificial Intelligence School of Computer Science and Technology, Soochow University, China {ysyang, zqhe}@stu.suda.edu.cn {wlchen, zhli13, minzhang}@suda.edu.cn Abstract A bottleneck problem with Chinese named entity recognition (NER) in new domains is the lack of annotated data. One solution is to utilize the method of distant supervision, which has been wide- ly used in relation extraction, to automatically populate annotated training data without human- cost. The distant supervision assumption here is that if a string in text is included in a predefined dictionary of entities, the string might be an entity. However, this kind of auto-generated data suf- fers from two main problems: incomplete and noisy annotations, which affect the performance of NER models. In this paper, we propose a novel approach which can partially solve the above problems of distant supervision for NER. In our approach, to handle the incomplete problem, we apply partial annotation learning to reduce the effect of unknown labels of characters. As for noisy annotation, we design an instance selector based on reinforcement learning to distinguish positive sentences from auto-generated annotations. In experiments, we create two datasets for Chinese named entity recognition in two domains with the help of distant supervision. The exper- imental results show that the proposed approach obtains better performance than the comparison systems on both two datasets. 1 Introduction In recent years, deep learning approaches have achieved great progress in the task of named entity recog- nition (NER) (Collobert et al., 2011; Chiu and Nichols, 2015). The standardized approach is that using BiLSTMs for encoding and then applying CRF for jointly label decoding (Huang et al., 2015; Lample et al., 2016). In addition, BiLSTMs and CNNs are employed to model character- or word-level representa- tions (Ma and Hovy, 2016). Most previous studies on NER focus on a certain set of predefined NER types, such as organization, location, person, date, and so on, where a certain amount of labeled data is provided to train the models. However, different applications require particular entity types, such as “Brand” and “Product” in E- commerce domain, and “Company” for finance industry. Considering the high cost of human annotation, it may not be feasible to annotate large amounts of labeled data for each new NER type, but small-scale data is available at some time. As an alternative solution, distant supervision can automatically generate large-scale labeled data for new-type NER without human-cost. The idea of distant supervision has widely used in the task of relation extraction (Mintz et al., 2009; Riedel et al., 2010; Zeng et al., 2015). For relation extraction, at first we have a knowledge base. If two entities e1 and e2 have relation r according to the knowledge base, then we populate this knowledge and assume the relation between e1 and e2 is r in …

Content truncated to 500 words