Bitext Name Tagging for Cross-lingual Entity Annotation Projection

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 461–470, Osaka, Japan, December 11-17 2016. Bitext Name Tagging for Cross-lingual Entity Annotation Projection Dongxu Zhang1, Boliang Zhang2, Xiaoman Pan2, Xiaocheng Feng3, Heng Ji2, Weiran Xu1 1Beijing University of Posts and Telecommunications, Beijing, China zhangdongxuu@gmail.com, xuweiran@bupt.edu.cn 2Rensselaer Polytechnic Institute, NY, USA {zhangb8,panx2,jih}@rpi.edu 3Harbin Institute of Technology, Harbin, China xcfeng@ir.hit.edu.cn Abstract Annotation projection is a practical method to deal with the low resource problem in incident languages (IL) processing. Previous methods on annotation projection mainly relied on word alignment results without any training process, which led to noise propagation caused by word alignment errors. In this paper, we focus on the named entity recognition (NER) task and propose a weakly-supervised framework to project entity annotations from English to IL through bitexts. Instead of directly relying on word alignment results, this framework combines advantages of rule-based methods and deep learning methods by implementing two steps: First, generates a high-confidence entity annotation set on IL side with strict searching methods; Second, uses this high-confidence set to weakly supervise the model training. The model is finally used to accomplish the projecting process. Experimental results on two low-resource ILs show that the proposed method can generate better annotations projected from English-IL parallel corpora. The performance of IL name tagger can also be improved significantly by training on the newly projected IL annotation set. 1 Introduction Annotation projection task aims to deal with low resource issues where human annotations are limited or unavailable in incident languages or domains. Since supervised learning algorithms can not work without annotation sets, annotation projection methods could automatically generate annotations from another language or domain where rich annotation sets are available, such as English. Yarowsky and Ngai (2001) proposed a method of using parallel text with word alignment results to project annotations. Fig. 1 shows an example of entity projection with word alignment results from English to Turkish. On English side, Voice of America Radio and Congo are tagged as an organization (ORG) and a location (LOC) respectively by an English name tagger. The dashed lines represent word alignment results generated by a word alignment tool. Following alignment results, we can project labels to Amerikann Sesi and Congo on Turkish side automatically. A major problem of this framework is that it suffers from noises produced by word alignment errors. Thus, some de-noising methods have been proposed (Kim et al., 2010; Wang and Manning, 2014). Though promising, this framework has several disadvantages. One shortcoming is that it totally de- pends on word alignment results. In this case, noise propagation from word alignment errors is heavily troublesome. To alleviate this problem, there are post-processing methods for annotation correction (Kim et al., 2010) and soft expectations to make use of more probabilistic information inside word alignment results (Wang and Manning, 2014). Although post-processing correction is efficient to filter out wrong labels, it can hardly find back labels which have been lost in the word alignment step. The soft expecta- tion …

Content truncated to 500 words