Learning a Unified Named Entity Tagger from Multiple Partially Annotated Corpora for Efficient Adaptation

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 515–527 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 515 Learning A Unified Named Entity Tagger From Multiple Partially Annotated Corpora For Efficient Adaptation Xiao Huang1,2, Li Dong3∗, Elizabeth Boschee1, Nanyun Peng1,2 1Information Sciences Institute, University of Southern California 2Department of Computer Science, University of Southern California 3Outreach, Inc. {huan183, lidong}@usc.edu, {boschee, npeng}@isi.edu Abstract Named entity recognition (NER) identifies typed entity mentions in raw text. While the task is well-established, there is no universally used tagset: often, datasets are annotated for use in downstream applications and accord- ingly only cover a small set of entity types relevant to a particular task. For instance, in the biomedical domain, one corpus might an- notate genes, another chemicals, and another diseases—despite the texts in each corpus con- taining references to all three types of enti- ties. In this paper, we propose a deep struc- tured model to integrate these “partially an- notated” datasets to jointly identify all entity types appearing in the training corpora. By leveraging multiple datasets, the model can learn robust input representations; by build- ing a joint structured model, it avoids po- tential conflicts caused by combining several models’ predictions at test time. Experiments show that the proposed model significantly outperforms strong multi-task learning base- lines when training on multiple, partially an- notated datasets and testing on datasets that contain tags from more than one of the training corpora.1 1 Introduction Named Entity Recognition (NER), which identi- fies the boundaries and types of entity mentions from raw text, is a fundamental problem in natu- ral language processing (NLP). It is a basic com- ponent for many downstream tasks, such as rela- tion extraction (Hasegawa et al., 2004; Mooney and Bunescu, 2005), coreference resolution (Soon et al., 2001), and knowledge base construction (Craven et al., 1998; Craven and Kumlien, 1999). ∗ Work done while the author was at USC ISI. 1The code and the datasets will be made available at https://github.com/xhuang28/NewBioNer One problem in NER is the diversity of en- tity types, which vary in scope for different do- mains and downstream tasks. Traditional NER for the news domain focuses on three coarse- grained entity types: person, location, and organi- zation (Tjong Kim Sang and De Meulder, 2003). However, as NLP technologies have been applied to a broader set of domains, many other entity types have been targeted. For instance, Ritter et al. (2011) add seven new entity types (e.g., product, tv-show) on top of the previous three when an- notating tweets. Other efforts also define differ- ent but partially overlapping sets of entity types (Walker et al., 2006; Ji et al., 2010; Consortium, 2013; Aguilar et al., 2014). These non-unified annotation schemas result in partially annotated datasets: each dataset is only annotated with a subset of possible entity types. One approach to this problem is to train individ- ual NE taggers for each partially annotated dataset and combine their results using some heuristics. Figure 1 shows …

Content truncated to 500 words